Method and system for predicting resource aggregation capability of microgrid

Abstract

The present application relates to the technical field of micro-grid resource scheduling, and particularly relates to a micro-grid resource aggregation capability prediction method and system, the method comprising: obtaining source-load-storage load data of the micro-grid, the data being used to represent the operating state of the distributed power supply, flexible load and energy storage device; constructing an overall aggregation feasible interval according to the individual feasible interval of the three types of resources, and creating an aggregated power energy boundary model accordingly; inputting the model into a deep neural network ultra-short-term prediction framework, predicting the future dynamic evolution trend by the framework, and outputting the predicted aggregation feasible interval, so as to accurately depict the power-energy joint feasible region of the source-load-storage integrated micro-grid within the minute scale, quantify the nonlinear cumulative effect of the micro-grid evolution over time, and improve the reliability and timeliness of subsequent scheduling decisions.

Description

Technical Field

[0001] This invention relates to the field of microgrid resource scheduling technology, and in particular to a method and system for predicting the resource aggregation capability of a microgrid. Background Technology

[0002] Distributed photovoltaic (PV) power, energy storage, and flexible loads are experiencing explosive growth on the distribution network side. Microgrids are gradually evolving from single "source-load" balancing units into complex, controllable clusters with multi-energy complementarity and spatiotemporal coupling. Theoretically, such clusters can present a dispatchable extension approximating a "virtual power plant" within minutes to hours, providing ancillary services such as frequency regulation, voltage regulation, and backup power to the upper-level grid. However, after long-term field testing and simulation verification, the inventors found that the aforementioned "extension" still faces significant bottlenecks in its engineering implementation: (1) Distributed resources have small individual capacity and scattered locations. Their power-energy feasible boundary is driven by multiple random variables such as weather, user behavior, and market electricity price, exhibiting strong time-varying and non-convex characteristics. Traditional aggregation models based on static envelope or linear extrapolation will experience boundary drift within a minute scale, resulting in a deviation of more than 30% between the planned and actual available amount for the day and day.

[0003] (2) Existing research on “flexible resource aggregation feasible region” mainly focuses on demand-side response scenarios, with prediction steps generally greater than 1 hour, and takes “individual modeling-convex hull approximation-offline aggregation” as the main line. This paradigm exposes two major defects in the ultra-short-term (≤4h) scheduling of microgrids: First, in order to ensure conservatism, the convex hull approximation continuously shrinks the feasible region inward, resulting in a waste of effective capacity redundancy; Second, offline aggregation cannot capture the dynamic cumulative effects such as photovoltaic fluctuations and energy storage cycle efficiency decay, so the closer to real-time operation, the more significant the spatiotemporal misalignment between the aggregation boundary and the real feasible region.

[0004] (3) If the dispatch center issues instructions based on the distorted aggregation boundary, it is very easy to trigger multiple safety constraints such as overcharging of energy storage, curtailment of photovoltaic power or over-limit of tie line power, which will ultimately force the automatic generation control (AGC) to frequently reverse, weakening the ability of the microgrid cluster to provide rapid backup to the outside world. Summary of the Invention

[0005] In view of the above-mentioned deficiencies or disadvantages, the present invention provides a method and system for predicting the resource aggregation capability of microgrids, which can solve at least one of the above technical problems.

[0006] This invention provides a method for predicting the resource aggregation capability of a microgrid, comprising: Acquire source-load-storage load data of the microgrid, which includes the operating status data of distributed power sources, flexible loads, and energy storage devices; The overall aggregated feasible range is constructed based on the individual feasible ranges of distributed power sources, flexible loads, and energy storage devices. Each individual feasible range is determined by the corresponding operating status data. Create a convergent power energy boundary model based on the overall feasible range of convergence; The aggregated power energy boundary model is input into a pre-defined deep neural network ultra-short-term prediction framework. The ultra-short-term prediction framework predicts the future dynamic evolution trend of the overall aggregated feasible interval and outputs the predicted aggregated feasible interval.

[0007] According to a second aspect, the present invention provides a microgrid resource aggregation capability prediction system, comprising: The multi-dimensional load data acquisition module is used to acquire the source-load-storage load data of the microgrid. The source-load-storage load data includes the operating status data of distributed power sources, flexible loads, and energy storage devices. The aggregated feasible range setting module is used to construct the overall aggregated feasible range based on the individual feasible ranges of distributed power sources, the individual feasible ranges of flexible loads, and the individual feasible ranges of energy storage devices. Each individual feasible range is determined by the corresponding operating status data. The energy boundary model creation module is used to create a pooled power energy boundary model based on the overall pooling feasible range. The short-term aggregation capability prediction module is used to input the aggregation power energy boundary model into a preset deep neural network ultra-short-term prediction framework. The ultra-short-term prediction framework predicts the future dynamic evolution trend of the overall aggregation feasible range and outputs the predicted aggregation feasible range.

[0008] According to a third aspect, the present invention provides an electronic device comprising: At least one processor; and The memory that is communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to execute the resource aggregation capability prediction method for any microgrid in the embodiments of the present invention.

[0009] According to another aspect of the present invention, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute a resource aggregation capability prediction method for any microgrid in the embodiments of the present invention.

[0010] By adopting the technical solution of this invention, the source-load-storage load data of the microgrid is obtained. First, an overall aggregated feasible region is constructed based on the individual feasible regions of distributed power sources, flexible loads, and energy storage devices. Based on this, an aggregated power-energy boundary model is created. Then, the model is input into an ultra-short-term prediction framework of a deep neural network. The framework predicts the future dynamic evolution trend of the overall aggregated feasible region and outputs the predicted aggregated feasible region. This enables the accurate capture of the adjustable range changes of controllable resources of the microgrid within an ultra-short-term time scale. It achieves accurate characterization of the power-energy joint feasible region of the source-load-storage integrated microgrid within a minute scale and quantifies the nonlinear cumulative effect of the microgrid's evolution over time, thereby improving the reliability and timeliness of subsequent scheduling decisions. Attached Figure Description

[0011] Figure 1 This is a flowchart of a microgrid resource aggregation capability prediction method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the sliding window algorithm of the present invention; Figure 3 This is a schematic diagram of the Transformer neural network structure of the present invention; Figure 4 This is a schematic diagram of the deep reinforcement learning training loss result of the present invention; Figure 5 This is a schematic diagram comparing the MSE results of reinforcement learning models with different depths according to the present invention; Figure 6 This is a structural block diagram of a microgrid resource aggregation capability prediction system according to an embodiment of the present invention; Figure 7 This is a block diagram of an electronic device used to implement embodiments of the present invention. Detailed Implementation

[0012] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0013] This invention provides a method for predicting the resource aggregation capability of a microgrid, based on a first aspect. This method can be applied to a controllable resource aggregation capability prediction system for a microgrid (hereinafter referred to as the "system"), such as... Figure 1 As shown, the method may include: Step S110: Obtain the source-load-storage load data of the microgrid.

