Energy router and distributed energy storage collaborative planning method considering load growth

Abstract

The application discloses a kind of energy router and distributed energy storage collaborative planning method considering load growth, comprising: obtaining the historical source load data of target power distribution network, construct historical net load time series after preprocessing, decompose into different central frequency intrinsic mode function, and divide into low-frequency trend term, medium-frequency periodic term and high-frequency random term;Respectively, low-frequency trend prediction component, medium-frequency periodic prediction component and high-frequency random prediction component are obtained;Based on the three prediction components, obtain the typical day scene set of future target year and corresponding weight;As input, build a double-layer optimization model;The upper model decides equipment location and size, with the goal of minimizing investment cost and voltage deviation, the lower model carries out time sequence operation simulation and feeds back the result;Riemann manifold evolution algorithm is used to solve output pareto optimal planning scheme set and determine the final planning scheme through evaluation. The application can significantly improve the accuracy, robustness, economy and operation reliability of the planning scheme.

Description

Technical Field

[0001] This invention belongs to the field of power system planning and energy internet technology, specifically relating to a collaborative planning method for energy routers and distributed energy storage that takes into account load growth. Background Technology

[0002] With the advancement of the "dual carbon" goals, the penetration rate of distributed power sources, represented by photovoltaics and wind power, in distribution networks has been increasing year by year, driving the evolution of traditional power grids into an energy internet. However, the randomness and volatility of distributed power sources exacerbate the mismatch between source and load distribution in time and space, posing severe challenges to voltage stability, power flow control, and power quality of distribution networks. Against this backdrop, energy routers, as core hub devices of the energy internet, have become key to solving these problems due to their multi-port energy management and flexible power flow control capabilities. Meanwhile, distributed energy storage, as a key resource for improving grid flexibility, has also seen extensive research on its planning methods.

[0003] However, existing distribution network planning studies are mostly based on current load and renewable energy output levels, and generally suffer from the following shortcomings: First, scenario construction is lagging. Traditional planning input data is often based on typical daily snapshots or simple growth rate extrapolation, lacking consideration of the nonlinearity and uncertainty brought about by potential source load growth (such as large-scale electric vehicle access), resulting in insufficient capacity or voltage exceeding limits in the early stages of operation, lacking robustness. Second, coordination mechanisms are lacking. Existing studies often configure EER (Electrical / Electric Energy Router) and DES (Distributed Energy System) independently, failing to fully explore their complementary benefits in the spatiotemporal scale. DES focuses on energy shifting in the time domain, while EER has flexible power flow transfer capabilities in the spatial domain. If the configuration scheme of the two lacks deep coupling, it is difficult to achieve the optimal balance between investment economy and operational flexibility. Third, there are bottlenecks in solution efficiency. Distribution network collaborative planning is a typical two-level nonlinear integer programming problem, involving massive time-series scenarios and complex non-convex constraints. Traditional heuristic algorithms suffer from low sampling efficiency and are prone to getting trapped in local optima when facing high-dimensional search spaces. They are also computationally expensive and cannot accurately capture the Pareto front within a limited number of simulations.

[0004] Therefore, how to extract evolutionary patterns from historical data, generate forward-looking future scenarios, and establish a collaborative planning model that can balance economic efficiency and power quality while being highly efficient in solving problems is a theoretical support and technical challenge that urgently needs to be addressed in the current construction of power distribution networks. Summary of the Invention

[0005] To overcome the problems of delayed scenario prediction, insignificant synergistic effects, and low algorithm efficiency in existing power distribution network planning, this invention provides a collaborative planning method for energy routers and distributed energy storage that takes load growth into account. The technical problem to be solved by this invention is achieved through the following technical solution: A collaborative planning method for energy routers and distributed energy storage that takes into account load growth includes: The historical source-load data of the target distribution network for several consecutive years is obtained. After preprocessing, a historical net load time series is constructed. The variational mode decomposition algorithm is used to decompose the historical net load time series into eigenmode functions with different center frequencies, and divides them into low-frequency trend terms, medium-frequency periodic terms and high-frequency random terms. The low-frequency trend term is predicted for future trends to obtain a low-frequency trend prediction component; the mid-frequency periodic term is predicted based on its historical periodic structure; the high-frequency random term is reconstructed using a conditional diffusion model to obtain a high-frequency random prediction component; based on the three prediction components, a set of typical daily scenarios and their corresponding weights for the future target year are obtained through superposition clustering; the future target year is the year following the latest year in the historical source load data. Using the typical daily scenario set and corresponding weights as input, a two-layer optimization model for collaborative planning is constructed. The upper-layer model is used to decide on the site selection and capacity determination scheme of energy routers and distributed energy storage, with the goal of minimizing the annualized investment cost and the cumulative voltage deviation of the system. The lower-layer model is used to perform time-series operation simulation of the distribution network under the given upper-layer scheme, and feeds back the simulation results to the upper layer for scheme evaluation. The Riemannian manifold evolution algorithm is used to solve the two-level optimization model, and the Pareto optimal planning scheme set is output. The final planning scheme is determined by evaluating the Pareto optimal planning scheme set.

[0006] The present invention can have the following beneficial effects: 1. It can generate high-fidelity forward-looking source-load scenarios that include potential load surge trends over the next few years and retain random fluctuation characteristics, enabling planning schemes to have redundancy to cope with the large-scale access of new loads in the future, and significantly enhancing engineering robustness.

[0007] 2. By deeply coupling the spatial power flow flexibility control of the energy router (EER) with the time energy shift characteristics of distributed energy storage (DES) through a two-layer model, voltage fluctuations are effectively mitigated without significantly increasing investment costs, thereby greatly improving the distribution network's ability to absorb distributed power sources.

[0008] 3. The Riemannian Manifold Evolution Algorithm (RMEA), designed with information geometry theory, can automatically sense the electrical coupling relationship between decision variables. It overcomes the shortcomings of traditional algorithms, such as blind searching in high-dimensional non-convex spaces and easy getting trapped in local optima, and achieves faster convergence speed and more uniform Pareto solution set distribution.

[0009] 4. While minimizing investment costs, the system ensures optimal voltage deviation across the entire network through precise lower-level power flow simulation, achieving the best balance between investment economy and operational power quality. This effectively avoids redundant construction caused by over-investment in power assets or insufficient capacity. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a collaborative planning method for energy routers and distributed energy storage that takes into account load growth, provided in an embodiment of the present invention. Figure 2 The flowchart shows the algorithm solution for the collaborative planning method of energy router and distributed energy storage that takes into account load growth, as provided in the embodiments of the present invention. Figure 3 The diagram illustrates the power distribution network source-load evolution characteristics and data-driven forward-looking scenario generation mechanism in the energy router and distributed energy storage collaborative planning method considering load growth provided in this embodiment of the invention. Figure 4 The logic diagram of the two-layer optimization model in the energy router and distributed energy storage collaborative planning method considering load growth provided in the embodiments of the present invention; Figure 5 This is a schematic diagram of the random component reconstruction principle based on the diffusion model in the energy router and distributed energy storage collaborative planning method that takes into account load growth provided in the embodiments of the present invention. Detailed Implementation

[0011] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0012] This invention provides a collaborative planning method for energy routers and distributed energy storage that takes into account load growth. Please refer to [link / reference]. Figure 1 and Figure 2 Understood, this method may include the following steps S1 to S5, specifically: S1. Obtain historical source-load data of the target distribution network for several consecutive years. After preprocessing, construct a historical net load time series. Use variational mode decomposition algorithm to decompose the historical net load time series into eigenmode functions with different center frequencies, and divide them into low-frequency trend terms, medium-frequency periodic terms and high-frequency random terms. S2, perform future trend prediction on the low-frequency trend term to obtain the low-frequency trend prediction component; obtain the medium-frequency cycle prediction component based on the historical cycle structure of the medium-frequency cycle term; reconstruct the high-frequency random term using the conditional diffusion model to obtain the high-frequency random prediction component; based on the three prediction components obtained, obtain the typical daily scene set and corresponding weights for the future target year through superposition clustering; the future target year is the year following the latest year in the historical source load data; S3. Using the typical daily scenario set and corresponding weights as input, a two-layer optimization model for collaborative planning is constructed. The upper-layer model is used to decide on the location and capacity scheme of energy routers and distributed energy storage, with the goal of minimizing the annualized investment cost and the cumulative voltage deviation of the system. The lower-layer model is used to perform distribution network time-series operation simulation under the given upper-layer scheme, and feeds the simulation results back to the upper layer for scheme evaluation. S4. The Riemannian manifold evolution algorithm is used to solve the two-level optimization model and output the Pareto optimal planning scheme set. S5. The final planning scheme is determined by evaluating the Pareto optimal planning scheme set.

