Intelligent photovoltaic grid-connected regulation method and system

Abstract

The application discloses an intelligent photovoltaic grid-connected regulation method and system, relates to the photovoltaic grid-connected regulation field, discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a kind of intelligent photovoltaic grid-connected regulation method and system, and discloses a

Description

TECHNICAL FIELD

[0001] The present application relates to the field of photovoltaic grid-connected regulation, and more particularly, to an intelligent photovoltaic grid-connected regulation method and system. BACKGROUND

[0002] With the transformation of global energy structure, new energy generation forms represented by photovoltaic are developing at an unprecedented speed. However, the inherent properties of photovoltaic generation, i.e. the intermittency and volatility of its output power caused by weather conditions (especially cloud shadow), bring serious challenges to the safe and stable operation of the power grid. When large-scale cloud clusters move rapidly and shadow large photovoltaic power stations, it will cause a sharp power drop in a short time. This rapid power change rate, i.e. power ramping event, will seriously impact the stability of power grid frequency and voltage if not effectively regulated. Therefore, how to intelligently and smoothly regulate the grid-connected power of photovoltaic power stations has become a key technical problem to be solved in this field.

[0003] To address this challenge, existing technical solutions mainly develop in two directions. One is to rely on expensive external equipment for prediction, such as deploying all-sky cameras, weather radars or satellite cloud map analysis systems to see the movement trajectory of clouds, so as to predict their impact on power station power generation. This solution can provide certain prediction information, but the equipment investment and later operation and maintenance cost are high, and the data processing is complex, which is difficult to popularize in all power stations on a large scale. Another solution is to rely heavily on energy storage systems for passive compensation, i.e. when power drop is detected, energy storage is quickly called to fill in and achieve the principle of filling in as much as possible. The fundamental defect of this mode lies in its reaction lag, the power impact has already occurred, and the control action is started, which cannot fundamentally eliminate the impact itself. At the same time, this way puts forward very high requirements for the capacity and response speed of energy storage, leading to excessive dependence on energy storage and increasing the overall cost of the system. In addition, traditional methods often ignore the power distribution difference within large photovoltaic power stations, i.e. the spatial inconsistency caused by the movement of cloud shadow within the power station, and cannot achieve fine-grained partitioned collaborative control.

[0004] In view of this, it is urgent to propose a new intelligent photovoltaic grid-connected regulation scheme. SUMMARY

[0005] To solve the defects of the existing scheme, according to an aspect of the present application, a kind of intelligent photovoltaic grid-connected regulation method is provided, it includes: based on power station zoning map, the total power of each sub-region is obtained by aggregating inverter original power data stream;The total power of each sub-region is normalized to obtain the normalized power map consisting of the normalized power indicator of each sub-region;Based on power station zoning map, cloud shadow event detection and motion vector modeling are carried out to the time series of normalized power map to obtain cloud motion vector field;Based on cloud motion vector field and the normalized power map of current time step, downstream area power impact prediction is carried out to obtain all sub-region impact prediction;The spatiotemporal transfer strategy optimization under global power target is carried out to the all sub-region impact prediction to obtain optimal scheduling plan;Optimal scheduling plan is decomposed and issued to scheduling instruction and closed-loop correction to obtain inverter power set value.

[0006] According to another aspect of the present application, an intelligent photovoltaic grid-connected regulation system is provided, which includes: a power data stream aggregation module for aggregating inverter original power data stream based on power station zoning map to obtain the total power of each sub-region;A power data normalization module for normalizing the total power of each sub-region to obtain a normalized power map consisting of the normalized power indicator of each sub-region;A cloud motion vector field generation module for detecting cloud shadow events and modeling motion vectors based on power station zoning map to obtain cloud motion vector field by time series of normalized power map;An impact prediction module for predicting downstream area power impact based on cloud motion vector field and the normalized power map of current time step to obtain all sub-region impact prediction;An optimal scheduling module for optimizing spatiotemporal transfer strategy under global power target to the all sub-region impact prediction to obtain optimal scheduling plan;An inverter power setting module for decomposing and issuing scheduling instruction to optimal scheduling plan and closed-loop correction to obtain inverter power set value.

[0007] Compared with the prior art, the intelligent photovoltaic grid-connected regulation method and system provided by the application converts the photovoltaic power station into a large distributed cloud shadow sensor network, discards the expensive external observation equipment, and instead deeply excavates the power data of the inverter inside the power station. Specifically, by aggregating and normalizing the real-time power data stream, a power map capable of intuitively reflecting the position and intensity of the cloud shadow is constructed. Based on the time series analysis of the map, the moving track of the power valley (i.e. the cloud shadow) is detected and tracked, so as to establish a cloud motion vector field model and realize accurate prediction of the future power impact. This active prediction mechanism solves the control lag problem of the traditional scheme. Finally, through global space-time optimization scheduling, the predicted centralized power impact is distributed in advance and smoothly to the inverters in different areas of the power station, effectively solving the spatial inconsistency problem, realizing the transformation from passive lag compensation to active prediction and collaborative regulation, and greatly improving the grid friendliness. BRIEF DESCRIPTION OF DRAWINGS

[0008] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application taken in conjunction with the accompanying drawings.

[0009] Figure 1 is a flow chart of the intelligent photovoltaic grid-connected regulation method according to the embodiment of the application.

[0010] Figure 2 is a data flow diagram of the intelligent photovoltaic grid-connected regulation method according to the embodiment of the application.

[0011] Figure 3 is a flow chart of step S3 in the intelligent photovoltaic grid-connected regulation method according to the embodiment of the application.

[0012] Figure 4 is a block diagram of the intelligent photovoltaic grid-connected regulation system according to the embodiment of the application. DETAILED DESCRIPTION

[0013] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. It should be understood that the accompanying drawings and embodiments of the present disclosure are only for exemplary purposes, and are not intended to limit the protection scope of the present disclosure.

[0014] In the face of the above-mentioned defects in the prior art, the application provides an intelligent photovoltaic grid-connected regulation method. Figure 1 is a flow chart of the intelligent photovoltaic grid-connected regulation method according to the embodiment of the application. Figure 2 is a data flow diagram of the intelligent photovoltaic grid-connected regulation method according to the embodiment of the application. As Figure 1 and Figure 2As shown, the intelligent photovoltaic grid-connected control method according to an embodiment of this application includes: Step S1, aggregating the original power data stream of the inverter based on the power plant partition map to obtain the total power of each sub-region; Step S2, normalizing the total power of each sub-region to obtain a normalized power map composed of the normalized power index of each sub-region; Step S3, performing cloud shadow event detection and motion vector modeling on the time series of the normalized power map based on the power plant partition map to obtain a cloud motion vector field; Step S4, performing downstream region power impact prediction based on the cloud motion vector field and the normalized power map at the current time step to obtain impact prediction for all sub-regions; Step S5, optimizing the spatiotemporal transfer strategy under the global power target for the impact prediction for all sub-regions to obtain the optimal scheduling plan; Step S6, performing scheduling instruction decomposition and closed-loop correction on the optimal scheduling plan to obtain the inverter power setpoint.