[0014] Among them, the source-load-storage load data includes the operating status data of distributed power sources, flexible loads, and energy storage devices.

[0015] Specifically, distributed power sources can refer to renewable energy power generation equipment, represented by photovoltaic power generation units, that are connected to voltage levels of 35 kV and below, located on the user side or distribution network side. The output power of distributed power sources can be adjusted as needed within the rated capacity range; for example, it can be a photovoltaic power generation unit. Flexible loads can refer to electrical equipment that has the ability to be interrupted, transferred, or reduced, and can quickly adjust its power consumption within a specified power range by receiving external signals; for example, it can be a 380-volt central air conditioning unit with a remote start / stop interface. Energy storage devices can refer to complete sets of equipment that use electrochemical batteries as the core and realize the storage and release of electrical energy through charge and discharge control. Its state of charge and charge / discharge power are directly controlled by the energy management system; for example, it can be a 100 kWh lithium iron phosphate energy storage integrated cabinet connected to a 380-volt bus. Therefore, source-load-storage load data can include the active power, reactive power, and port voltage of the photovoltaic power generation unit; the active power consumption, reactive power consumption, and reduceable capacity flag of the flexible load; and the state of charge, charge / discharge power, and maximum allowable charge / discharge power of the energy storage device.

[0016] The system can continuously record the electrical quantities of each node at a sampling rate of no less than 10 Hz through an embedded synchronous data acquisition terminal with a unified clock reference and support for synchronous sampling of multiple electrical quantities, thereby constructing a source-load-storage load dataset.

[0017] For example, the synchronous data acquisition terminal can be installed in an outdoor stainless steel enclosure at the microgrid grid connection point. This enclosure has a protection rating of at least IP54 and is secured internally using DIN (Deutsches Institut für Normung rail) guide rails to ensure the terminal is vertical and away from high-voltage live parts. In IP54, "I" indicates a dustproof rating of 5 (not completely preventing dust ingress, but the amount of dust will not affect normal operation), and "P" indicates a waterproof rating of 4 (preventing harmful effects from splashes of water from all directions). The terminal connects to the local monitoring and control module via an RS485 bus (a serial communication standard for long-distance, multi-node data exchange), with a communication baud rate set to 115200 bits per second. It uses the Modbus-RTU protocol (a master-slave serial communication protocol) to read the aforementioned operating status data. Simultaneously, the terminal incorporates a temperature-compensated real-time clock chip (model PCF8563, accuracy ±5ppm, or ±5 parts per million) and uses GPS (Global Positioning System) for data acquisition. Using the Global Positioning System (GPS) second pulse (also known as 1PPS) as a reference, the system calibrates its time once per second to achieve millisecond-level timestamp alignment, forming a time-stamped source-load storage load data frame.

[0018] Furthermore, assuming that the system reads the photovoltaic active power as 850 kW, reactive power as 120 kvar, port voltage as 10.2 kV, flexible load active power consumption as 600 kW, reduceable capacity flag as 1 (indicating it can participate in regulation), energy storage state of charge as 65%, and maximum allowable discharge power as 500 kW within a certain sampling period, then the above data, after being encapsulated by the synchronous data acquisition terminal, is uploaded to the edge computing node in JSON (JavaScript Object Notation) format via a 4G wireless module (referring to an industrial-grade wireless transmission unit based on fourth-generation mobile communication technology) for subsequent steps to call.

[0019] Step S120: Construct an overall aggregated feasible region based on the individual feasible regions of distributed power sources, flexible loads, and energy storage devices.

[0020] The feasible intervals for each individual device are determined by the corresponding operational status data. Specifically, an individual feasible interval can refer to the set of boundaries that a single device is allowed to operate on in the power-energy two-dimensional plane, including the upper power limit, lower power limit, upper energy limit, and lower energy limit. In this specification, "limit" and "boundary" are used interchangeably, both referring to the allowable extreme values ​​of the corresponding physical quantities in terms of numerical values. Therefore, "upper power limit" is the same as "upper power limit," and "lower energy limit" is the same as "lower energy limit," and so on.

[0021] For example, for photovoltaic power generation units, based on the current light intensity and temperature correction curve, the system obtains an upper limit of 1000 kW and a lower limit of 0 kW for their active power output, and an upper limit of the cumulative power generation of the theoretical power generation of the day and a lower limit of the power generation already generated, forming an individual feasible range for distributed power sources; for flexible loads, based on the user's interruptible load contract, the system obtains an upper limit of 200 kW and a lower limit of 0 kW for their power reduction, and an upper limit of the cumulative power reduction of the contracted power and a lower limit of the power reduction already reduced, forming an individual feasible range for flexible loads; for energy storage devices, based on the current state of charge and maximum charging and discharging power, the system obtains an upper limit of 500 kW for discharge power and an upper limit of -400 kW for charging power (the negative sign - only indicates the direction of charging, and the dimension is still positive), an upper limit of the cumulative discharge energy of the remaining dischargeable energy and an upper limit of the cumulative charging energy of the remaining rechargeable energy, forming an individual feasible range for energy storage devices.

[0022] Next, the system can use a linear superposition method to add the upper power bounds of the three types of individuals to obtain the upper power bound of the overall aggregate feasible interval as 1000 + 0 + 500 = 1500 kW, and the lower power bound as 0 - 200 - 400 = -600 kW (the negative sign - only indicates the charging direction, and the dimension is still positive). Similarly, the upper and lower energy bounds are superimposed to form a closed quadrilateral overall aggregate feasible interval. Among them, the overall aggregate feasible interval is used to characterize the total power-energy boundary that all controllable resources in the microgrid can jointly provide in the next time period.