[0013] The core design point of the energy router and distributed energy storage collaborative planning method considering load growth provided in this embodiment of the invention is as follows: 1. A forward-looking source-load scenario generation technique based on variational mode decomposition (VMD) and conditional diffusion model is proposed, which can accurately capture the long-term trend term and short-term random fluctuations of the load.

[0014] 2. A two-layer optimization model that takes into account both life-cycle cost and voltage offset index was constructed, realizing deep synergy between EER spatial power flow regulation and DES time energy shift.

[0015] 3. By introducing information geometry theory, the Riemannian manifold evolution algorithm (RMEA) was designed. Through natural gradient updates and Fisher information matrix-guided search, the solution efficiency and solution set distribution of complex planning problems are significantly improved.

[0016] The following explains each step separately.

[0017] For step S1, historical source-load data of the target distribution network for several consecutive years are obtained. After preprocessing, a historical net load time series is constructed. The variational mode decomposition algorithm is used to decompose the historical net load time series into intrinsic mode functions with different center frequencies, and divides them into low-frequency trend terms, medium-frequency periodic terms and high-frequency random terms. The target distribution network in this embodiment of the invention mainly refers to a modern distribution network facing challenges such as a high proportion of new energy access and new loads, such as urban commercial area or high-tech park distribution networks, industrial (park) distribution networks, and integrated photovoltaic, energy storage and charging community distribution networks.

[0018] As an example, the target distribution network of this invention can be a distribution network with a high proportion of distributed power sources and electric vehicle loads. The purpose of the method of this invention is to generate forward-looking scenarios through data-driven approaches and to use information geometry algorithms to plan the collaborative site selection and capacity determination of energy routers (EER) and distributed energy storage (DES).

[0019] This distribution network, which includes a high proportion of distributed generation and electric vehicle loads, is a typical form of modern active distribution network. Its core characteristics are: On the power supply side: A large number of distributed photovoltaic (PV) and distributed wind power (WT) sources are connected. PV output exhibits a clear diurnal cycle with strong weather correlation, characterized by "high output during the day and low output on sunny days and low output on cloudy days"; wind power output, on the other hand, shows even stronger intermittency and random fluctuations. Together, they constitute a high-proportion, highly uncertain power source.

[0020] On the load side: In addition to traditional residential, commercial and industrial infrastructure loads, large-scale electric vehicle (EV) charging loads have become an important component. Their charging behavior has a dual spatiotemporal randomness: in terms of time, it is related to users' commuting and work-rest habits, with charging peaks occurring at night or during specific time periods during the day; in terms of space, the distribution of charging piles is closely related to the layout of residential areas, commercial areas and fast charging stations.

[0021] Step S1 involves acquiring the original source payload data and extracting prospective features. In an optional implementation, S1 may specifically include the following steps S11 to S14: S11, Obtain historical source and load data of the target distribution network over the past several years, including historical load data and distributed generation output data; The cumulative years corresponding to historical source load data can be set according to planning requirements. As an example, the following text will use the acquisition of historical source load data for the past 5 consecutive years of the target distribution network as an example.

[0022] Historical load data is used to characterize load-side power changes in residential, industrial, commercial, and electric vehicle charging sectors. It includes not only the baseline load curve but also electric vehicle charging power data, exhibiting a stronger peak-to-valley difference, a faster load ramp-up rate, and a unique fluctuation pattern influenced by travel habits. The data for a single moment in historical load data represents the load power.

[0023] Distributed generation output data is used to characterize the output changes of distributed power sources such as photovoltaic (PV) and / or wind power. It can be understood that distributed generation output data records the intermittent power output of PV and wind power affected by weather and seasons. The data for a single moment in the distributed generation output data represents the output power of the distributed generation.

[0024] S12, after aligning the historical load data and the distributed power output data by time, preprocessing is performed. The preprocessing includes missing value imputation, outlier removal, noise reduction and smoothing, thereby constructing a historical net load time series. The historical load data and the distributed power output data are time-aligned to construct an accurate and reliable historical net load time series.

[0025] The preprocessing performed includes missing value imputation, outlier removal, noise reduction, and smoothing. The core purpose of preprocessing is to purify the raw data, improve data quality, and provide a reliable, consistent, and physically meaningful data foundation for subsequent decomposition, prediction, and scenario generation. Preprocessing ensures data integrity, guarantees data authenticity, and extracts valid signals. Data that has undergone rigorous preprocessing can more realistically reflect the objective laws of power distribution network source-load interaction, thereby significantly improving the scientific rigor and reliability of the final planning scheme.

[0026] The constructed historical net load time series can be represented as: ; in, For a moment Historical net load time series; For a moment The load power; For a moment The output power of distributed power sources.

[0027] S13, The variational mode decomposition algorithm is used to decompose the historical net load time series into multiple eigenmode functions with different center frequencies; The variational mode decomposition (VMD) algorithm was used to analyze the historical net load time series. Decompose to obtain Each with a different center frequency eigenmode functions As a modal component, where , The value of can be 5 to 6, with 6 being the preferred value. To achieve the above decomposition, the constructed optimization problem expression is as follows: ; In this expression, The imaginary unit; Indicates time; For a moment Dirac distribution over time; For the first The center frequency of each mode; For the first One intrinsic mode function.

[0028] Among them, the above decomposition is performed using the variational mode decomposition algorithm. Figure 2 The historical load VMD decomposition process is detailed in the relevant technical explanations and will not be elaborated here.

[0029] S14. Based on the magnitude of multiple center frequencies and two preset frequency thresholds, the multiple intrinsic mode functions are divided into low-frequency trend terms, mid-frequency periodic terms, and high-frequency random terms.

[0030] After obtaining each modal component, it is divided into low-frequency trend terms based on the magnitude of the center frequency. Mid-frequency periodic term and high-frequency random terms ,correspond Figure 3 The components are "low-frequency trend component", "medium-frequency periodic component" and "high-frequency residual component".

[0031] In one alternative implementation, two frequency thresholds are set. and ,and And divided according to the following rules: when When, it is classified as a low-frequency trend term; when When, it is divided into intermediate frequency periodic terms; when When, it is divided into high-frequency random items.

[0032] Among them, low-frequency trend items Reflects the long-term growth or decline trend of net load; mid-frequency periodic term. Reflecting intraday, weekly, or seasonal cyclical fluctuations; high-frequency random terms It reflects short-term random disturbances and residual characteristics.

[0033] For step S2, future trend prediction is performed on the low-frequency trend term to obtain the low-frequency trend prediction component; the mid-frequency cycle prediction component is obtained based on the historical cycle structure of the mid-frequency cycle term; the high-frequency random term is reconstructed by combining the conditional diffusion model to obtain the high-frequency random prediction component; based on the three prediction components obtained, the typical daily scene set and corresponding weights for the future target year are obtained through superposition clustering; the future target year is the year following the latest year in the historical source load data; Step S2 corresponds to Figure 2 The trend prediction, time-series load reconstruction, and typical daily extraction of K-Means are included in this part.

[0034] In one optional implementation, in step S2, future trend prediction is performed on the low-frequency trend term to obtain a low-frequency trend prediction component; a mid-frequency cycle prediction component is obtained based on the historical cycle structure of the mid-frequency cycle term; and the high-frequency random term is reconstructed using a conditional diffusion model to obtain a high-frequency random prediction component. This may include the following steps a1 to a3: Step a1: The low-frequency trend term is predicted using a method combining linear extrapolation and logistic regression to obtain the low-frequency trend prediction component for the future target year. like Figure 3 As shown on the right, for the low-frequency trend term A method combining linear extrapolation and logistic regression is used to predict future trends, simulating potential load growth caused by factors such as increased electric vehicle penetration and changes in the scale of distributed energy resources. This yields a low-frequency trend prediction component for the 6th year (the future target year). .