[0015] In step S1, based on the power plant zoning map, the raw power data stream of the inverters is aggregated to obtain the total power of each sub-region. It should be understood that large photovoltaic power plants cover a vast area, with tens of thousands of inverters within them. When clouds move, their shading effect exhibits significant spatial inconsistencies within the power plant. Directly analyzing the massive, discrete power data of individual inverters would not only present a huge computational burden but also make it difficult to effectively extract macroscopic, directional cloud shadow movement trends from the complex details. Therefore, this application aggregates the raw power data stream of the inverters based on the power plant zoning map to obtain the total power of each sub-region, transforming the raw, high-granularity data points into structured data units with clear geospatial attributes that better reflect local overall power changes. This lays a solid and efficient data foundation for subsequent accurate identification of cloud shadow events and modeling of their motion vector fields.

[0016] Optionally, the implementation of step S1 is as follows: To achieve the aggregation of the raw power data stream of inverters, first a power plant zoning map needs to be constructed and loaded. This map is a pre-defined static data structure, generated in the planning and design phase or digital modeling phase of the power plant. It divides the entire power plant into a number of geographically continuous sub-regions according to the physical layout of the power plant, the electrical wiring such as the ownership of combiner boxes or box-type transformers, and the management unit. For example, a 100-megawatt photovoltaic power plant can be divided into 20 sub-regions, numbered Z01 to Z20, each corresponding to a 5-megawatt installed capacity. The specific form of the power plant zoning map can be a mapping table, which details the unique identifier (ID) of each inverter and its corresponding sub-region number. For example, the map contains corresponding relationships such as (inverter ID: INV0001, sub-region ID: Z01), (inverter ID: INV0002, sub-region ID: Z01),..., (inverter ID: INV2500, sub-region ID: Z20).

[0017] At the same time, the raw power data stream of inverters needs to be obtained in real time. This data stream is provided by the power plant's supervisory control and data acquisition (SCADA) facility. The data acquisition unit deployed on site will read the current active power value on the DC or AC side of each inverter in the field at a fixed, high-frequency time interval, such as every 5 seconds. These data, attached with precise timestamps and inverter IDs, form a continuous data stream. At any sampling time point T, this data stream appears as a data set, for example: {(timestamp: T, inverter ID: INV0001, power: 45.2 kW), (timestamp: T, inverter ID: INV0002, power: 46.1 kW),...}.

[0018] The aggregation process is triggered to execute at each sampling time point T. The processing program iterates through all sub-regions defined in the power plant zoning map, i.e. from Z01 to Z20. For each sub-region, for example Z01, the program first queries the list of all inverter IDs belonging to this region according to the zoning map. Then, the program retrieves the power values corresponding to these IDs in the raw power data stream of inverters at the current time point T, and performs an arithmetic sum on them. For example, if sub-region Z01 contains inverters INV0001 to INV125, then the total power of Z01 at time T is P_INV0001(T) + P_INV0002(T) +... + P_INV125(T). This calculation process is performed sequentially for all sub-regions.

[0019] Finally, the output at each time point T is a data structure containing the total power of all sub-regions. Its form is a list of power indexed by time stamp T, for example: {(time stamp: T, sub-region ID: Z01, total power: 4.8 MW), (time stamp: T, sub-region ID: Z02, total power: 4.9 MW),..., (time stamp: T, sub-region ID: Z20, total power: 2.1 MW)}.

[0020] In step S2, the total power of each sub-region is normalized to obtain a normalized power map composed of normalized power indicators of each sub-region. Accordingly, the total power of each sub-region is itself affected by various non-cloud factors such as solar zenith angle, atmospheric transparency, seasonal change, etc. For example, at noon on a clear day, the power output of a sub-region is much higher than its output in the morning or evening; similarly, the clear sky power in summer is significantly higher than in winter. If the cloud shadow event is directly judged based on the fluctuation of such original total power, it is easy to produce a false judgment and cannot separate the key disturbance factor of cloud cover from the complex background changes. Therefore, it is necessary to normalize the total power of each sub-region to eliminate all predictable power baseline changes determined by the illumination geometry and astronomical period, thereby generating a standardized indicator that only reflects the instantaneous meteorological condition (especially the amount of cloud cover) influence, the normalized power index, so that the affected degree of different time and different sub-regions has comparability.

[0021] In an optional implementation, step S2, the total power of each sub-region is normalized to obtain a normalized power map composed of normalized power indicators of each sub-region, comprising: step S21, inputting the ID and time of each sub-region into the clear sky power model to obtain the theoretical clear sky power of each sub-region; step S22, based on the theoretical clear sky power of each sub-region, the total power of each sub-region is normalized to obtain the normalized power map.

[0022] In the optional implementation manner, the implementation process of step S2 is as follows: first, step S21 is performed. The clear-sky power model is a deterministic calculation model based on physical principles, which integrates the knowledge of astronomy, atmospheric physics and photovoltaic engineering, and mainly consists of a solar position algorithm module, an atmospheric radiation transmission module and a photovoltaic array performance module. Taking the calculation of the theoretical clear-sky power of the sub-region Z01 at the time point T (time stamp: T, sub-region ID: Z01, total power: 4.8 MW) as an example. First, the program processes the sub-region ID Z01 and the time stamp T as inputs. After the model receives the ID, it queries the physical properties related to Z01 from a preset power station static parameter database. The database is part of the digital construction of the power station, which contains the accurate central geographical longitude and latitude of each sub-region, for example, longitude 116.4°E and latitude 39.9°N, the total rated installed capacity of all photovoltaic components in the region, for example, 5 MW, the unified tilt angle of the photovoltaic array, for example, 28 degrees, and the azimuth angle, for example, 0 degrees, that is, facing south, and a comprehensive performance coefficient, for example, 0.97, which comprehensively reflects the fixed efficiency loss caused by the aging attenuation of components due to long-term operation, line loss and dust shielding, etc. Then, the solar position algorithm module is started. The module receives the time stamp T and the longitude and latitude coordinates of Z01. It calculates the specific position of the sun in the sky at that time and at that location according to the standard astronomical algorithm such as the SPA algorithm. Its output is two key angles: the solar zenith angle, that is, the angle between the sun's rays and the normal line of the ground; and the solar azimuth angle, that is, the direction angle of the sun in the horizontal plane. For example, at time T, the model may calculate that the solar zenith angle is 35 degrees and the azimuth angle is 150 degrees.