[0023] Step S130: Create a power energy boundary model based on the overall feasible range of aggregation.

[0024] Among them, the aggregated power-energy boundary model is used to convert the four-dimensional boundaries (power upper bound, power lower bound, energy upper bound, and energy lower bound) of the overall aggregated feasible interval into a standardized data structure for deep learning frameworks to read.

[0025] Specifically, the system can use a pre-defined Python script to call the Pandas library (an open-source data analysis tool) to write the four-dimensional boundary of the overall aggregated feasible interval into a CSV (Comma-Separated Values) file. The field order is [Power Upper Bound, Power Lower Bound, Energy Upper Bound, Energy Lower Bound]. The file encoding uses UTF-8 (Universal Character Set Conversion Format 8-bit encoding), and the row frequency is synchronized with the sampling period. Subsequently, the script uses the MinMaxScaler function (MinMaxScaler function: a standardization tool that linearly maps data to the 0-1 interval) to normalize the four-dimensional boundary, generating an aggregated power and energy boundary model file with the file extension .pkl, which is stored on the solid-state drive of the edge computing node.

[0026] For example, the overall aggregate feasible range obtained by the system at a certain moment is an upper power limit of 1500 kW, a lower power limit of -600 kW, an upper energy limit of 800 kWh, and a lower energy limit of -400 kWh. After being normalized by MinMaxScaler, the values ​​in the aggregated power-energy boundary model are mapped to the interval of 0 to 1, forming a dimensionless model input vector [0.82, 0.00, 1.00, 0.00] for subsequent neural network reading; where the negative sign - only indicates the charging direction, and the dimension is still positive.

[0027] Step S140: Input the aggregated power energy boundary model into the preset deep neural network ultra-short-term prediction framework, and the ultra-short-term prediction framework predicts the future dynamic evolution trend of the overall aggregated feasible interval and outputs the predicted aggregated feasible interval.

[0028] Among them, the ultra-short-term prediction framework adopts the Transformer encoder structure (a deep learning network unit based on self-attention mechanism) to capture the temporal dependencies of the aggregation boundary; the predicted aggregation feasible interval refers to the set of power-energy boundaries after model deduction in the future time period.

[0029] Next, the system can load the .pkl file generated in step S130 into the prediction service process via memory mapping. The process is built on the PyTorch framework (an open-source deep learning framework), and the model weight file is saved in .pt format. The number of stacked layers of the Transformer encoder is set to 2, the number of attention heads is set to 4, the feedforward hidden dimension is set to 64, the length of the input sequence corresponds to the length of the historical time period, and the length of the output sequence corresponds to the length of the future time period. After the model performs forward inference, the normalized prediction boundary vector is output, and then restored to the actual dimensions by the inverse MinMaxScaler (a minimum-maximum normalization scaler) to form the prediction aggregation feasible interval.

[0030] Then, the system performs a validity check on the predicted aggregated feasible interval: if the upper bound of the predicted power is less than the lower bound of the power or the upper bound of the predicted energy is less than the lower bound of the energy, an anomaly flag is triggered, and the aggregated feasible interval of the previous time step is used as a substitute output to ensure that the result is physically reasonable.

[0031] Alternatively, when the load rate of edge computing nodes exceeds a preset threshold, the system can upload the aggregated power energy boundary model to the cloud inference service via the MQTT protocol (Message Queuing Telemetry Transport, a lightweight IoT message transmission protocol). The cloud will then return the predicted aggregated feasible range, achieving cloud-edge collaborative prediction.

[0032] Therefore, according to the above implementation method, the system can acquire the source-load-storage load data of the microgrid, first construct the overall aggregated feasible range based on the individual feasible ranges of distributed power sources, flexible loads and energy storage devices, and create an aggregated power-energy boundary model accordingly. Then, the model is input into the ultra-short-term prediction framework of the deep neural network, which predicts the future dynamic evolution trend of the overall aggregated feasible range and outputs the predicted aggregated feasible range. This enables the system to accurately capture the adjustable range changes of the microgrid's controllable resources within an ultra-short-term time scale, accurately characterize the power-energy joint feasible domain of the integrated source-load-storage microgrid within a minute-level scale, and quantify the nonlinear cumulative effect of the microgrid's evolution over time (i.e., the prediction of the microgrid's ultra-short-term controllable resource aggregation capability), thereby improving the reliability and timeliness of subsequent scheduling decisions.

[0033] In some embodiments, the above method further includes: Based on the source-load-storage load data, individual power boundaries and individual energy boundaries are constructed for distributed power sources, flexible loads, and energy storage devices.

[0034] A source-load-storage power-energy boundary model is created based on the power boundary and energy boundary of each volume.

[0035] In the source-load-storage power energy boundary model: Individual power boundaries and individual energy boundaries of distributed generation are used to characterize the feasible mapping range between the output power and cumulative power generation of distributed generation; individual power boundaries and individual energy boundaries of flexible loads are used to characterize the feasible mapping range between the power consumption and cumulative power consumption of flexible loads; individual power boundaries and individual energy boundaries of energy storage devices are used to characterize the feasible mapping range between the charging and discharging power and the range of state of charge changes of energy storage devices.

[0036] For example, the source-load-storage power energy boundary model includes the photovoltaic power-energy boundary model, the load power-energy boundary model, and the energy storage power-energy boundary model.