[0035] This step utilizes existing linear extrapolation combined with logistic regression as a predictive technique to quantify and extrapolate the long-term evolution patterns implicit in the low-frequency trend term of historical data into the future. In short, the low-frequency trend term is a smooth time series that has been freed from periodicity and random fluctuations, primarily reflecting long-term, slow changes. First, the shape of this low-frequency trend term is analyzed, and linear extrapolation, a simple and direct trend extension method, is used to capture and extend the explicit, approximately linear growth portion of the series. Logistic regression modeling is then combined to more precisely characterize and predict the non-linear growth with saturation characteristics driven by factors such as increased electric vehicle penetration and changes in the scale of distributed energy. The combination can be achieved by, for example, using a logistic regression model to fit the historical trend to determine the shape and saturation point of the growth curve, and then extrapolating the trend determined by the model; or by using more suitable methods for different stages of the series. Finally, the low-frequency trend prediction component for the 6th year (the future target year) is obtained. This is a time series of 8760 hours (one year). It no longer contains any periodicity or random fluctuations, but purely represents the net load baseline level or "baseline" determined by long-term, slowly changing driving factors (such as the slowly increasing number of electric vehicles and the gradually increasing installed capacity of new energy vehicles) in the future target year.

[0036] For details on the specific processing steps of step a1, please refer to relevant technical explanations; further details will not be provided here.

[0037] Step a2: Retain the historical periodic structure of the intermediate frequency periodic term, and perform periodic mapping and phase alignment according to the calendar information of the future target year to obtain the intermediate frequency periodic prediction component of the future target year; This step involves the intermediate frequency periodic term. Preserving its historical periodic structure, and performing periodic mapping and phase alignment according to the calendar information of the 6th year in the future, the mid-frequency periodic component of the 6th year in the future is obtained. .

[0038] Mid-frequency periodic term It primarily reflects the regular and repetitive fluctuation patterns in net load. It can include various cycles; for example, the intraday cycle represents the daily load peak-to-valley variations (such as morning and evening peaks). The weekly cycle reflects the significant differences in load patterns between weekdays and weekends. The seasonal cycle reflects the differences in electricity consumption patterns caused by variations in temperature and sunshine duration across different seasons.

[0039] During processing, preserving the historical periodic structure means not altering the fluctuation patterns, amplitudes, and relative phase relationships inherent in the historical mid-frequency periodic terms. Period mapping and phase alignment involve applying the preserved historical periodic structure to the calendar timeline of the future target year (e.g., the 6th year). Mapping refers to determining the corresponding future time points for historical periodic patterns. For example, mapping all historical "Monday" periodic fluctuation patterns to the dates of all "Mondays" in the future target year. Alignment ensures that the starting point and phase of the periodic fluctuations precisely match the actual future time.

[0040] The result obtained from the processing is the mid-frequency periodic component of the future target year. This is a time series of 8760 hours (one year) in length. It contains all deterministic cyclical fluctuations that can be extrapolated from historical patterns in the coming year, but does not include long-term growth trends and short-term random noise.

[0041] Step a3: Using the high-frequency random terms as training data, train a one-dimensional conditional diffusion model; the conditional diffusion model uses time feature information as a conditional vector, and the time feature information includes at least month, weekday, hour and workday identifiers; based on the trained conditional diffusion model and the time feature information of the future target year, generate a high-frequency random prediction component corresponding to the future target year.

[0042] Step a3 is high-fidelity future scene generation based on the conditional diffusion model. This part corresponds to... Figure 3 diffusion models and Figure 5 The diagram shown illustrates the principle of random component reconstruction based on the diffusion model.

[0043] It is understandable that the high-frequency random term is the complete high-frequency residual sequence obtained after VMD decomposition.

[0044] The obtained high-frequency random items As the training object for the conditional diffusion model. To facilitate model training, high-frequency random terms are first used... The data is sliced ​​according to a preset time window, and the extracted single sample is denoted as the original random item sample. , to be used as training samples.

[0045] In one specific implementation, high-frequency random items can be sliced ​​according to time windows of 24 hours, 168 hours, or 720 hours to form a sample set, represented as follows: ; in, For the sample set; M The number of samples in the sample set; For the first m A sample of original random items, This is the condition vector corresponding to the sample. The condition vector includes at least one or more of the following: month information, weekday information, hour information, weekday / non-weekday identifier, low-frequency trend forecast value, mid-frequency periodic value, and meteorological related features. The condition vector is used to constrain the model to generate high-frequency random fluctuations that match the future time context.

[0046] A one-dimensional conditional diffusion model based on the U-Net architecture is constructed to learn the statistical distribution characteristics of high-frequency random terms. The one-dimensional conditional diffusion model is a complete generative model framework for generating high-frequency random components. The diffusion model includes two processes: forward diffusion (noise addition) and backward denoising (generation).

[0047] Its forward diffusion process is as follows Figure 5 The distribution gradually diffuses from "original high-frequency components" to "Gaussian noise." During the training phase, the original random item samples... Gaussian noise is gradually added, and the first step is defined using the backsampling formula. Step state for: ; in, For the first The variance scheduling parameter of step 1 represents the variance scheduling parameter of step 2. The noise intensity of the step satisfies: . It is the identity matrix; This indicates that the forward diffusion process is in the first stage. Step to the first The conditional probability distribution of the step; This indicates a Gaussian distribution.

[0048] In a preferred embodiment, the total number of diffusion steps Set to 1000, variance scheduling parameter A linear scheduling method is adopted, with values ​​ranging from 0.0001 to 0.02, i.e.: The intermediate steps are determined by linear interpolation, as expressed by the formula: ; Furthermore, the forward diffusion process can be written as: ; in, , It is the first The proportion of the original signal retained in the step. From step 0 to step 1 The proportion of original samples retained in the first step. During the reverse denoising training phase, a conditional U-Net network is constructed. It is specifically responsible for the reverse denoising process and is the core neural network component in this one-dimensional conditional diffusion model, namely the denoising network. Specifically, it is used to denoise given noisy samples. diffusion steps and condition vector Predicting noise in the case of Its loss function is: ; in, Indicates network parameters; Represents the mathematical expectation; This represents the square of the L2 norm.

[0049] The training process of the one-dimensional conditional diffusion model is shown in the following steps: 1) From high-frequency random items The original random item samples are extracted by slicing the data according to a preset time window. ; 2) Construct a corresponding condition vector for each original random item sample. ; 3) Number of random sampling diffusion steps ; 4) Scheduling based on variance parameters and Gaussian noise , the original random item sample Diffusion ; 5) Input conditional U-Net network, output noise prediction value ; 6) Update network parameters based on the mean square error between the actual noise and the predicted noise. ; 7) Repeat the above steps until the loss function converges to obtain the trained conditional diffusion model.

[0050] When generating the 8760-hour high-frequency random components for the 6th year, we first construct the conditional sequences corresponding to the 8760 times based on the full-year calendar information and trend prediction results for the 6th year. (i.e., condition vector) The conditional sequence includes at least the month, weekday, hour, and workday attributes corresponding to each future time point, as well as the corresponding values ​​in the low-frequency trend prediction component and the mid-frequency period prediction component.

[0051] Then, with Gaussian white noise sequence As the initial input, according to The reverse denoising process is executed in reverse order, and the current state is changed at each step. Current number of steps and corresponding condition vector Input the trained conditional U-Net network to obtain the noise estimate for the current step, and gradually recover the state from the previous time step according to the backsampling formula. .when At that time, the generated result is obtained. That is, the high-frequency random prediction component sample of the future.

[0052] In one specific implementation, the 8760-hour high-frequency random prediction component for the entire sixth year can be obtained by generating and stitching the components daily. In another implementation, the 8760-hour high-frequency random prediction component for the entire sixth year can be obtained by continuously generating the components using a sliding time window and smoothly stitching them together in overlapping intervals. The final result is a sequence of high-frequency random prediction components for the sixth year, i.e., the high-frequency random prediction component sequence. .

[0053] For the handling of low-frequency trend terms, mid-frequency periodic terms, and high-frequency random terms, please refer to [link to documentation]. Figure 3 The diagram illustrates the evolution characteristics of power distribution network sources and loads and the data-driven, forward-looking scenario generation mechanism.