[0023] Then, the calculated solar position angles are passed to an atmospheric radiation transfer module built based on a neural network. This module is a pre-trained deep feedforward neural network that accurately simulates the penetration of solar radiation through the atmosphere and its final intensity on the plane of a specific tilt photovoltaic array under ideal clear sky conditions in a data-driven manner. The input layer of this neural network consists of three neurons that receive the solar zenith angle, the solar azimuth angle, and the annual day number representing the seasonal atmospheric variations, for example, from 1 to 365. These input features collectively provide the model with the core spatiotemporal information needed for prediction. This is followed by two hidden layers, for example, the first hidden layer contains 64 neurons, and the second hidden layer contains 32 neurons, each using a rectified linear unit (ReLU) as the activation function to give the network powerful nonlinear fitting capabilities, enabling it to learn and express the highly complex relationships between solar position, season, and final radiation intensity. The output layer of the network is a single neuron with a linear activation function, and its only output value is the predicted plane global irradiance. This neural network undergoes a rigorous offline training phase before deployment. During the training process, the solar position angles and annual day numbers from historical data are used as inputs, and the corresponding, quality-controlled measured values of clear sky plane global irradiance are used as the target output. Through the backpropagation algorithm and optimizers such as the Adam optimizer, the thousands of weight and bias parameters in the network are continuously adjusted to minimize the mean squared error between the network's predicted values and the true measured values. After training is complete, these fixed weights and bias parameters contain the complex rules of clear sky radiation transfer at that specific location, forming an efficient and accurate calculation model. In actual operation, the trained model performs fast forward propagation calculations. When the solar zenith angle at time T is 35 degrees, the solar azimuth angle is 150 degrees, and the corresponding annual day number is input into the network, the data flows through each layer. Each neuron performs a weighted sum of the signals it receives and adds a bias, then performs a nonlinear transformation through the activation function, passes it layer by layer until the output layer. Finally, the output neuron gives an accurate numerical value. For example, after a series of matrix operations that contain a large amount of historical clear sky data rules, the model finally outputs a plane global irradiance of 920 watts per square meter. Finally, the photovoltaic array performance module performs the final power conversion calculation. It receives the plane global irradiance of 920 watts per square meter calculated by the previous module and combines the rated installed capacity of Z01 of 5 megawatts and the comprehensive performance coefficient of 0.97 found in the database. The calculation formula is: theoretical clear sky power = (plane global irradiance / standard test condition irradiance) x rated installed capacity x comprehensive performance coefficient. Among them, the standard test condition irradiance is an industry-recognized constant, which is 1000 watts per square meter. Therefore, the theoretical clear sky power of Z01 at time T is calculated as: (920 / 1000) x 5 megawatts x 0.97 ≈ 4.46 megawatts.This complete calculation process will be repeated for all sub-regions Z02, Z03, Z20, etc. at the same time point T, because the slight difference in position or performance coefficient of each sub-region may cause a slight difference in its theoretical clear sky power. The final output is a list of theoretical clear sky powers corresponding to the input sub-region list, in the form of: {(timestamp: T, sub-region ID: Z01, theoretical clear sky power: 4.46 MW), (timestamp: T, sub-region ID: Z02, theoretical clear sky power: 4.47 MW),...}.

[0024] Next, step S22 is performed. As for sub-region Z01, the program extracts its actual total power 4.8 MW and theoretical clear sky power 4.46 MW at time T from the input list. The normalization calculation is simply the division of the two: normalized power index = 4.8 / 4.46 ≈ 1.076. This index is greater than 1, which may reflect the cloud edge enhancement effect or a slight deviation in model calibration. For sub-region Z20, if its total power is 2.1 MW and the corresponding theoretical clear sky power is 4.40 MW, then its normalized power index is 2.1 / 4.40 ≈ 0.477, which is significantly lower than 1, indicating that this region is being shaded by clouds. To avoid calculation errors due to the theoretical power approaching zero at sunrise and sunset, a power threshold is preset, and when the theoretical clear sky power is lower than this value, the normalized index is set to invalid or a default value. After the calculation of all sub-regions is completed, the final output is the normalized power map at time point T, with the data structure: {(timestamp: T, sub-region ID: Z01, normalized power index: 1.076),..., (timestamp: T, sub-region ID: Z20, normalized power index: 0.477)}.

[0025] In step S3, based on the power station partition map, cloud shadow event detection and motion vector modeling are performed on the time series of the normalized power map to obtain the cloud motion vector field. It can be understood that the time series of the normalized power map is a series of continuous snapshots taken of the photovoltaic power station, accurately recording the spatial distribution of cloud shadows at each instant. However, these static snapshots themselves do not contain direct prediction information and cannot tell the control center how fast and in which direction the cloud shadows will move. In order to solve the control lag problem of the traditional scheme and realize the transition from passive response to active pre-control, it is necessary to understand the dynamic evolution law of the cloud shadows. Therefore, by performing cloud shadow event detection and motion vector modeling on the time series of the normalized power map, the discrete and static power map snapshots are converted into a continuous dynamic model cloud motion vector field that can describe the motion trend of the entire field of clouds. This vector field provides the core and decisive input for the subsequent accurate prediction of the power impact on downstream regions.

[0026] In an optional implementation, Figure 3The flow chart of step S3 in the intelligent photovoltaic grid-connected regulation method according to the embodiments of the present application is shown. As shown in Figure 3 Step S3, based on the power station zoning map, performs cloud shadow event detection and motion vector modeling on the time sequence of the normalized power map to obtain a cloud motion vector field, including: step S31, performing power drop event detection on the time sequence of the normalized power map based on a sliding window and threshold judgment to obtain a detected event list; step S32, based on the power station zoning map, performing event pairing and local motion vector calculation based on space-time correlation on the detected event list to obtain an original motion vector set; and step S33, performing continuous vector field generation based on spatial interpolation on the original motion vector set to obtain the cloud motion vector field.

[0027] In the optional implementation manner described above, the implementation process of step S3 is as follows: first, the normalized power map generated by step S2 at each time step, for example, every 5 seconds, needs to be continuously received and cached, so as to form a dynamically updated time sequence of the normalized power map in the memory. The time sequence stores the normalized power index history records of all sub-regions for a period of time, for example, the past 1 minute, that is, 12 records. The processing program iteratively checks each sub-region defined in the power station zoning map, for example, Z01 to Z20, in a sliding time window.