[0037] For example, when creating a photovoltaic power-energy boundary model, the individual power boundary and individual energy boundary of a distributed generation are determined in the following manner: The upper limit of photovoltaic power = current irradiance (W / m²) × total area of ​​photovoltaic modules (m²) × module conversion efficiency (dimensionless, taken as 18%) × maximum allowable output coefficient of inverter (dimensionless, taken as 98%); the lower limit of photovoltaic power is 0 kilowatts. Alternatively, the upper limit of photovoltaic power can be determined by the following formula: ; Where G represents the current irradiance (W / m²), and A represents the total area of ​​the photovoltaic modules (W / m²). ), η 表示组件转换效率（取18%），k表示逆变器最大允许出力系数（取98%）。

[0038] 光伏能量上边界=当日太阳辐射总量（千瓦时每平方米）×光伏组件总面积（平方米）×组件转换效率×逆变器效率（无量纲，取96%）－已发电量（千瓦时）；光伏能量下边界为0千瓦时。

[0039] 具体而言，系统可以通过下述的公式化条件创建光伏功率-能量边界模型：；；；；；；其中，为时刻光伏的实际功率；、分别为时刻光伏功率的下边界和上边界；为时刻光伏的实际能量；、分别为时刻光伏能量的下边界和上边界。

[0040] 示例性地，在创建负荷功率-能量边界模型时，柔性负荷的个体功率边界与个体能量边界通过下述方式确定：负荷功率上边界=用户合同最大需求（千瓦，取配电营销系统当月核定值）；负荷功率下边界=可中断负荷最小运行功率（千瓦，取设备铭牌额定功率的10%，或5千瓦，二者取大者）；负荷能量上边界=合同可削减电量（千瓦时，取需求响应协议中"日可削减总量”）－已削减电量（千瓦时，由终端电表累计）；负荷能量下边界为0千瓦时。

[0041] 具体而言，系统可以通过下述的公式化条件创建负荷功率-能量边界模型：；；；；；；式中，为时刻负荷的实际功率；、分别为时刻负荷功率的下边界和上边界；为时刻负荷的实际能量；、分别为时刻负荷能量的下边界和上边界。

[0042] 示例性地，在创建储能功率-能量边界模型时，储能装置的个体功率边界与个体能量边界通过下述方式确定：储能功率上边界=储能变流器铭牌最大放电功率（千瓦，取设备参数表值）；储能功率下边界=－储能变流器铭牌最大充电功率（千瓦，负号表示吸收功率）；储能能量上边界=当前荷电状态（%）×电池额定容量（千瓦时）×放电效率（无量纲，取92%）；储能能量下边界=－(100%－当前荷电状态)×电池额定容量（千瓦时）×充电效率（无量纲，取92%)；或者，储能能量边界可以通过下述公式来定义：能量上边界：；能量下边界：；其中，SOC表示当前荷电状态（0~1），C表示电池额定容量（kWh），η、η分别表示放电与充电效率（均取92%），负号（-）仅表示可充电方向，量纲仍为正数千瓦时。

[0043] 初始荷电状态由上一控制周期末端SOC（State of Charge，荷电状态）实测值继承；额定容量与预测时长直接读取电池管理系统BMS的静态参数区与运行日历，无需额外输入。

[0044] 具体而言，系统可以通过下述的公式化条件创建储能功率-能量边界模型：；；；；；；，；式中，为时刻储能的实际充 / 放电功率；、分别为时刻储能放电功率和充电功率的最大值；为时刻储能的实际能量、分别为时刻储能能量的下边界和上边界；分别为时刻储能充 / 放电效率；为储能初始时刻荷电状态；为储能额定容量；T为预测的时间长度。

[0045] 因此，根据上述实施方式，系统能够将分布式电源、柔性负荷及储能装置的运行状态数据转化为统一的功率-能量边界集合，形成源荷储功率能量边界模型，为后续聚合与预测提供结构化输入基础。

[0046] 在一些实施例中，根据分布式电源、柔性负荷及储能装置的个体可行区间构建整体聚合可行区间，包括：分别提取分布式电源、柔性负荷及储能装置的个体功率边界与个体能量边界。

[0047] 其中，个体功率边界指单台设备在下一控制时段内可输出的最大功率与可吸收的最小功率，个体能量边界指单台设备在下一控制时段内可释放的最大能量与可存储的最大能量。

[0048] 将分布式电源、柔性负荷及储能装置的个体功率边界进行叠加，得到整体聚合可行区间的功率上界与功率下界。

[0049] 具体而言，上述的叠加操作可通过下述的计算方式完成：功率上界=Σ分布式电源功率上界+Σ柔性负荷功率上界+Σ储能装置功率上界；其中，该（柔性负荷）功率上界为可削减量，叠加时已取负方向；功率下界=Σ分布式电源功率下界+Σ柔性负荷功率下界+Σ储能装置功率下界；单位均为千瓦。

[0050] 将分布式电源、柔性负荷及储能装置的个体能量边界进行叠加，计算得到整体聚合可行区间的能量上界与能量下界。

[0051] 具体而言，上述的叠加操作可通过下述的计算方式完成：能量上界=Σ分布式电源能量上界+Σ柔性负荷能量上界+Σ储能装置能量上界；能量下界=Σ分布式电源能量下界+Σ柔性负荷能量下界+Σ储能装置能量下界；单位均为千瓦时。

[0052] 基于整体聚合可行区间的功率上界、功率下界、能量上界及能量下界，构建整体聚合可行区间。

[0053] 具体而言，整体聚合可行区间用于表征微电网内所有可控资源在下一时段可共同提供的总功率-能量边界，为后续聚合功率能量边界模型提供输入。

[0054] 因此，根据上述实施方式，系统能够将分散的个体边界线性叠加为统一的四维边界，实现资源集群可调能力的结构化描述，并为超短期预测框架提供标准化输入。

[0055] 在一些实施例中，根据整体聚合可行区间创建聚合功率能量边界模型，包括：基于整体聚合可行区间的功率上界与功率下界，生成聚合功率能量边界模型的聚合功率上边界与聚合功率下边界。

[0056] 系统可以将聚合功率上边界取值为功率上界数值，聚合功率下边界取值为功率下界数值，单位均为千瓦，无需额外变换。

[0057] 基于整体聚合可行区间的能量上界与能量下界，生成聚合功率能量边界模型的聚合能量上边界与聚合能量下边界。

[0058] 系统可以将聚合能量上边界取值为能量上界数值，聚合能量下边界取值为能量下界数值，单位均为千瓦时，保持量纲一致。

[0059] 其中，聚合功率能量边界模型中：聚合功率上边界与聚合功率下边界用于界定微电网在预测时段内的总功率可行范围。聚合能量上边界与聚合能量下边界用于界定微电网在预测时段内的总能量可行范围。总功率可行范围与总能量可行范围共同构成四维边界向量，用于后续深度神经网络输入。