[0054] In one optional implementation, in step S2, based on the obtained three prediction components, a set of typical daily scenarios and their corresponding weights for the future target year is obtained through superposition clustering. This may include the following steps b1~b2: Step b1: The low-frequency trend prediction component, the mid-frequency periodic prediction component, and the high-frequency random prediction component are superimposed hourly to reconstruct a forward-looking net load sequence for the future target year. Step b2: Perform K-Means clustering on the prospective net load sequence to extract multiple typical daily scenarios, and calculate the occurrence weight of each typical daily scenario throughout the year, thereby obtaining the set of typical daily scenarios and their corresponding weights for the future target year.

[0055] like Figure 3 As shown in the "Reconstruction" section, the low-frequency trend prediction component for the 6th year in the future will be... Mid-frequency periodic prediction component and high-frequency random prediction components By overlaying the data hourly, a forward-looking net load sequence for the next 6 years is obtained, represented as follows: ; Subsequently, the reconstructed forward-looking net load sequence Clustering is performed using the K-Means algorithm, and the output is... Extracted from A typical daily scene set and its corresponding weights This serves as the input for subsequent planning models.

[0056] Among them, multiple typical daily scenarios constitute a typical daily scenario set, which refers to a limited number of scenarios selected, clustered, and summarized from the 8760-hour forward-looking net load sequence throughout the year. indivual, A representative 24-hour net load curve (a natural number greater than 0) is used to characterize different types of daily load operation patterns in the future. Specifically, the K-Means algorithm groups 8760 "points" throughout the year (each "point" is actually a 24-hour curve) based on their shape similarity. The final center of each group (or the most representative curve) is defined as a "typical daily scenario." Directly using the 8760-hour sequence for planning would result in a massive model size and computational infeasibility. This embodiment of the invention extracts a set of typical daily scenarios, using the operating status of a few representative days to approximate the complex changes throughout the year. This significantly reduces the computational complexity and time cost of the subsequent bi-level planning model while preserving the main characteristics of future loads.

[0057] Weight Indicates the first The probability or representative percentage of a typical daily scenario throughout the year (here) This represents the sequence number of a typical daily scene, and its value range is... It quantifies the first The weights represent the "number of days" or "time proportion" of a typical daily scene within the 8760 hours of a year. For example, in K-Means clustering, the algorithm assigns each day of the year (each 24-hour curve) to the category of the most similar typical daily scene. That is equivalent to being classified as the first The proportion of "days" under each typical daily scenario category to the total number of days in the year. For example, if 120 days in a year are classified as "Typical Daily Scenario A," then the weight of this scenario is approximately 120 / 365 ≈ 0.329. This weighting ensures that in subsequent planning model evaluations, the importance of different typical daily scenarios is proportional to their actual frequency of occurrence throughout the year. Scenarios with high frequency (high weight) have a greater impact on planning objectives (such as total investment cost, cumulative voltage deviation); scenarios with low frequency (low weight) have a smaller impact. This makes the planning scheme not targeted at a few special days, but rather a statistically representative optimization of the entire year's operational status.

[0058] The above process corresponds to Figure 2 The term "Time-series load reconstruction and K-Means typical day extraction" is used in this context.

[0059] For step S3, a two-layer optimization model for collaborative planning is constructed using the typical daily scenario set and corresponding weights as input. The upper-layer model is used to decide on the site selection and capacity determination scheme of energy routers and distributed energy storage, with the goal of minimizing the annualized investment cost and the cumulative voltage deviation of the system. The lower-layer model is used to perform distribution network time-series operation simulation under the given upper-layer scheme, and feeds back the simulation results to the upper layer for scheme evaluation. Step S3 involves constructing the upper-level decision model (hereinafter referred to as the upper-level model) and the lower-level operational optimization model (hereinafter referred to as the lower-level model) of the two-level optimization model.

[0060] The following will explain each point separately.

[0061] (a) Construction of the upper-level model: This step corresponds to Figure 4 The planning layer logic in the previous step uses the typical daily scene set and its weights as input to the upper-level model.

[0062] In one optional implementation, the construction process of the upper-level model in the two-layer optimization model may include the following steps c1 to c5: Step c1: Using the typical daily scene set and corresponding weights as input to the upper-level model, define the upper-level decision vector as follows: ; in, and Distributed energy storage and energy routers at the nodes, respectively. The addressing Boolean variable is usually represented by 1 to indicate "yes" (i.e., the device is installed on this node) and 0 to indicate "no" (i.e., the device is not installed on this node). and These are the rated power and rated energy storage capacity of distributed energy storage, respectively. The rated switching power of the energy router; a set of decision values ​​of the upper-layer decision vector constitutes an addressing and capacity determination scheme; Step c2, establish the first objective function at the upper level, with the goal of minimizing the annualized investment cost, expressed as: ; in, This is the first objective function of the upper layer; This is the annualized coefficient. , These are the unit power cost and unit capacity cost of distributed energy storage, respectively. Cost per unit power of the energy router; This represents the set of nodes for Distributed Energy Storage (DES). This represents the set of candidate installation nodes for the energy router; Step c3: Establish the upper-level second objective function, with the goal of minimizing the cumulative voltage deviation of the system, expressed as: ; in, This is the second objective function of the upper layer; Typical daytime scene Next node At any moment The per-unit voltage value is fed back from the lower-level model. This is the per-unit value of the rated voltage, which is a fixed value. This represents the number of intrinsic mode functions with different center frequencies obtained by decomposing the historical net load time series; Indicates the first Weighting of a typical daily scenario; This represents a set of discrete time periods within a single typical day scenario, preferably a set of time periods corresponding to 24 hours. This represents the set of all nodes in the target distribution network; and All a subset of and They can be the same or different; This indicates the calculation of absolute value; Step c4: Set upper-level constraints related to the upper-level decision vector, including at least constraints on the number of equipment locations, equipment port configuration, and upper and lower limits on the rated power and capacity of the equipment, to ensure that the planning scheme meets the requirements of engineering feasibility; wherein: The constraints on the number of equipment locations are as follows: ; Device port configuration constraints are as follows: ; The upper and lower limits of the equipment's rated power and rated capacity are as follows: ; in: Indicates candidate sites Should charging stations be built? This indicates the construction of charging stations. This indicates that no charging stations will be built; The maximum number of sites allowed to be built; Indicates site The number of charging ports configured; Represents the field of real numbers greater than or equal to 0; Indicates site The planned capacity; Represents the field of real numbers greater than or equal to 0; This represents the total number of candidate sites, which is the set of all nodes in the target distribution network; For the site The minimum number of ports that should be configured after the website is built; For the site Configurable maximum number of ports; For the site The minimum capacity allowed after construction; For the site The maximum capacity that can be configured after construction.

[0063] Step c5: The upper-level model is composed of the upper-level decision vector, the upper-level first objective function, the upper-level second objective function, and the upper-level constraints.

[0064] (II) Construction of the lower-level model: This step corresponds to Figure 4 The lower-level model's inputs include the equipment location and capacity determination results, typical daily scenario net load curves, and equipment operating boundaries (equipment operating boundaries are the state-of-charge evolution constraints of distributed energy storage and the cross-node energy conservation constraints of energy routers mentioned later); the outputs include node voltage, branch power flow, energy storage SOC trajectory, and EER power exchange trajectory for each scenario.

[0065] In one optional implementation, the construction process of the lower-level model in the two-layer optimization model may include the following steps d1 to d4: Step d1: Obtain the site selection and capacity determination schemes decided by the upper-layer model, as well as the typical daily scene set; As mentioned earlier, a set of decision values ​​from the upper-level decision vector constitutes a location and capacity allocation scheme; the upper-level model can determine multiple location and capacity allocation schemes. However, during the execution of the lower-level model, each time a location and capacity allocation scheme is independently called and processed, this is a specific candidate scheme that needs to be evaluated. The task of the lower-level model is to perform a full-scenario time-series simulation of this specific candidate scheme and calculate its performance indicators.