[0028] In an optional implementation, the step S31 of performing the power drop event detection based on the sliding window and threshold judgment on the time series of the normalized power map to obtain the detected event list comprises: a step S311 of extracting the normalized power index of a previous time step and the normalized power index of a current time step of a first sub-region from the time series of the normalized power map; a step S312 of judging whether the absolute value between the normalized power index of the current time step and the normalized power index of the previous time step is greater than a preset drop threshold to obtain a first judgment result; a step S313 of judging whether the normalized power index of the current time step exceeds a preset state threshold to obtain a second judgment result; and a step S314 of determining that a drop event occurs in the first sub-region in response to the first judgment result and the second judgment result being true. Taking the sub-region Z15 as an example, at the current time point T, the program first performs the step S311 to extract the normalized power index of Z15 at the current time step T, for example, 0.70, and the normalized power index of Z15 at the previous time step T-5 seconds, for example, 0.95, from the time series. Then, the step S312 is performed. The program calculates the absolute difference between the two indexes, that is, |0.70-0.95|=0.25. The difference is compared with a preset drop threshold. The threshold is set by statistical analysis on the historical sunny and rainy day data of the power station, and is intended to effectively filter out normal fluctuations caused by device noise or weak light changes, for example, which can be set to 0.2. Since the calculated difference 0.25 is greater than the preset threshold 0.2, the first judgment result is true. Next, the step S313 is performed. The program judges whether the normalized power index 0.70 of the current time step has entered a significant power suppression interval. This is done by comparing it with a preset state threshold, which defines the boundary of the non-sunny state, which is determined by statistical analysis on a large amount of historical operation data of the power station, for example, which can be set to 0.85. Since 0.70 is less than 0.85, it indicates that the sub-region Z15 is indeed in the cloud shadow state, and therefore the second judgment result is true. Finally, the step S314 is performed, and in response to the first and second judgment results being true, the program formally determines that a power drop event occurs in the sub-region Z15 at the time point T. A record containing event details is created, for example: {event ID: E20230520T1400S_Z15, occurrence time: T, sub-region ID: Z15, geographic coordinates: (longitude, latitude)}, wherein the geographic coordinates are obtained from the power station partition map. The complete judgment process is performed in parallel for all sub-regions. Finally, a detected event list is output, which collects all successfully identified power drop event records at T.

[0029] Next, step S32 is performed. A continuously updated history buffer is obtained, which stores all detected events in the past time period. The size of the buffer is a preset parameter, for example, it can be set to store all event records in the past 60 seconds to ensure that there is enough data for matching while avoiding interference from outdated information. Specifically, the processing program iterates through each new event in the detected event list at the current time point T. Take a new event in the list as an example, which occurs in sub-region Z15 at time T. The goal of the processing program is to find a most reasonable predecessor event for this new event in the history event buffer, i.e. the event triggered by the same cloud shadow at an earlier time and at an upstream location. This search process follows strict spatio-temporal correlation principles. First, the program sets a physically reasonable maximum cloud movement speed, which can be determined based on historical meteorological statistics of the region where the site is located, for example, it is preset to 30 meters / second. This maximum speed value defines a dynamic search range. For the event occurring in Z15 at T, the program iterates through each historical event in the history event buffer. For example, one of the historical events occurs in sub-region Z08 at time Th. The program makes the following judgments: First, calculate the spatial distance between the two event occurrence locations. The program queries the center geographic coordinates of Z15 and Z08 from the power station partition map respectively, and then calculates the Euclidean distance between the two points, denoted as D. Second, calculate the time interval between the two event occurrences, denoted as Δt = T - Th. Third, based on the spatial distance D and the time interval Δt, calculate an implied movement speed V = D / Δt. Fourth, compare this implied speed with the preset maximum cloud speed of 30 meters / second. Only when V is less than or equal to 30 meters / second, the event of Z08 is considered as a physically reasonable and possible predecessor, and is added to the candidate list. This constraint excludes irrelevant events that are too far apart or too short in time interval, resulting in unrealistic speeds. After iterating through all historical events, there may be zero, one or more candidate predecessor events. If there is no candidate event, the current event of Z15 is considered as an isolated event and cannot be used to calculate the vector. If there are multiple candidate events, a selection strategy is needed to determine the best pairing. A commonly used and effective strategy is to select the candidate event that is closest in space. That is, among all candidate events that satisfy the speed constraint, the one with the smallest spatial distance D from Z15 is selected as the final pairing. For example, after searching and screening, the program finds that an event occurring in sub-region Z08 at T-25 seconds is the best pairing of the current event of Z15. At this time, an original motion vector can be calculated. The displacement of the vector is determined by the difference between the center geographic coordinates of the two sub-regions, i.e. from the coordinates of Z08 to the coordinates of Z15. The time difference is 25 seconds. The vector is calculated by dividing each component of the displacement vector, such as eastward displacement and northward displacement, by the time difference of 25 seconds.The result is a two-dimensional vector with a well-defined direction and magnitude, which represents the average motion of the cloud shadow in the past 25 seconds within the local area from Z08 to Z15. This pairing and calculation process is repeated for all detected events at the current time point T. When all detected events have been processed, the final output is a set of original motion vectors, which contains all the local motion vectors successfully calculated at the current time step. Each vector is a discrete observation of the cloud shadow motion.

[0030] Finally, step S33 is performed. In an optional implementation, step S33, the spatial interpolation based continuous vector field generation on the original motion vector set to obtain the cloud motion vector field, comprises: step S331, extracting a first grid point from the power grid grid definition; step S332, calculating the distance between each original motion vector in the original motion vector set and the first grid point to obtain a distance set; step S333, determining a set of original vector weights based on the distance set; step S334, calculating the weighted sum of the original motion vector set based on the set of original vector weights to obtain the motion vector of the first grid point.