[0060] 示例性地，系统可以通过下述的公式化条件创建聚合功率能量边界模型：；；；；；；其中，为时刻可调资源聚合功率；、分别为时刻可调资源聚合功率的下边界和上边界；为时刻可调资源聚合能量；、分别为时刻可调资源聚合能量的下边界和上边界；为功率基线；为能量基线。

[0061] 因此，根据上述实施方式，系统能够将整体聚合可行区间的四维边界无损映射为聚合功率能量边界模型，实现边界数据的标准化封装，为超短期预测框架提供可直接读取的输入向量。

[0062] 在一些实施例中，超短期预测框架包括时序数据集构建模块与预测模型模块；将聚合功率能量边界模型输入预设的深度神经网络的超短期预测框架，包括：对聚合功率能量边界模型进行数据预处理，生成归一化聚合边界数据。

[0063] 其中，时序数据集构建模块是一种用于将归一化聚合边界数据按固定时间窗口滑动采样，生成"历史输入—未来输出”样本对的程序单元，输出为{X→Y}序列。预测模型模块是基于Transformer编码器结构的轻量化神经网络程序单元，用于接收{X→Y}样本，执行前向推理并输出未来时段归一化聚合边界预测向量。

[0064] 系统在进行数据预处理时，可以依次采用缺失值线性插补、异常值截断及最小-最大归一化的步骤，令归一化结果映射至0～1区间，得到四维归一化向量：[P_max_norm,P_min_norm,E_max_norm,E_min_norm]。

[0065] 将归一化聚合边界数据输入时序数据集构建模块，生成输入输出样本对，输入输出样本对表征归一化聚合边界数据在历史时段与未来时段之间的映射关系。

[0066] 示例性地，时序数据集构建模块可以采用滑动窗口机制（滑动窗口机制：在时间序列上固定长度采样并步进一格的构造方法），窗口长度取16个采样点，步长取1个采样点；每个窗口内16组归一化向量作为输入X，对应未来1个采样点的归一化向量作为输出Y，形成输入输出样本对{X→Y}；该样本对即表征归一化聚合边界数据在历史时段与未来时段之间的映射关系。

[0067] 在一些实施例中，系统可以采用如图2所示的滑动窗口算法，对归一化聚合边界数据进行时序切片，以生成输入输出样本对。该算法的核心参数包括：采用滑动窗口机制构建样本集：窗口长度：16个采样点（输入序列）；滑动步长：1个采样点，确保高样本密度与序列连续性；输出目标：紧邻窗口后的第1个采样点（4维边界向量）。

[0068] 具体而言，系统按时间顺序依次截取[X(t−15),X(t−14),...,X(t)]作为输入，X(t+1)作为输出，形成单个样本{X→Y}；其中X为四维向量[P_max_norm,P_min_norm,E_max_norm,E_min_norm]，Y为下一时刻的四维预测目标。通过循环滑动，整个时段可被切分为N−16个样本（N为总采样点数），满足深度神经网络对时序依赖建模的需求。该滑动窗口算法实现简单、计算开销低，可直接嵌入时序数据集构建模块，为后续Transformer模型提供标准化训练样本。

[0069] 将输入输出样本对输入预测模型模块，以启动资源聚合能力的预测步骤。

[0070] 示例性地，预测模型模块可以基于轻量级Transformer编码器（一种仅含自注意力与前馈网络、无解码器的深度学习单元），层数2、注意力头数4、隐藏维度64，接收样本对{X→Y}后执行前向推理，输出未来时段的归一化聚合边界预测向量，供后续反归一化与调度决策使用。

[0071] 因此，根据上述实施方式，系统能够将原始边界数据转化为标准化样本，并通过深度学习框架完成超短期演化趋势推理，实现对微电网资源聚合能力的在线预测。

[0072] 在一些实施例中，超短期预测框架对整体聚合可行区间的未来动态演化趋势进行预测，输出预测聚合可行区间的步骤，包括：通过预测模型模块对输入输出样本对进行特征提取，得到高维时序特征。

[0073] 具体而言，系统可以令预测模型模块采用双层Transformer编码器结构，每层包含4头自注意力机制，隐藏维度64；输入输出样本对经嵌入层后形成16×64维张量，经自注意力计算后输出平均池化向量，即高维时序特征，维度64。

[0074] 将高维时序特征映射为预测聚合可行区间的聚合功率上边界、聚合功率下边界、聚合能量上边界及聚合能量下边界。

[0075] 接着，系统可以将高维时序特征映射为预测聚合可行区间的聚合功率上边界、聚合功率下边界、聚合能量上边界及聚合能量下边界。

[0076] 具体而言，映射通过全连接层实现，权重矩阵64×4，偏置向量4维，输出向量记为[P_max_pred,P_min_pred,E_max_pred,E_min_pred]，随后经逆MinMaxScaler还原为实际量纲，单位为千瓦或千瓦时。

[0077] 将聚合功率上边界、聚合功率下边界、聚合能量上边界及聚合能量下边界封装为预测聚合可行区间并输出。

[0078] 具体而言，系统的封装格式可以采用JSON对象，键名依次为"power_max","power_min","energy_max","energy_min"，数值保留两位小数，通过MQTT（Message QueuingTelemetry Transport，消息队列遥测传输协议）主题"mg / forecast"发布，供调度主站订阅。

[0079] 因此，根据上述实施方式，系统能够将历史边界序列转化为未来四维边界预测，完成资源聚合可行区间的超短期输出，为后续个体功率分配方案生成提供高置信度输入。

[0080] 在一些实施例中，在输出预测聚合可行区间之后，上述方法还包括：基于预测聚合可行区间中的聚合功率上边界、聚合功率下边界、聚合能量上边界及聚合能量下边界，构建聚合可行域分解模型。

[0081] 其中，聚合可行域分解模型可以指以四维边界为约束空间、以功率-能量耦合关系为规则的数学模型，用于将聚合曲线逆向映射至各设备功率曲线。

[0082] 在预测聚合可行区间内，通过聚合可行域分解模型对预测聚合可行区间进行约束保持型采样，生成可行聚合曲线。

[0083] 具体而言，系统在进行约束保持型采样时，可以采用等间隔均匀抽样，功率步长取50千瓦，能量步长取25千瓦时，每一步均校验累积能量是否位于能量上下边界之间，若超出则丢弃并重新抽样，直至形成连续的可行聚合曲线。