[0066] Step d2: Using the LinDistflow power flow model, construct the distribution network time-series operation simulation architecture and set operation constraints. The operation constraints include: power flow balance constraints, voltage drop constraints, state of charge evolution constraints of distributed energy storage, and cross-node energy conservation constraints of energy routers. The power flow balance constraint is expressed as:

[0067] in, Represents any node in the target distribution network; Represented by node For all branches at the power outflow end End node A set; Indicates at time From node Flow to Node branch road Active power; Represented by node For all branches at the power inflow end The first node A set; and These are all mathematical set symbols; Indicates at time From node Flow to Node branch road Active power; Indicates a branch The resistance; Indicates at time Flowing through a branch road The current; This indicates the time in the historical load data. Access Node Total load power; This indicates the output data of distributed power sources at time [time]. Access Node The output power of distributed photovoltaic power sources; Indicates at time Installed on node The charging power of distributed energy storage; Indicates at time Installed on node The discharge power of distributed energy storage; Indicates at time Installed on node The active power exchanged between the energy router (EER) and the power grid is positive when injected into the grid and negative when absorbed from the grid. The voltage drop constraint is expressed as follows: ; in, Indicates at time ,node The per-unit voltage value; Indicates at time ,node The per-unit voltage value; Indicates a branch The resistance; Indicates a branch The reactance; Indicates at time From node Flow to Node branch road reactive power; The charge state evolution constraint of the distributed energy storage is expressed as: ; in, Indicates that it is installed on the node Distributed energy storage at any time The state of charge; Indicates that it is installed on the node Distributed energy storage at any time The state of charge; This indicates the charging efficiency of distributed energy storage; Indicates at time Installed on node The charging power of distributed energy storage; Indicates the simulation time step; Indicates that it is installed on the node The rated energy storage capacity of distributed energy storage; Indicates at time Installed on node The discharge power of distributed energy storage; This indicates the discharge efficiency of distributed energy storage; The cross-node energy conservation constraint of the energy router is expressed as: ; in, This indicates the injected power of the EER at different ports; Indicates at time The energy router from its port Active power injected into the power grid; This represents the power absorbed by the EER at different ports; Indicates at time The energy router from its port Active power absorbed from the power grid; This represents the set of all nodes in the target distribution network; This represents the conversion efficiency of the energy router.

[0068] Step d3: For each typical day scenario in the typical day scenario set, under the conditions of the candidate site selection and capacity setting scheme, perform 24-hour time-series operation simulation calculation to obtain the per-unit voltage value of all nodes of the target distribution network, the power flow of all branches, the state of charge and charging / discharging power of each distributed energy storage, and the port switching power of each energy router in each time period under the typical day scenario. This step involves optimizing the operating power of EER and DES at different times based on the LinDistflow power flow model for each typical daily scenario, given the upper-level planning results, and solving for node voltage, branch power flow, and equipment output.

[0069] Specifically, all parameters in a given candidate site selection and capacity sizing scheme provided by the upper-level model are taken as fixed values. For the typical daily scenario currently being processed, its 24-hour net load curve is used as known time-series data input. Based on the LinDistflow power flow model constructed in step d2 and all operational constraints, the goal is to find a feasible, and usually optimal, operational strategy by solving for the variables. , , , , The optimal value is obtained. In actual solutions, mathematical programming solvers (such as CPLEX, Gurobi) or specially designed optimization algorithms are used for calculation. Please refer to the relevant technologies for details, which will not be elaborated here.

[0070] The output specifically includes: This typical daytime scene Per-unit voltage values ​​of all nodes in the target distribution network during each time period , representing all nodes At all times The voltage; All branch roads At all times trend As mentioned above, Indicates at time Installed on node The active power injected into the grid by the DES; Distributed energy storage at all times State of charge As mentioned above, Indicates that it is installed on the node Distributed energy storage at any time The state of charge; Distributed energy storage at all times charging and discharging power As mentioned above, Indicates at time Installed on node The active power injected into the grid by the distributed energy storage (DES); Each energy router at all times port switching power As mentioned above, Indicates at time Installed on node The energy router (EER) exchanges active power with the power grid.

[0071] Step d4: The per-unit voltage values ​​of each node obtained from the time-series simulation calculation are fed back to the upper-level model as input data for calculating the objective function of the system's cumulative voltage deviation, so as to evaluate the operating performance of the location and capacity setting scheme.

[0072] The node voltages (per-unit voltage values) obtained by the lower-level model are fed back to the upper-level model as feedback to evaluate the performance of different planning schemes in typical future scenarios. The remaining outputs of the lower-level model serve as scheduling instructions for the corresponding scenarios, which can be retained for interpretability analysis or to verify the operational results of the planning on typical days.

[0073] Specifically, the objective function for the system's cumulative voltage deviation is the upper-level second objective function. , which contains It is obtained from feedback from the lower-level model, and the upper-level model receives the feedback. Substitute the pre-established second objective function The formula is used to calculate and obtain a specific result. The output value represents the objective function value of the system's cumulative voltage deviation. This value is a key performance evaluation indicator, quantitatively reflecting the severity of the overall voltage deviation caused by the currently evaluated site selection and capacity designation scheme under various typical future operating scenarios. The smaller the output value, the better the improvement in distribution network voltage quality and the superior operational performance of the proposed scheme. This indicator is related to the investment cost target. Together, they constitute the dual criteria for evaluating the merits of planning schemes, thereby guiding optimization algorithms (such as the Riemannian manifold evolution algorithm) to find collaborative planning schemes that are more economical in terms of investment and have better power quality.

[0074] For step S4, the Riemannian manifold evolution algorithm is used to solve the two-level optimization model and output the Pareto optimal planning scheme set; Step S4 is implemented using the Riemannian Manifold Evolution Algorithm (RMEA). Corresponding to... Figure 2 The RMEA solution process in the middle, namely Figure 2 The process within the dashed box.

[0075] In one optional implementation, S4 may include the following steps: S41, the solution problem of the two-level optimization model is mapped to a statistical manifold composed of a family of probability distributions, and the weight vector of the multi-objective decomposition and the distribution parameters defining the probability distribution of candidate solutions are initialized to construct the initial search state of the algorithm; wherein, the statistical manifold ; It is the upper-level decision vector that serves as the input to the upper-level model, representing the candidate planning schemes for the location and capacity determination of energy routers and distributed energy storage. Represents the probability density function; These are distribution parameters; Represents the parameter space; Step S41 is the preparation and initialization stage of the entire Riemannian manifold evolution algorithm. It mainly involves initializing the weight vector and distribution parameters to construct the initial Riemannian metric field. Step S41 corresponds to... Figure 2 The first box in the dashed box on the right is “Sampling of Planning Schemes Based on Probabilistic Manifolds” and the second box is “DistFlow Simulation and Target Evaluation in Typical Daily Scenarios”.

[0076] In general, the solution implementation based on the Riemannian manifold evolution algorithm (RMEA) first samples planning schemes based on probabilistic manifolds, evaluates these planning schemes under typical daily scenarios, constructs the Fisher information matrix based on the evaluation results, and calculates the natural gradient. Then, iteratively corrects the momentum of the parallel shift operator and updates the local metrics, calculates the evolutionary state entropy to adjust the step size, and finally updates the population distribution parameters and re-evaluates.

[0077] Here, the family of probability distributions refers to a family of probabilistic models that can be used to generate different candidate planning solutions. Each specific probability model (such as a multivariate Gaussian distribution) can generate various solutions with a certain probability. The statistical manifold is the space or set of all these possible probability models. The algorithm no longer directly operates on individual solutions, but instead searches for an optimal model in this space of probability models, ensuring that this model generates good solutions with a high probability. The concept of weight vectors here comes from the multi-objective evolutionary algorithm decomposition framework, and is relevant to the bi-objective problem of this invention (minimizing investment cost). and voltage deviation In this problem, a set of uniformly distributed weight vectors is used to decompose the original problem into multiple single-objective subproblems. Each subproblem is... and A specific weighted combination represents a different cost-quality trade-off preference. Initializing the weight vector prepares for parallel searching of a set of solutions covering the entire range of trade-offs.

[0078] The distribution parameters are the core decision variables of the RMEA algorithm. Taking the most common multivariate Gaussian distribution as an example, its distribution parameters... Typically includes a mean vector and a covariance matrix. Initialization This means setting an initial search distribution for the algorithm.

[0079] Initialize distribution parameters This is to define a probability distribution from which to sample and generate an initial set of candidate solutions, i.e., the initial population. During the algorithm's iteration process, each step is based on the updated distribution parameters. A new population is generated, or an existing population is screened and recombined. This constantly updated set of schemes is called the current population in the current iteration step.

[0080] After initialization, an iterative loop is performed. S41 to S45 constitute one iterative loop.