[0031] Specifically, the original motion vector set, for example, at the current time point, the set can contain three successfully calculated original motion vectors: V1, whose starting point is located at the center of sub-region Z08, and the vector value is (12 m / s, 5 m / s); V2, whose starting point is located at the center of Z12, and the vector value is (11 m / s, 6 m / s); V3, whose starting point is located at the center of Z03, and the vector value is (13 m / s, 4.5 m / s). The vector values here represent the speed components in the east and north directions, respectively. First, in step S331, the processing program loads a pre-defined power station grid definition and extracts the first grid point therefrom. The power station grid is a virtual two-dimensional coordinate grid that covers the entire geographical range of the power station and is generated once during the digital modeling stage of the power station. The purpose of the grid is to discretize the continuous physical space for calculation. For example, for a square power station with a size of 1 km by 1 km, a 100x100 grid can be defined, which means that each grid cell has a size of 10 m by 10 m. The goal of the processing program is to calculate a motion vector for the center point of each of the 10000 grid cells, i.e., the grid points. The program starts from the first grid point, for example, the grid point G(1, 1) located at the top left corner of the power station. Next, step S332 is executed to calculate the spatial distances between the grid point G(1, 1) and the starting points of each original motion vector in the original motion vector set. The program obtains the geographical coordinates of G(1, 1) and calculates the Euclidean distances between it and the starting points of V1 (the center of Z08), V2 (the center of Z12), and V3 (the center of Z03), respectively. For example, the calculated distances can be: d1=350 m, d2=500 m, d3=200 m. The output is a distance set for G(1, 1), i.e., {350, 500, 200}. Then, step S333 is executed to determine a weight for each original motion vector based on the distance set. Here, the spatial interpolation method, inverse distance weighting, is used. That is, the influence of a known original motion vector on an unknown point is inversely proportional to the distance between them, i.e., the closer the distance, the greater the influence. The weight calculation formula is: wi=1 / (di^p), where di is the distance and p is a pre-set power parameter used to control the speed of weight decay with distance. The value of p is set to 2, which means that the weight is inversely proportional to the square of the distance, making the influence of nearby points significantly amplified. According to this formula, the original weights are calculated: w1=1 / 350^2, w2=1 / 500^2, w3=1 / 200^2. Subsequently, the original weights are normalized to ensure that the sum of all weights is 1. The normalized weights w'i=wi / (w1+ w2+w3). After calculation, an original vector weight set for G(1, 1) is obtained, for example, {0.15, 0.06, 0.79}.Finally, according to step S334, the final motion vector of grid point G(1, 1) is calculated by weighted summation of the original motion vector set based on the normalized weight set. This is a weighted average process of vectors. The eastward velocity component V_x of G(1, 1) is 0.15*12 + 0.06*11 + 0.79*13 ≈ 12.73 m / s. The northward velocity component V_y of G(1, 1) is 0.15*5 + 0.06*6 + 0.79*4.5 ≈ 4.67 m / s. Therefore, the motion vector of grid point G(1, 1) is determined as (12.73 m / s, 4.67 m / s). This complete interpolation calculation process will be applied to each grid point in the power station grid in turn, from G(1, 1) to G(100, 100). When the motion vectors of all grid points are calculated, the final output is a complete and continuous cloud motion vector field, which can be a two-dimensional array or a hash table, storing the mapping relationship between each grid point coordinate and its corresponding two-dimensional motion vector. This vector field comprehensively describes the cloud motion trend over the entire power station at the current time.

[0032] In step S4, based on the cloud motion vector field and the normalized power map of the current time step, downstream area power impact prediction is performed to obtain the overall sub-area impact prediction. It should be understood that the cloud motion vector field is like a real-time wind direction map drawn for the entire power station, accurately describing the current motion law of the cloud layer on a macroscopic scale. However, this vector field itself only describes the current dynamic and is not a prediction of the future state. In order to truly realize the fundamental change from passive lag compensation to active prediction and coordinated regulation, the known current state and motion law need to be used to deduce the future evolution result. Therefore, based on the cloud motion vector field and the normalized power map of the current time step, the present application performs downstream area power impact prediction to combine the static power distribution map with the dynamic motion vector field, thereby converting the abstract dynamic model into a specific and quantitative power prediction sequence, providing an indispensable data basis for subsequent formulation of the optimal and forward-looking scheduling strategy.

[0033] In an optional implementation, step S4, based on the cloud motion vector field and the normalized power map of the current time step, downstream area power impact prediction is performed to obtain the overall sub-area impact prediction, including: step S41, based on the cloud motion vector field, performing inverse space-time trajectory tracing based on a prediction time domain for each sub-area to obtain a source point request list; step S42, based on the normalized power map of the current time step, performing source point power sampling and prediction sequence construction for each source point request in the source point request list to obtain the overall sub-area impact prediction.

[0034] In the optional implementation manner, the implementation process of step S4 is as follows: first, step S41 is performed. In an optional implementation manner, step S41, based on the cloud motion vector field, performs prediction time domain-based inverse space-time trajectory tracing on each sub-region to obtain a source point request list, comprising: step S411, obtaining the center geographic coordinates of a first sub-region; step S412, obtaining the running vector of the first sub-region from the cloud motion vector field; step S413, based on the running vector of the first sub-region, performing inverse tracing calculation on the center geographic coordinates of the first sub-region to obtain a first source point request, the first source point request comprising the ID, source point coordinates and future time point of the first sub-region.

[0035] Specifically, before starting the calculation, a prediction time domain needs to be set in advance, which defines the range and granularity of the prediction. For example, according to the time scale requirement of the power grid for power smoothing control, the prediction time domain can be set to 300 seconds in the future, and the prediction time step is set to 10 seconds. This means that for each sub-region in the power station, the source point position at 30 time points of T+10 seconds, T+20 seconds,..., T+300 seconds in the future needs to be calculated. The entire implementation process is performed in cycles for sub-regions, and all future time points in each sub-region are traced and calculated. Take Z15 in the power station as an example of the first sub-region.

[0036] First, step S411 is executed to obtain the center geographic coordinates of the first sub-region Z15. This information is static and is obtained directly from the power station partition map data structure determined during the power station planning phase. The map is established on a two-dimensional Cartesian coordinate system with a fixed point of the power station as the origin, and the unit is meters. For example, the query result is that the center coordinates of Z15 are (800, 650), that is, located at a position of 800 meters east and 650 meters north in the power station coordinate system. Next, step S412 is executed to obtain the movement vector at the position of Z15 from the input cloud movement vector field. The cloud movement vector field is a grid data structure covering the entire power station, and each grid point is associated with a movement vector. Since the center coordinates (800, 650) of Z15 are likely to not exactly fall on a certain grid point, it is necessary to obtain the vector at the precise position through spatial interpolation. One implementation is to use bilinear interpolation: first locate the nearest four grid points surrounding the point (800, 650) in the vector field grid, and then obtain the movement vectors of the four grid points. Finally, according to the relative position of the point (800, 650) in the rectangular unit formed by the four grid points, the four vectors are weighted and averaged. In this way, a precise movement vector specific to the center position of Z15 can be obtained. For example, after interpolation calculation, the movement vector of Z15 is (12.2 m / s, 5.1 m / s), where the first component represents the eastward speed and the second component represents the northward speed. Then, step S413 is executed to generate a series of first source point requests based on the obtained movement vector and the center geographic coordinates (800, 650) of Z15. The calculation process traverses each future time point in the prediction time domain. For the first time point in the prediction time domain, that is, 10 seconds in the future (T+10s), the source point coordinates are calculated. The core formula of the reverse tracing is: source point coordinates = target point coordinates - movement vector x time difference. The specific calculation is: eastward source point coordinates = 800 - (12.2 m / s x 10 s) = 678 m. Northward source point coordinates = 650 - (5.1 m / s x 10 s) = 599 m. The physical meaning of this calculation is that the cloud that will move to the center of Z15 in 10 seconds is currently located at (678, 599) in the power station coordinate system at the current time T. After the calculation is completed, the program generates a structured source point request record, that is, the first source point request, whose content is: {target sub-region ID: "Z15", source point coordinates: (678, 599), future time point: T+10s}. This reverse tracing calculation is repeated for all subsequent time points in the prediction time domain. For example, for 20 seconds in the future (T+20s), the time difference is 20 seconds, and the source point coordinates are calculated as: eastward source point coordinates = 800 - (12.2 m / s x 20 s) = 556 m. Northward source point coordinates = 650 - (5.1 m / s x 20 s) = 548 m.and a corresponding source point request is generated: {target sub-region ID: "Z15", source point coordinate: (556, 548), future time point: T+20s}. This loop will continue until the end of the prediction time horizon, i.e. T+300s, resulting in a total of 30 independent source point requests for sub-region Z15, each of which precisely indicates the geographical location of the cloud cluster affecting the power at a certain specific future time point at the current time point. The above complete tracing process (steps S411 to S413) will also be applied to all other sub-regions in the power plant, i.e. Z01 to Z14 and Z16 to Z20, in turn. When the tracing calculation for all sub-regions at all future time points is completed, the final source point request list is obtained, which collects all source point request records of all sub-regions in the entire prediction time horizon, with a total number of entries (number of sub-regions x number of prediction time horizon steps).