[0084] 以最小化功率偏差为目标，将可行聚合曲线分解为分布式电源、柔性负荷及储能装置的个体功率曲线。

[0085] 具体而言，系统以通过下述的公式化条件创建分解模型：；；；；；其中，、分别为时刻聚合功率上偏差与下偏差；为时刻采样的功率；为资源r在时刻的分解功率。

[0086] 所述绝对误差评价指标：；所述相对误差评价指标：；根据分布式电源、柔性负荷及储能装置的个体功率曲线生成个体功率分配方案。

[0087] 具体而言，分布式电源、柔性负荷及储能装置的各自的个体功率分配方案的格式可以采用JSON数组，数组元素包含设备ID、出力序列、起止时间戳，通过MQTT主题"mg / dispatch"下发至就地测控单元，实现闭环控制。

[0088] 因此，根据上述实施方式，系统能够将预测得到的聚合可行区间精准拆解为每台设备的可执行出力计划，完成从聚合预测到个体调度的无缝衔接。

[0089] 在一些实施例中，系统分别采用LSTM（Long Short-Term Memory，长短期记忆网络）、GRU（Gated Recurrent Unit，门控循环单元）、CNN（Convolutional Neural Network，卷积神经网络）及Transformer模型的这四种神经网络架构，对同一微电网可调资源聚合可行域的功率-能量边界进行预测，并在相同测试集上对比预测精度。对比结果以均方误差MSE（Mean Squared Error，均方误差）作为评价指标，MSE数值越低，表示预测边界与真实边界的偏差越小，预测精度越高。

[0090] 其中，系统可以利用如图3所示的Transformer模型，对滑动窗口生成的输入输出样本对进行特征提取与预测。该模型采用纯编码器（Encoder-Only）架构，核心组件包括：输入层，用于接收四维归一化边界向量如下所示：[P_max_norm,P_min_norm,E_max_norm,E_min_norm]；并通过线性投影映射到64维隐藏空间。

[0091] 位置编码（Positional Encoding）：使用正弦-余弦函数为16个时间步的向量序列添加顺序信息，保持时序顺序敏感性。

[0092] 多头注意力机制（Multi-Head Attention）：头数设为4，每头维度16；通过并行计算自注意力权重，捕捉任意两个时间步之间的长程依赖关系。

[0093] 归一化与残差连接：在子层输出后依次执行层归一化（LayerNorm）与残差相加，缓解深层梯度退化问题。

[0094] 前馈网络（Feed-Forward Network）：由两层全连接组成：隐藏维度64→256→64，激活函数采用ReLU，用于增强非线性表达能力。

[0095] 输出层：将编码后的高维时序特征经线性变换压缩为4维向量，对应的预测聚合可行区间为：[P_max_pred,P_min_pred,E_max_pred,E_min_pred]。

[0096] 模型采用2层堆叠的轻量级Transformer编码器，参数量约37k（千），可在边缘计算节点实现单次前向推理延迟低于200毫秒，满足实时预测需求。该Transformer结构通过并行化注意力计算，提升了微电网聚合功率-能量边界动态演化的捕捉精度，为后续可行域分解与个体功率分配提供高置信度预测结果。

[0097] 如图4所示，各模型在训练阶段的损失曲线均呈单调下降趋势，其中Transformer模型的训练损失下降速率最快，且在150轮次后趋于稳定，最终损失值为0.03（均方误差，MSE，样本量N=8640），显著低于同测试集下LSTM模型的0.20（相同损失函数、相同N），GRU（0.15）及CNN（0.08）。这表明Transformer模型在捕捉聚合边界动态演化特征方面具有更强的拟合能力。

[0098] 如图5所示，在测试集上的MSE对比结果进一步验证了上述结论。Transformer模型的MSE为0.00026（即0.026%），远低于LSTM（0.00060，即0.060%）、GRU（0.00047，即0.047%）及CNN（0.00039，即0.039%）。因此，采用Transformer架构的超短期预测框架在微电网聚合功率-能量可行域预测任务中具有更高的预测精度，可为后续调度决策提供更可靠的边界依据。

[0099] 图6是本发明一实施例的微电网的资源聚合能力预测系统的结构框图。

[0100] 如图6所示，该微电网的资源聚合能力预测系统，包括：多维负载数据获取模块210，用于获取微电网的源荷储负载数据，源荷储负载数据包括分布式电源、柔性负荷及储能装置的运行状态数据。

[0101] 聚合可行区间设定模块220，用于根据分布式电源的个体可行区间、柔性负荷的个体可行区间以及储能装置的个体可行区间构建整体聚合可行区间，各个体可行区间由对应的运行状态数据确定。

[0102] 能量边界模型创建模块230，用于根据整体聚合可行区间创建聚合功率能量边界模型。

[0103] 短期聚合能力预测模块240，用于将聚合功率能量边界模型输入预设的深度神经网络的超短期预测框架，由超短期预测框架对整体聚合可行区间的未来动态演化趋势进行预测，输出预测聚合可行区间。

[0104] 本发明实施例的装置的各模块、子模块的具体功能和示例的描述，可以参见上述方法实施例中对应步骤的相关描述，在此不再赘述。

[0105] 根据本发明的实施例，本发明的上述方法可以应用到一种电子设备和一种可读存储介质中。

[0106] 图7示出了可以用来实施本发明实施例的电子设备600的示意性框图。电子设备旨在表示各种形式的数字计算机，诸如，膝上型计算机、台式计算机、工作台、个人数字助理、服务器、刀片式服务器、大型计算机和其他适合的计算机。电子设备还可以表示各种形式的移动装置，诸如，个人数字助理、蜂窝电话、智能电话、可穿戴设备和其他类似的计算装置。本文所示的部件、它们的连接和关系、以及它们的功能仅仅作为示例，并且不意在限制本文中描述的和 / 或者要求的本发明的实现。

[0107] 如图7所示，电子设备600包括计算单元601，其可以根据存储在只读存储器（ROM）602中的计算机程序或者从存储单元608加载到随机访问存储器（RAM）603中的计算机程序，来执行各种适当的动作和处理。在RAM603中，还可存储电子设备600操作所需的各种程序和数据。计算单元601、ROM602以及RAM603通过总线604彼此相连。输入 / 输出（I / O）接口605也连接至总线604。