[0081] S42, based on the initialized distribution parameters, calculate the Fisher information matrix of the current probability distribution on the statistical manifold, and calculate the natural gradient based on the Fisher information matrix to guide the update of the search direction; wherein, the formula for calculating the Fisher information matrix is: ; The formula for calculating the natural gradient is: ; Wherein, step S42 corresponds to Figure 2 The third box in the dashed box on the right is titled "Fisher Information Matrix Construction and Natural Gradient Calculation". Represents the first element in the Fisher information matrix. u Okay, number v Column elements; For mathematical expectation operators; Indicates the first u One distribution parameter; Indicates the first v One distribution parameter; Indicates the ordinary gradient; This represents the inverse matrix of the Fisher information matrix; Represents the natural gradient; S43, a Riemann momentum term is introduced to iteratively update the natural gradient, and the update formula is as follows: ; Step S43 corresponds to Figure 2 The fourth box in the dashed box on the right is "Parallel translation operator momentum correction and local metric update".

[0082] in, Indicates the first The Riemann momentum term at the next iteration; Indicates the momentum decay coefficient; The covariance matrix is ​​used to characterize the correlation between decision variables. The covariance matrix is ​​calculated from the sample dispersion of each planning variable in the current population and is used to characterize the correlation between different node location variables, power capacity variables and electrical coupling relationships caused by power flow constraints. S44, monitoring the evolutionary state entropy during algorithm iteration. ,when Lower than the preset reference entropy At this time, the search step size is adaptively increased to encourage the algorithm to escape the local optimum. Step S44 corresponds to Figure 2 The fifth box in the dashed box on the right is "Evolutionary State Entropy Calculation and Step Size Negative Feedback Adjustment".

[0083] In evolutionary computation, entropy is a commonly used metric to measure the diversity or dispersion of a population. Evolutionary state entropy. Quantified in the first In the next iteration, the distribution of all candidate planning schemes in the current population within the decision space (i.e., the space composed of various site selection and occupancy schemes) is discrete. A high entropy value indicates large individual differences and good diversity within the population; a low entropy value indicates convergence and poor diversity among individuals. A preset reference entropy is used. The value can be set according to your needs or based on experience. For example, under normalized caliber, you can choose 0.6.

[0084] This step provides an adaptive adjustment mechanism to maintain population diversity and prevent the algorithm from prematurely converging to a local optimum. Understandably, when entropy decreases (population diversity declines), increasing the step size can cause newly generated candidate solutions to deviate more significantly from the current population center, thus helping the algorithm escape the current local search region, explore a wider unknown portion of the solution space, and increase the likelihood of finding a globally better Pareto front.

[0085] S45. Based on the initialized weight vector, the upper-level multi-objective problem of the two-level optimization model is decomposed into multiple scalar quantum problems and solved in parallel using a multi-objective evolutionary algorithm decomposition framework. Iterative optimization is then performed, and finally, a Pareto optimal programming scheme set composed of non-dominated solutions is output.

[0086] Step S45 corresponds to Figure 2 The last box in the dashed frame on the right is labeled "Population Distribution Parameter Update." This step, combined with the MOEA / D framework, decomposes the upper-level multi-objective problem into multiple scalar subproblems, searches for the optimal solutions to each subproblem in parallel, and ultimately obtains a Pareto-optimal programming scheme set composed of non-dominated programming schemes. The Pareto-optimal programming scheme set refers to a group of EER and DES location and capacity determination schemes where there is no mutual dominance between the two objectives of annualized investment cost and system voltage deviation. Each scheme corresponds to a set of equipment configurations and their operating results that satisfy the constraints.

[0087] The processing in S45 mainly involves: firstly, problem decomposition. Specifically, based on a set of uniformly distributed weight vectors obtained in step S41, the bi-objective optimization problem of the upper-level model is decomposed into multiple scalar optimization sub-problems. Each weight vector corresponds to a sub-problem, representing the degree of preference of that sub-problem for the two optimization objectives; different weight vectors have different values, thus corresponding to different objective trade-offs. In other words, each sub-problem is essentially a single-objective optimization problem formed by weighted scalarization of the bi-objective function under the action of the corresponding weight vector, thereby enabling multiple sub-problems to search different regions on the Pareto front.

[0088] Then, candidate solutions are generated or updated in parallel. Specifically, for each of the multiple subproblems, the algorithm maintains a current candidate solution or the current optimal candidate solution corresponding to that subproblem. In the current iteration, step S45 uses the updated search direction, momentum information, and step size information output from previous steps S42 to S44 to iteratively update the candidate solutions corresponding to each subproblem, thereby generating new candidate solutions. Since different subproblems correspond to different weight vectors, the generated new candidate solutions will reflect different optimization emphases in the target space, thereby driving the overall population to conduct a distributed search along the Pareto front.

[0089] Finally, evaluation and updates are performed. Specifically, for each newly generated candidate solution in a subproblem, the lower-level model is invoked to perform time-series simulations, and its two objective functions are calculated. and The value is then determined. This new solution is then compared and competed with the current solution for the subproblem and the solutions for neighboring subproblems, updating the currently maintained optimal solutions for the subproblem and its neighboring subproblems.

[0090] After the current iteration completes one S45 step, we obtain an updated "new generation population" consisting of all candidate solutions corresponding to multiple subproblems. This population is expected to be closer to the Pareto optimal frontier overall compared to the previous generation population.

[0091] Those skilled in the art will understand that S41 constructs the initial population. Each S42-S45 iteration is one iteration cycle. This process is repeated, and the population evolves generation by generation under the geometric guidance of RMEA and the decompositional cooperation of MOEA / D. When a preset termination condition is met (such as reaching the maximum number of iterations), the algorithm stops iterating. At this point, the final generation of the population contains a large number of high-quality candidate solutions found by the algorithm. From all the solutions in this final generation, all non-dominated solutions are selected. The set of these non-dominated solutions is the final Pareto optimal programming solution set. Each solution represents a defined set of EER and DES location and capacity configurations and their corresponding performance metrics. , ).

[0092] S5. The final planning scheme is determined by evaluating the Pareto optimal planning scheme set.

[0093] Step S5 involves outputting the proposed solution and conducting a multi-dimensional evaluation.

[0094] In one optional implementation, S5 may include: The grey relational analysis method is used to determine the final planning scheme from the Pareto optimal planning scheme set.

[0095] Specifically, this implementation method may include the following steps: S51, for each planning scheme in the Pareto optimal planning scheme set, construct an evaluation index set; the evaluation index set includes at least the initial investment cost, system voltage offset, and redundancy for adapting to future load growth. This step yields a set of Pareto optimal programming solutions, where each solution contains its... , Specific numerical values ​​are required. An evaluation vector is constructed for each scheme, which includes at least the initial investment cost, system voltage offset, and redundancy for adapting to future load growth.

[0096] The initial investment cost is based on the first objective function of the upper level. The value is obtained from the capacity. The unit price is derived from the initial investment cost. For details, please refer to the relevant technical understanding in this field, which will not be explained in detail here. The system voltage offset is the upper-level secondary objective function The value; The redundancy for adapting to future load growth is calculated as follows: ; in, Redundancy to accommodate future load growth; The planned available capacity; For current or predicted load demand.

[0097] Therefore, each scheme will obtain a set of evaluation index values, which includes the above three evaluation indexes.

[0098] S52, Dimensionless processing is performed on each evaluation index in the evaluation index set to obtain normalized index values; Understandably, since the dimensions and orders of magnitude of each indicator are different, normalization is required to eliminate the influence of dimensions and convert all indicator values ​​to [0,1] or a comparable range.

[0099] S53, determine a reference sequence consisting of the optimal values ​​of each evaluation index; This step involves selecting the optimal value for each evaluation index from all options and combining these optimal values ​​into a virtual, ideal sequence of options, called the reference sequence.

[0100] S54. For each planning scheme, calculate the grey relational coefficient between its normalized evaluation index values ​​and the corresponding index values ​​in the reference sequence. Specifically, for each scheme, the distance or difference between the normalized value of each evaluation indicator of that scheme and the corresponding indicator value in the reference sequence is calculated. This difference is then substituted into the grey relational coefficient formula to obtain a value between 0 and 1. The closer this value is to 1, the closer the scheme is to the ideal value for that indicator. Therefore, the grey relational coefficient matrix for each scheme with respect to each evaluation indicator can be obtained.

[0101] S55. Based on the preset weights of each evaluation index, calculate the weighted sum of the grey relational coefficients of all evaluation indexes for each planning scheme to obtain the grey relational degree of the planning scheme. For each scheme, the grey relational coefficients of each indicator are weighted and summed. This weighted sum is the comprehensive evaluation score of the scheme, called the grey relational degree.