[0037] Step S42 is then executed. In an optional implementation, step S42, based on the normalized power map of the current time step, performs source point power sampling and prediction sequence construction on each source point request in the source point request list to obtain the overall sub-region impact prediction, including: step S421, extracting a first source point request from the source point request list; step S422, extracting a source point coordinate from the first source point request; step S423, based on the normalized power map of the current time step, performing spatial interpolation sampling on the source point coordinate to obtain the normalized power index prediction value of the source point coordinate at the current time step. The entire implementation process is a loop process that traverses each record in the source point request list. In order to maintain data flow consistency with the previous steps, the prediction for sub-region Z15 will be taken as an example.

[0038] First, step S421 is executed. For example, it extracts the first record in the list, which is generated for predicting the state of sub-region Z15 at T+10 seconds in the future, and its content is: {target sub-region ID: "Z15", source point coordinate: (678, 599), future time point: T+10s}. Next, step S422 is executed, and the key information, i.e., the source point coordinate, is parsed from the extracted source point request. In this example, the extracted coordinate is (678, 599). This coordinate represents that the cloud state above it at the current time T will determine the power state of Z15 after 10 seconds. Then, step S423 is executed, and based on the source point coordinate (678, 599), spatial interpolation sampling is performed on the normalized power map at the current time point T to obtain the normalized power index prediction value of the exact position at the current time. Since the data of the normalized power map is discrete, i.e., each sub-region such as Z01, Z02, etc. has only one power index value associated with its center point, and the calculated source point coordinate (678, 599) is almost impossible to coincide with the center of any sub-region. Therefore, the power index of this arbitrary point is estimated by spatial interpolation. Here, the bilinear interpolation method is used to achieve this. First, the nearest four sub-region center points that surround the source point coordinate (678, 599) need to be found in the power station zoning map, and these four points form a rectangular or approximately rectangular unit. For example, by querying the zoning map, it is found that the source point (678, 599) is surrounded by the centers of Z09, Z10, Z13, and Z14, whose center coordinates are: Z09(600, 500), Z10(700, 500), Z13(600, 600), and Z14(700, 600). Subsequently, the normalized power indices of these four sub-regions are queried from the normalized power map at the current time point T. For example, the query results are: normalized power index Z09=0.50, normalized power index Z10=0.45, normalized power index Z13=0.80, and normalized power index Z14=0.75. Next, linear interpolation is performed in two steps. First, linear interpolation is performed twice along the east-west direction (X-axis). On the southern line (Y=500), interpolation is performed for point (678, 500). The point is between Z09 and Z10, and its relative position on the X-axis is (678-600) / (700-600)=0.78. This value indicates that point (678, 500) has traveled 78% of the distance on the line segment from Z09 to Z10, so it is closer to Z10 and farther from Z09. Therefore, the interpolation power P_south is = normalized power index Z09×(1-0.78)+normalized power index Z10×0.78=0.461.Using the same logic, the north edge line formed by Z13 and Z14 is interpolated, and the point (678, 600) on the north edge line (Y=600) is interpolated. Since the east coordinate of this point is the same as the virtual point on the south edge line, its X-axis relative position is also 0.78. Therefore, the interpolation power P_north of this point is 0.761. In the second step, the two interpolation points P_south and P_north obtained in the previous step are linearly interpolated along the north-south direction (Y-axis). The relative position of the source point (678, 599) on the Y-axis is (599-500) / (600-500)=0.99. Therefore, the final normalized power index prediction value of the source point is P_south x (1-0.99) + P_north x 0.99 = 0.461 x 0.01 + 0.761 x 0.99 ≈ 0.758. This calculated value 0.758 is determined as the normalized power index prediction value of sub-region Z15 at the future T+10s. The processing program associates this prediction value with the corresponding target information (sub-region ID: "Z15", future time point: T+10s) as the first data point of the prediction sequence of Z15. Then, the program continues to process the next record in the source point request list, which is the prediction request for Z15 at T+20s, and the source point coordinates are (556, 548). The program repeats the above interpolation sampling process to obtain the prediction value of Z15 at T+20s. This process continues until all 30 source point requests for Z15 (from T+10s to T+300s) are processed, thus constructing a complete future power prediction sequence for Z15, which contains 30 data points. When the prediction sequence of a sub-region is constructed, the program will then repeat the entire process for the next sub-region (e.g. Z16). Finally, when all requests in the source point request list are processed, the final output is the impact prediction of all sub-regions. This is a comprehensive data structure, which can be a dictionary with sub-region ID as the key and its corresponding future power prediction time sequence as the value.

[0039] In step S5, the spatiotemporal transfer strategy optimization under the global power target is performed on the impact prediction of all sub-regions to obtain an optimal scheduling plan. Accordingly, the previous steps have successfully generated a detailed and high-precision power prediction sequence for all sub-regions in the power station within the next few minutes. However, this prediction reveals a future that will occur without intervention, which is often full of severe power fluctuations caused by cloud movement, and direct grid connection will cause serious impact on the power grid. It is far from enough to just know that the impact is coming, the core goal is to actively eliminate these impacts. Therefore, the spatiotemporal transfer strategy optimization under the global power target is needed to perform on the impact prediction of all sub-regions to convert this passive prediction into an active, executable, and optimal control blueprint. It intelligently transfers the power surplus that will occur in one region to another region that will soon have a power shortage, thereby ironing the total output power of the whole station on a macro level and achieving a power output that is friendly to the power grid and smooth and stable.