[0108] 电子设备600中的多个部件连接至I / O接口605，包括：输入单元606，例如键盘、鼠标等；输出单元607，例如各种类型的显示器、扬声器等；存储单元608，例如磁盘、光盘等；以及通信单元609，例如网卡、调制解调器、无线通信收发机等。通信单元609允许电子设备600通过诸如因特网的计算机网络和 / 或各种电信网络与其他设备交换信息 / 数据。

[0109] 计算单元601可以是各种具有处理和计算能力的通用和 / 或专用处理组件。计算单元601的一些示例包括但不限于中央处理单元（CPU）、图形处理单元（GPU）、各种专用的人工智能（AI）计算芯片、各种运行机器学习模型算法的计算单元、数字信号处理器（DSP）、以及任何适当的处理器、控制器、微控制器等。计算单元601执行上文所描述的各个方法和处理，例如一种微电网的资源聚合能力预测方法。例如，在一些实施例中，一种微电网的资源聚合能力预测方法可被实现为计算机软件程序，其被有形地包含于机器可读介质，例如存储单元608。在一些实施例中，计算机程序的部分或者全部可以经由ROM602和 / 或通信单元609而被载入和 / 或安装到电子设备600上。当计算机程序加载到RAM603并由计算单元601执行时，可以执行上文描述的一种微电网的资源聚合能力预测方法的一个或多个步骤。备选地，在其他实施例中，计算单元601可以通过其他任何适当的方式（例如，借助于固件）而被配置为执行一种微电网的资源聚合能力预测方法。

[0110] 本文中以上描述的系统和技术的各种实施方式可以在数字电子电路系统、集成电路系统、现场可编程门阵列（FPGA）、专用集成电路（ASIC）、专用标准产品（ASSP）、芯片上系统的系统（SOC）、负载可编程逻辑设备（CPLD）、计算机硬件、固件、软件、和 / 或它们的组合中实现。这些各种实施方式可以包括：实施在一个或者多个计算机程序中，该一个或者多个计算机程序可在包括至少一个可编程处理器的可编程系统上执行和 / 或解释，该可编程处理器可以是专用或者通用可编程处理器，可以从存储系统、至少一个输入装置、和至少一个输出装置接收数据和指令，并且将数据和指令传输至该存储系统、该至少一个输入装置、和该至少一个输出装置。

[0111] 用于实施本发明的方法的程序代码可以采用一个或多个编程语言的任何组合来编写。这些程序代码可以提供给通用计算机、专用计算机或其他可编程数据处理装置的处理器或控制器，使得程序代码当由处理器或控制器执行时使流程图和 / 或框图中所规定的功能 / 操作被实施。程序代码可以完全在机器上执行、部分地在机器上执行，作为独立软件包部分地在机器上执行且部分地在远程机器上执行或完全在远程机器或服务器上执行。

[0112] 在本发明的上下文中，机器可读介质可以是有形的介质，其可以包含或存储以供指令执行系统、装置或设备使用或与指令执行系统、装置或设备结合地使用的程序。机器可读介质可以是机器可读信号介质或机器可读储存介质。机器可读介质可以包括但不限于电子的、磁性的、光学的、电磁的、红外的、或半导体系统、装置或设备，或者上述内容的任何合适组合。机器可读存储介质的更具体示例会包括基于一个或多个线的电气连接、便携式计算机盘、硬盘、随机存取存储器（RAM）、只读存储器（ROM）、可擦除可编程只读存储器（EPROM或快闪存储器）、光纤、便捷式紧凑盘只读存储器（CD-ROM）、光学储存设备、磁储存设备、或上述内容的任何合适组合。

[0113] 为了提供与用户的交互，可以在计算机上实施此处描述的系统和技术，该计算机具有：用于向用户显示信息的显示装置（例如，CRT（阴极射线管）或者LCD（液晶显示器）监视器）；以及键盘和指向装置（例如，鼠标或者轨迹球），用户可以通过该键盘和该指向装置来将输入提供给计算机。其它种类的装置还可以用于提供与用户的交互；例如，提供给用户的反馈可以是任何形式的传感反馈（例如，视觉反馈、听觉反馈、或者触觉反馈）；并且可以用任何形式（包括声输入、语音输入、或者触觉输入）来接收来自用户的输入。

[0114] 可以将此处描述的系统和技术实施在包括后台部件的计算系统（例如，作为数据服务器）、或者包括中间件部件的计算系统（例如，应用服务器）、或者包括前端部件的计算系统（例如，具有图形用户界面或者网络浏览器的用户计算机，用户可以通过该图形用户界面或者该网络浏览器来与此处描述的系统和技术的实施方式交互）、或者包括这种后台部件、中间件部件、或者前端部件的任何组合的计算系统中。可以通过任何形式或者介质的数字数据通信（例如，通信网络）来将系统的部件相互连接。通信网络的示例包括：局域网（LAN）、广域网（WAN）和互联网。

[0115] 计算机系统可以包括客户端和服务器。客户端和服务器一般远离彼此并且通常通过通信网络进行交互。通过在相应的计算机上运行并且彼此具有客户端-服务器关系的计算机程序来产生客户端和服务器的关系。服务器可以是云服务器，也可以为分布式系统的服务器，或者是结合了区块链的服务器。

[0116] 应该理解，可以使用上面所示的各种形式的流程，重新排序、增加或删除步骤。例如，本发明中记载的各步骤可以并行地执行也可以顺序地执行也可以不同的次序执行，只要能够实现本发明公开的技术方案所期望的结果，本文在此不进行限制。

[0117] 上述具体实施方式，并不构成对本发明保护范围的限制。本领域技术人员应该明白的是，根据设计要求和其他因素，可以进行各种修改、组合、子组合和替代。任何在本发明的原则之内所作的修改、等同替换和改进等，均应包含在本发明保护范围之内。