[0102] Therefore, each solution can obtain a corresponding gray relational degree, which is a numerical value.

[0103] S56. From the Pareto optimal planning scheme set, select the planning scheme with the largest grey relational degree as the final planning scheme to represent the final site selection and sizing plan.

[0104] This step compares the grey relational degree of all options and selects the option with the highest value. This option means that, under the given index and weight system, its overall performance is closest to the ideal reference option.

[0105] Those skilled in the art will understand that the process described in S51-S56 above is a typical multi-attribute decision-making process. It compares the solutions in the Pareto solution set with a reference sequence serving as the ideal point, and comprehensively considers the importance (i.e., weight) of each indicator. Ultimately, it quantitatively selects the solution with the highest grey relational degree as the final EER and DES location and capacity planning scheme. This is a planning scheme that is most balanced in terms of investment, voltage quality, and development adaptability, and best suits the decision-maker's preferences, providing planners with a clear basis for decision-making.

[0106] The final output results of this invention embodiment include: 1) EER installation nodes and rated switching capacity; The installation node of EER is explained in step c1, and it is determined by the upper-level decision vector. express, Indicates that the energy router is at the node The location selection Boolean variable; in the final determined planning scheme, The node is the installation node of EER.

[0107] The rated switching capacity of the EER is determined by the upper-level decision vector. express, This refers to the rated switching power of the energy router. In the finalized planning scheme, for... The node, its The specific value is the rated switching capacity of the node.

[0108] 2) Installation nodes, rated power, and rated capacity of DES; The installation nodes of DES are determined by the upper-level decision vector. express; For distributed energy storage at nodes The location selection Boolean variable. In the final determined planning scheme, The node is the DES installation node.

[0109] The rated power and rated capacity of DES are respectively determined by the upper-level decision vector. and express, and These refer to the rated power and rated capacity of distributed energy storage, respectively; in the finalized planning scheme, for The specific values ​​of these two variables are the rated power and rated capacity of the DES at the node.

[0110] 3) Voltage improvement effect in typical scenarios; It is based on the upper-level second objective function The calculated voltage improvement effect in typical scenarios corresponds to the system voltage offset mentioned earlier. 4) The ability to adapt to potential future load growth.

[0111] The ability to adapt to potential future load growth is what was mentioned earlier as the redundancy for adapting to future load growth.

[0112] It should be noted that, under non-leap year conditions, the forward net load series for the 6th year in the future is constructed as 8760 hours; if the 6th year in the future is a leap year, it is expanded to an 8784-hour series.

[0113] The present invention has the following beneficial effects: Proactive source-load scenario representation capability. Compared with existing technologies, this invention utilizes VMD deconstruction and conditional diffusion models to generate high-fidelity scenarios that include potential load surge trends and retain stochastic fluctuation characteristics for the next 5-10 years. This allows the planning scheme to go beyond simply meeting the current grid conditions, but also possesses redundancy to cope with the large-scale integration of new loads such as electric vehicles in the future, significantly enhancing the engineering robustness of the planning scheme.

[0114] The invention leverages the spatiotemporal complementarity of resource utilization. In its planning scheme description, the invention deeply couples the spatial power flow flexibility control of the Energy Flow Regulator (EER) with the temporal energy shifting characteristics of the Distribution Execution System (DES) through a two-layer model. Compared to single-device planning schemes, this invention can release energy through the DES to smooth out peak loads during peak periods, while simultaneously utilizing the EER to direct surplus power to weak nodes, effectively mitigating voltage fluctuations. This collaborative mechanism significantly enhances the distribution network's capacity to absorb distributed generation without substantially increasing investment costs.

[0115] The algorithm demonstrates high efficiency and reliability. By introducing information geometry theory, the Riemannian manifold evolution algorithm designed in this invention completely overcomes the shortcomings of traditional evolutionary algorithms, such as blind searching and susceptibility to local optima in high-dimensional non-convex spaces. Utilizing the Fisher information matrix and natural gradients, the algorithm can automatically perceive the strong electrical coupling relationships between decision variables, guiding the population to search along the steepest descent direction. In practical engineering calculations, this manifests as a faster convergence speed and a more uniform Pareto solution set distribution, providing power planners with a more scientific decision-making reference.

[0116] Excellent economic and technical balance. This invention not only minimizes investment costs but also ensures optimal voltage deviation across the entire network through precise lower-level power flow simulation. This planning paradigm of "overall planning and dynamic optimization" effectively avoids redundant construction caused by excessive investment in power assets or insufficient capacity, and has extremely high practical value.

[0117] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A collaborative planning method for energy routers and distributed energy storage that takes into account load growth, characterized in that, include: The historical source-load data of the target distribution network for several consecutive years is obtained. After preprocessing, a historical net load time series is constructed. The variational mode decomposition algorithm is used to decompose the historical net load time series into eigenmode functions with different center frequencies, and divides them into low-frequency trend terms, medium-frequency periodic terms and high-frequency random terms. The low-frequency trend term is predicted for future trends to obtain a low-frequency trend prediction component; the mid-frequency periodic term is predicted based on its historical periodic structure; the high-frequency random term is reconstructed using a conditional diffusion model to obtain a high-frequency random prediction component; based on the three prediction components, a set of typical daily scenarios and their corresponding weights for the future target year are obtained through superposition clustering; the future target year is the year following the latest year in the historical source load data. Using the typical daily scenario set and corresponding weights as input, a two-layer optimization model for collaborative planning is constructed. The upper-layer model is used to decide on the site selection and capacity determination scheme of energy routers and distributed energy storage, with the goal of minimizing the annualized investment cost and the cumulative voltage deviation of the system. The lower-layer model is used to perform time-series operation simulation of the distribution network under the given upper-layer scheme, and feeds back the simulation results to the upper layer for scheme evaluation. The Riemannian manifold evolution algorithm is used to solve the two-level optimization model, and the Pareto optimal planning scheme set is output. The final planning scheme is determined by evaluating the Pareto optimal planning scheme set.

2. The method according to claim 1, characterized in that, The process involves acquiring historical source-load data of the target distribution network over several consecutive years, preprocessing it to construct a historical net load time series, and then using a variational mode decomposition algorithm to decompose the historical net load time series into eigenmode functions with different center frequencies. These eigenmode functions are further divided into low-frequency trend terms, mid-frequency periodic terms, and high-frequency random terms, including: Obtain historical source and load data of the target distribution network over the past several years, including historical load data and distributed generation output data; The historical load data and the distributed power output data are time-aligned and then preprocessed. The preprocessing includes missing value imputation, outlier removal, noise reduction and smoothing, thereby constructing a historical net load time series. The historical net load time series is decomposed into multiple eigenmode functions with different center frequencies using a variational mode decomposition algorithm. Based on the magnitude of multiple center frequencies and two preset frequency thresholds, multiple intrinsic mode functions are divided into low-frequency trend terms, mid-frequency periodic terms, and high-frequency random terms.

3. The method according to claim 1, characterized in that, The low-frequency trend term is used to predict future trends, resulting in a low-frequency trend prediction component; the mid-frequency cycle prediction component is obtained based on the historical cycle structure of the mid-frequency cycle term. By reconstructing the high-frequency random term using a conditional diffusion model, high-frequency random prediction components are obtained, including: The low-frequency trend term is used to predict future trends by combining linear extrapolation with logistic regression, and the low-frequency trend prediction component for the future target year is obtained. The historical periodic structure of the intermediate frequency periodic term is preserved, and periodic mapping and phase alignment are performed according to the calendar information of the future target year to obtain the intermediate frequency periodic prediction component of the future target year. A one-dimensional conditional diffusion model is trained using the high-frequency random terms as training data. The conditional diffusion model uses time feature information as a conditional vector, and the time feature information includes at least month, weekday, hour and weekday identifiers. Based on the trained conditional diffusion model and the time feature information of the future target year, a high-frequency random prediction component corresponding to the future target year is generated.

4. The method according to claim 1, characterized in that, Based on the three predicted components, a set of typical daily scenarios and their corresponding weights for the future target year is obtained through superposition clustering, including: The low-frequency trend prediction component, the mid-frequency periodic prediction component, and the high-frequency random prediction component are superimposed hourly to reconstruct a forward-looking net load sequence for the future target year. K-Means clustering is performed on the prospective net load sequence to extract multiple typical daily scenarios, and the occurrence weight of each typical daily scenario in the whole year is calculated to obtain the set of typical daily scenarios and their corresponding weights for the future target year.