[0040] In an optional implementation, step S5, the spatiotemporal transfer strategy optimization under the global power target is performed on the impact prediction of all sub-regions to obtain an optimal scheduling plan, comprising: step S51, performing baseline power trajectory aggregation on each time point in the impact prediction of all sub-regions based on the clear sky power model to obtain an uncontrolled power trajectory; step S52, performing target trajectory smoothing on the uncontrolled power trajectory to obtain a smoothed target power trajectory; step S53, taking the pre-de-rating power and transient power increase of each sub-region at each future time point as a decision variable; step S54, constructing a target function based on the smoothed target power trajectory, the uncontrolled power trajectory, the pre-de-rating power, and the transient power increase, and constructing a constraint condition based on the inverter safety margin and the total power change limit; step S55, calling a solver to solve the target function and the constraint condition to obtain the optimal scheduling plan, the optimal scheduling plan comprising the optimal value of the pre-de-rating power and the transient power increase of each sub-region at each future time point.

[0041] In an alternative implementation, the implementation of step S5 is as follows: Firstly, step S51 is performed. The input of the process is a detailed data structure containing the normalized power index of all sub-areas (Z01 to Z20) at every predicted time point in the future, for example, every 10 seconds from T+10 seconds to T+300 seconds. The program will iterate through every time point in the prediction time domain, and perform power restoration calculation for all sub-areas at that time point, and finally aggregate them. Take T+60 seconds as an example. The program will first extract the normalized power index prediction value of all 20 sub-areas at T+60 seconds. For example, it is found from the input data that the prediction index of sub-area Z15 at T+60 seconds is 0.62. Next, in order to convert these dimensionless indices into actual power in megawatts, the program needs to calculate the theoretical clear sky power of each sub-area at that future time point. This calculation process completely relies on the internally integrated clear sky power model. Take the calculation of Z15 at T+60 seconds as an example. The program will call the clear sky power model with Z15's ID and the future timestamp T+60 seconds as input. Inside the model, first, the attributes of Z15 will be queried from the pre-set power station static parameter database, such as latitude and longitude, rated installed capacity, for example, 5 megawatts, comprehensive performance coefficient, for example, 0.96, and possible slight changes over time, etc. Next, the solar position algorithm module calculates the solar zenith angle and azimuth angle at that time according to the timestamp T+60 seconds and the latitude and longitude of Z15. Then, these two angle values, together with the annual day number representing seasonal atmospheric conditions, are passed to the neural network-based atmospheric radiation transfer module for a fast forward propagation, and finally the neuron outputs an accurate plane total radiation prediction value, for example, the model outputs 930 watts per square meter. Finally, the photovoltaic array performance module uses this radiation value to perform the final power conversion, and the calculation formula is: theoretical clear sky power = (930 / 1000) x 5 megawatts x 0.96 ≈ 4.46 megawatts. When the theoretical clear sky power of Z15 at T+60 seconds is calculated, the program can calculate its uncontrolled power prediction value at that time. The calculation method is: uncontrolled power = normalized power index x theoretical clear sky power = 0.62 x 4.46 megawatts ≈ 2.77 megawatts. Similarly, the program will calculate the uncontrolled power of all 20 sub-areas, and add them up to get the total uncontrolled power of the entire power station at T+60 seconds. This complete calculation and aggregation process will be performed for every time point in the prediction time domain, from T+10 seconds to T+300 seconds. Finally, the output is an uncontrolled power trajectory. It is a time series, and its data structure can be a list of (timestamp, total power) tuples, which accurately depicts the detailed situation of the total output power of the power station over time in the future 300 seconds without any control.

[0042] Step S52 is then executed. Before starting the calculation, a key constraint parameter needs to be determined: the total power ramp rate limit at the plant level. This limit is specified by the grid dispatching rules, for example, set to 10% of the total rated capacity of the plant per minute. If the total rated capacity of the plant is 100 MW, the maximum allowed ramp rate is 10 MW / min. Since the prediction time step in this application is 10 seconds, this limit needs to be converted to the maximum variation allowed in a single time step, i.e., ΔΡ_Ιιιτ = 10 MW / min ÷ (60 s / min ÷ 10 s / step) = 1.67 MW / step. That is, the ideal target power curve cannot vary more than 1.67 MW between any adjacent 10 seconds. The generation of the smoothed target power trajectory is a point-by-point iterative calculation process. The starting point of the trajectory is set to the actual total power at the current time T. For example, at time T, the actual total power obtained from the plant monitor is 86.0 MW. Then, Target(T) = 86.0 MW. Next, the program will start from the first future time point T+10 s and calculate the target power value at each time point in turn. The target power value Target(t) = Target(t-1) + sign(ΔΡ_natural) * min(|ΔΡ_natural|, ΔΡ_Ιιίτ) where Target(t-1) is the actual total power at the previous time, Target(t) is the actual total power at the current time, ΔΡ_natural is the natural variation of the uncontrolled trajectory, i.e., Uncontrolled(t) - Uncontrolled(t-1), Uncontrolled(t) is the value of the uncontrolled power trajectory at each time, and sign is the sign function. Taking a specific uncontrolled power trajectory segment as an example, the obtained uncontrolled power trajectory is: Uncontrolled(T) = 86.0 MW (current actual value), Uncontrolled(T+10 s) = 85.3 MW, Uncontrolled(T+20 s) = 82.5 MW, and Uncontrolled(T+30 s) = 88.1 MW (a dramatic power jump). The calculation process is as follows: Calculate Target(T+10 s): the target value at the previous time is Target(T) = 86.0 MW. Calculate the natural variation of the uncontrolled trajectory ΔΡ_natural = Uncontrolled(T+10 s) - Uncontrolled(T) = 85.3 - 86.0 = -0.7 MW. Target(T+10 s) = 86.0 + sign(-0.7) * min(|-0.7|, 1.67) = 86.0 + (-1) * 0.7 = 85.3 MW. At this step, since the natural variation 0.7 MW is smaller than the limit 1.67 MW, the target trajectory completely follows the natural trajectory.Target (T+20s): The target value at the previous time is Target (T+10s) = 85.3 MW. The natural change of the uncontrolled trajectory is AP_natural = Uncontrolled (T+20s) - Uncontrolled (T+10s) = 82.5 - 85.3 = -2.8 MW. The formula is applied: Target (T+20s) = 85.3 + sign (-2.8) x min (|-2.8|, 1.67) = 83.63 MW. At this step, since the natural change of 2.8 MW is greater than the limit (1.67 MW), the change of the target trajectory is forced to be limited within the maximum allowed range, thereby achieving the smoothing of the sharp drop. Target (T+30s): The target value at the previous time is Target (T+20s) = 83.63 MW. The natural change of the uncontrolled trajectory is AP_natural = Uncontrolled (T+30s) - Uncontrolled (T+20s) = 88.1 - 82.5 = +5.6 MW. The formula is applied: Target (T+30s) = 83.63 + sign (+5.6) x min (|+5.6|, 1.67) = 85.30 MW. At this step, in the face of a sharp power jump, the target trajectory is also smoothed to a steady rise by the ramp rate limit. This iterative calculation process continues until the entire 300-second prediction time domain is covered. Finally, the output is a new time series of smooth target power trajectories. This trajectory starts from the current actual power, and each step of its change strictly follows the pre-set ramp rate limit, showing a smooth and continuous characteristic.