Claims

1. A method for predicting the resource aggregation capability of a microgrid, characterized in that, include: Acquire source-load-storage load data of the microgrid, including the operating status data of distributed power sources, flexible loads, and energy storage devices; An overall aggregated feasible range is constructed based on the individual feasible ranges of the distributed power source, the individual feasible range of the flexible load, and the individual feasible range of the energy storage device. Each individual feasible range is determined by the corresponding operating status data. An aggregated power energy boundary model is created based on the overall aggregated feasible range. The aggregated power energy boundary model is generated based on the power upper bound, power lower bound, energy upper bound, and energy lower bound of the overall aggregated feasible range. It includes the aggregated power upper bound and aggregated power lower bound used to define the total power feasible range of the microgrid within the prediction period, and the aggregated energy upper bound and aggregated energy lower bound used to define the total energy feasible range. The aggregated power energy boundary model is input into a preset deep neural network ultra-short-term prediction framework, which includes a time series dataset construction module and a prediction model module based on a Transformer encoder structure. The aggregate power energy boundary model is preprocessed to generate normalized aggregate power upper boundary, aggregate power lower boundary, aggregate energy upper boundary, and aggregate energy lower boundary data. The normalized boundary data is input into the time series dataset construction module, and input-output sample pairs are constructed through a sliding window mechanism. The input-output sample pair is input into the prediction model module, and high-dimensional temporal features are extracted by the Transformer encoder. The high-dimensional temporal features are then mapped to the upper boundary of the aggregated power, the lower boundary of the aggregated power, the upper boundary of the aggregated energy, and the lower boundary of the aggregated energy of the predicted feasible interval. The ultra-short-term prediction framework encapsulates the upper boundary of the aggregation power, the lower boundary of the aggregation power, the upper boundary of the aggregation energy, and the lower boundary of the aggregation energy into the predicted aggregation feasible interval and outputs it. Within the predicted feasible aggregation interval, constraint-preserving sampling is performed using an equally spaced uniform sampling method. At each step, it is verified whether the accumulated energy is between the upper and lower energy boundaries to generate a feasible aggregation curve.

2. The method according to claim 1, characterized in that, The method further includes: Based on the source-load-storage load data, the individual power boundaries and individual energy boundaries of the distributed power source, the flexible load, and the energy storage device are constructed respectively. A source-load-storage power-energy boundary model is created based on each of the individual power boundaries and each of the individual energy boundaries; In the source-load-storage power energy boundary model: The individual power boundary and individual energy boundary of the distributed power source are used to characterize the feasible mapping range between the output power and the cumulative power generation of the distributed power source. The individual power boundary and individual energy boundary of the flexible load are used to characterize the feasible mapping range between the power consumption and the cumulative power consumption of the flexible load. The individual power boundary and individual energy boundary of the energy storage device are used to characterize the feasible mapping range between the charging and discharging power and the range of state of charge variation of the energy storage device.

3. The method according to claim 1, characterized in that, The construction of an overall aggregated feasible region based on the individual feasible regions of the distributed power source, the flexible load, and the energy storage device includes: Extract the individual power boundary and individual energy boundary of the distributed power source, the flexible load, and the energy storage device respectively; By superimposing the individual power boundaries of the distributed power source, the flexible load, and the energy storage device, the upper and lower power boundaries of the overall aggregated feasible interval are obtained. By superimposing the individual energy boundaries of the distributed power source, the flexible load, and the energy storage device, the upper and lower bounds of the overall aggregate feasible interval are calculated. The overall aggregate feasible interval is constructed based on the power upper bound, the power lower bound, the energy upper bound, and the energy lower bound of the overall aggregate feasible interval.

4. The method according to claim 1, characterized in that, After outputting the predicted aggregate feasible interval, the method further includes: Based on the upper boundary of aggregation power, the lower boundary of aggregation power, the upper boundary of aggregation energy, and the lower boundary of aggregation energy in the predicted aggregation feasible interval, an aggregation feasible region decomposition model is constructed; The aggregated feasible region decomposition model is used to perform the constraint-preserving sampling step within the predicted aggregated feasible interval; With the goal of minimizing power deviation, the feasible aggregation curve is decomposed into individual power curves of the distributed power source, the flexible load, and the energy storage device; Individual power allocation schemes are generated based on the individual power curves of the distributed power source, the flexible load, and the energy storage device.

5. A resource aggregation capability prediction system for a microgrid, characterized in that, include: The multi-dimensional load data acquisition module is used to acquire the source-load-storage load data of the microgrid, which includes the operating status data of distributed power sources, flexible loads and energy storage devices. The aggregated feasible range setting module is used to construct an overall aggregated feasible range based on the individual feasible ranges of the distributed power source, the individual feasible ranges of the flexible load, and the individual feasible ranges of the energy storage device, wherein each individual feasible range is determined by the corresponding operating status data. The energy boundary model creation module is used to create an aggregated power energy boundary model based on the overall aggregated feasible interval. The aggregated power energy boundary model is generated based on the power upper bound, power lower bound, energy upper bound, and energy lower bound of the overall aggregated feasible interval. It includes aggregated power upper bound and aggregated power lower bound for defining the total power feasible range of the microgrid within the prediction period, and aggregated energy upper bound and aggregated energy lower bound for defining the total energy feasible range. The short-term aggregation capability prediction module is used to input the aggregation power energy boundary model into a preset deep neural network ultra-short-term prediction framework. The ultra-short-term prediction framework includes a time series dataset construction module and a prediction model module based on the Transformer encoder structure. The aggregate power energy boundary model is preprocessed to generate normalized aggregate power upper boundary, aggregate power lower boundary, aggregate energy upper boundary, and aggregate energy lower boundary data. The normalized boundary data is input into the time series dataset construction module, and input-output sample pairs are constructed through a sliding window mechanism. The input-output sample pair is input into the prediction model module, and high-dimensional temporal features are extracted by the Transformer encoder. The high-dimensional temporal features are then mapped to the upper boundary of the aggregated power, the lower boundary of the aggregated power, the upper boundary of the aggregated energy, and the lower boundary of the aggregated energy of the predicted feasible interval. The ultra-short-term prediction framework encapsulates the upper boundary of the aggregation power, the lower boundary of the aggregation power, the upper boundary of the aggregation energy, and the lower boundary of the aggregation energy into the predicted aggregation feasible interval and outputs it. Within the predicted feasible aggregation interval, constraint-preserving sampling is performed using an equally spaced uniform sampling method. At each step, it is verified whether the accumulated energy is between the upper and lower energy boundaries to generate a feasible aggregation curve.

Images