5. The method according to claim 1, characterized in that, In the two-layer optimization model, the construction process of the upper-layer model includes: Using the typical daily scene set and corresponding weights as input to the upper-level model, the upper-level decision vector is defined as follows: ； in, and Distributed energy storage and energy routers at the nodes, respectively. The addressing Boolean variable, and These are the rated power and rated energy storage capacity of distributed energy storage, respectively. The rated switching power of the energy router; a set of decision values ​​of the upper-layer decision vector constitutes an addressing and capacity determination scheme; Establish the first objective function at the upper level, with the goal of minimizing the annualized investment cost, expressed as: ； in, This is the first objective function of the upper layer; This is the annualized coefficient. , These are the unit power cost and unit capacity cost of distributed energy storage, respectively. Cost per unit power of the energy router; This represents the set of candidate installation nodes for distributed energy storage. This represents the set of candidate installation nodes for the energy router; Establish a second objective function at the upper level, with the goal of minimizing the cumulative voltage deviation of the system, expressed as: ； in, This is the second objective function of the upper layer; Typical daytime scene Next node At any moment The per-unit voltage value, This is the per-unit value of the rated voltage; This represents the number of intrinsic mode functions with different center frequencies obtained by decomposing the historical net load time series; Indicates the first Weighting of a typical daily scenario; Represents a set of discrete time periods within a single typical day scenario; Represents the set of all nodes in the target distribution network. and All A subset of; This indicates the calculation of absolute value; Set upper-level constraints related to the upper-level decision vector, including at least constraints on the number of equipment locations, constraints on equipment port configuration, and upper and lower limits on the rated power and rated capacity of the equipment, so as to ensure that the planning scheme meets the requirements of engineering feasibility. The upper-level model is composed of the upper-level decision vector, the upper-level first objective function, the upper-level second objective function, and the upper-level constraints.

6. The method according to claim 1 or 5, characterized in that, In the aforementioned two-layer optimization model, the construction process of the lower-layer model includes: Obtain the site selection and capacity determination schemes decided by the upper-layer model, as well as the typical daily scene set; Using the LinDistflow power flow model, a time-series operation simulation architecture for the distribution network is constructed, and operational constraints are set. These constraints include: power flow balance constraints, voltage drop constraints, state of charge evolution constraints for distributed energy storage, and cross-node energy conservation constraints for energy routers. For each typical day scenario in the typical day scenario set, under the conditions of the site selection and capacity setting scheme, a 24-hour time-series operation simulation calculation is performed to obtain the per-unit voltage value of all nodes of the target distribution network, the power flow of all branches, the state of charge and charging / discharging power of each distributed energy storage, and the port switching power of each energy router in each time period under the typical day scenario. The per-unit voltage values ​​of each node obtained from the time-series simulation calculation are fed back to the upper-level model as input data for calculating the objective function of the system's cumulative voltage deviation, in order to evaluate the operational performance of the location and capacity setting scheme.

7. The method according to claim 6, characterized in that, The power flow balance constraint is expressed as: ； in, Represents any node in the target distribution network; Represented by node For all branches at the power outflow end End node A set; Indicates at time From node Flow to Node branch road Active power; Represented by node For all branches at the power inflow end The first node A set; and These are all mathematical set symbols; Indicates at time From node Flow to Node branch road Active power; Indicates a branch The resistance; Indicates at time Flowing through a branch road The current; Indicating a typical daytime scenario at a specific time Access Node Total load power; Indicating a typical daytime scenario at a specific time Access Node The output power of distributed photovoltaic power sources; Indicates at time Installed on node The charging power of distributed energy storage; Indicates at time Installed on node The discharge power of distributed energy storage; Indicates at time Installed on node The active power exchanged between the energy router (EER) and the power grid is positive when injected into the grid and negative when absorbed from the grid. The voltage drop constraint is expressed as follows: ； in, Indicates at time ,node The per-unit voltage value; Indicates at time ,node The per-unit voltage value; Indicates a branch The resistance; Indicates a branch The reactance; Indicates at time From node Flow to Node branch road reactive power; The charge state evolution constraint of the distributed energy storage is expressed as: ； in, Indicates that it is installed on the node Distributed energy storage at any time The state of charge; Indicates that it is installed on the node Distributed energy storage at any time The state of charge; This indicates the charging efficiency of distributed energy storage; Indicates at time Installed on node The charging power of distributed energy storage; Indicates the simulation time step; Indicates that it is installed on the node The rated energy storage capacity of distributed energy storage; Indicates at time Installed on node The discharge power of distributed energy storage; This indicates the discharge efficiency of distributed energy storage; The cross-node energy conservation constraint of the energy router is expressed as: ； in, This indicates the injected power of the EER at different ports; Indicates at time The energy router from its port Active power injected into the power grid; This represents the power absorbed by the EER at different ports; Indicates at time The energy router from its port Active power absorbed from the power grid; This represents the set of all nodes in the target distribution network; This represents the conversion efficiency of the energy router.

8. The method according to claim 1, characterized in that, The Riemannian manifold evolution algorithm is used to solve the two-level optimization model, outputting a set of Pareto optimal planning solutions, including: The solution problem of the bi-level optimization model is mapped onto a statistical manifold composed of a family of probability distributions, and the weight vector of the multi-objective decomposition and the distribution parameters defining the probability distribution of candidate solutions are initialized to construct the initial search state of the algorithm; wherein, the statistical manifold ; It is the upper-level decision vector that serves as the input to the upper-level model, representing the candidate planning schemes for the location and capacity determination of energy routers and distributed energy storage. Represents the probability density function; These are distribution parameters; Represents the parameter space; Based on the initialized distribution parameters, the Fisher information matrix of the current probability distribution on the statistical manifold is calculated, and the natural gradient is calculated based on the Fisher information matrix to guide the update of the search direction; wherein, the formula for calculating the Fisher information matrix is: ； The formula for calculating the natural gradient is: ； in, Represents the first element in the Fisher information matrix. u Okay, number v Column elements; For mathematical expectation operators; Indicates the first u One distribution parameter; Indicates the first v One distribution parameter; Indicates the ordinary gradient; This represents the inverse matrix of the Fisher information matrix; Represents the natural gradient; The natural gradient is iteratively updated by introducing a Riemann momentum term, and the update formula is as follows: ； in, Indicates the first The Riemann momentum term at the next iteration; Indicates the momentum decay coefficient; The covariance matrix is ​​used to characterize the correlation between decision variables; Evolutionary state entropy during the monitoring algorithm iteration process ,when Lower than the preset reference entropy At this time, the search step size is adaptively increased to encourage the algorithm to escape the local optimum. Based on the initialized weight vector, the upper-level multi-objective problem of the two-level optimization model is decomposed into multiple scalar quantum problems and solved in parallel using a multi-objective evolutionary algorithm decomposition framework. Iterative optimization is then performed, and finally, a Pareto optimal programming scheme set composed of non-dominated solutions is output.

9. The method according to claim 1, characterized in that, The step of determining the final planning scheme by evaluating the Pareto optimal planning scheme set includes: The grey relational analysis method is used to determine the final planning scheme from the Pareto optimal planning scheme set.

10. The method according to claim 9, characterized in that, The step of determining the final planning scheme from the Pareto optimal planning scheme set using grey relational analysis includes: For each planning scheme in the Pareto optimal planning scheme set, an evaluation index set is constructed; the evaluation index set includes at least the initial investment cost, system voltage offset, and redundancy for adapting to future load growth. The evaluation indicators in the set of evaluation indicators are dimensionless to obtain normalized indicator values. Determine a reference sequence consisting of the optimal values ​​of each evaluation index; For each planning scheme, calculate the grey relational coefficient between its normalized evaluation index values ​​and the corresponding index values ​​in the reference sequence; Based on the preset weights of each evaluation index, the weighted sum of the grey relational coefficients of all evaluation indexes for each planning scheme is calculated to obtain the grey relational degree of the planning scheme. From the Pareto optimal planning scheme set, the planning scheme with the largest grey relational degree is selected as the final planning scheme to represent the final site selection and sizing plan.

Images