[0043] Step 53 is then performed. The implementation of this process is to systematically define two core decision variables for each spatio-temporal node for the prediction time domain of 300 seconds (30 time points with a step of 10 seconds) and all 20 sub-regions (Z01 to Z20) in the power plant. The two variables are pre-reduction power and transient power increase. Among them, the pre-reduction power such as Curtailment_MW_j(t) refers to the instruction for a specific sub-region to actively reduce the power compared to its natural predicted power generation capacity, with the unit of megawatts; while the transient power increase such as Boost_MW_j(t) refers to the use of safety margin in inverter design to instruct it to output power exceeding the rated value in a short time, also with the unit of megawatts. That is, pre-reduction is the main way to deal with power surplus and suppress peaks, while transient power increase is the key tool to fill power gaps and deal with rapid drops by utilizing device potential. By abstracting these two complementary and power increase and decrease covering actual available behaviors into mathematical variables, a complete optimization model can be constructed, which can coordinate the deployment of these two resources in time and space within the entire station to achieve accurate tracking of the ideal power curve at the lowest cost.

[0044] Step 54 is performed next. First, the objective function is constructed. The objective function is to make the actual total output power curve of the whole station after the pre-curtailed and transient boosted power regulation of all sub-areas as close as possible to the ideal smooth target power trajectory. First, define the adjusted total power of the whole station at any future time point t, Adjusted_Total_Power(t), which is equal to the sum of the uncontrolled power of all sub-areas minus the total curtailed power of all sub-areas plus the total boosted power of all sub-areas. Its mathematical expression is: Adjusted_Total_Power(t)=∑j(uncontrolled_power_j(t)-Curtailment_MW_j(t)+Boost_MW_j(t)), where j traverses all sub-areas, i.e. Z01 to Z20. Based on this, the core objective function of the optimization problem is constructed to minimize the total absolute deviation between the adjusted total power and the ideal smooth target power in the entire prediction time domain, which is mathematically expressed as: Minimize∑|Adjusted_Total_Power(t)-Smoothed_Target_Power_Trajectory(t)|. This objective function intuitively expresses the intention to track the ideal trajectory, but due to the nonlinear nature of the absolute value function, in order to be able to use an efficient linear programming solver, it needs to be linearly converted. In order to facilitate the linear programming solver, the objective function of minimizing the sum of absolute values is transformed. A set of non-negative auxiliary variables error(t) is introduced, each corresponding to a time point t. The objective function is constructed as: Minimize∑error(t) where t traverses all prediction time points. At the same time, in order to make error(t) represent the absolute value deviation, two constraints need to be added: 1. Adjusted_Total_Power(t)-Smoothed_Target_Power_Trajectory(t)≤error(t). 2. Smoothed_Target_Power_Trajectory(t)-Adjusted_Total_Power(t)≤error(t). These two constraints together ensure that the value of error(t) is always greater than or equal to the absolute value of the difference between Adjusted_Total_Power(t) and the target value. Since the objective function is to minimize the sum of error(t), the solver will automatically push the value of error(t) to its lower limit, which is exactly equal to the absolute value. The second step is to construct the constraint conditions. The constraint conditions define the physically and procedurally feasible region for the optimization problem, ensuring that any solution given by the solver is realistic and feasible.1. Physical boundary constraints of decision variables, i.e. the quantification and materialization of inverter safety margins: for each sub-area j and each time point t, its pre- curtailment power Curtailment_MW_j(t) is non-negative and cannot exceed the natural available power of the sub-area at that time point. For example, for sub-area Z15 at T+60s, if its uncontrolled power forecast value is 2.77 MW, then its constraint is: 0≤Curtailment_MW_Z15(T+60s)≤2.77. Its transient boost power Boost_MW_j(t) is also non-negative and cannot exceed the maximum transient boost capacity Max_Boost_Capacity_j of the sub-area inverter. This value is an inherent property of the inverter and needs to be pre-set, typically 10% to 15% of its rated capacity. For example, if the rated capacity of Z15 is 5 MW, and its boost capacity is set to 10%, then Max_Boost_Capacity_Z15 is 0.5 MW. Its constraint is: 0≤Boost_MW_Z15(T+60s)≤0.5. These two types of constraints need to be established for all 1200 decision variables one by one. 2. Total power variation limit (ramp rate) constraints, which is a key requirement imposed by the grid on power plants for grid connection. The total power of the whole plant after regulation must vary within the allowed range between any adjacent time points. This constraint acts on Adjusted_Total_Power(t). According to the single-step maximum allowed variation ΔP_limit set in step S52 (e.g. 1.67 MW), the following constraint can be constructed: |Adjusted_Total_Power(t)-Adjusted_Total_Power(t-1)|≤1.67. In order to linearize, this absolute value constraint is split into two linear inequalities: Adjusted_Total_Power(t)-Adjusted_Total_Power(t-1)≤1.67, Adjusted_Total_Power(t)-Adjusted_Total_Power(t-1)≥-1.67. This pair of constraints needs to be established for every time point in the prediction horizon (starting from T+20s) to ensure that the entire adjusted power curve is smooth. Finally, a complete and structured mathematical optimization model is output. This model contains a linear objective function (minimize total deviation) and a series of linear constraints (including objective function linearization constraints, decision variable boundary constraints, and total power ramp rate constraints), which completely describe the scheduling problem to be solved.

[0045] In particular, if minimizing the instantaneous deviation between the total station power and the ideal target is taken as the sole objective, i.e., a simple norm form is used, control commands with drastic temporal jitter and spatial contradictions may be generated in order to accurately track the target at every moment. For example, the command for a certain zone may change drastically within consecutive time steps, or two adjacent zones may be subjected to completely opposite controls at the same time. This strategy not only causes unnecessary stress on power electronic equipment such as inverters, but also contradicts the inherent spatial continuity of physical phenomena such as cloud shadow movement. Therefore, constructing an objective function for electric field coordinated control optimization based on spatiotemporal smoothness constraints aims to find an optimal solution for the control strategy itself that is stable and smooth in temporal evolution and continuously coordinated in spatial distribution, while ensuring basic tracking accuracy. This results in a scheduling plan that is physically more reasonable, more equipment-friendly, and more robust to prediction errors.

[0046] Based on this, in a preferred implementation, the construction of the objective function includes: constructing a basic tracking objective function, namely: ;in, and They are time points The The uncontrolled power and the ideal smooth target power of the partition are given, while the L2 norm represents the sum of squares. This is a penalty control command. It should be understood that, in order to establish the core driving force of the entire optimization problem—that is, to ensure that the total output power of the photovoltaic power station after regulation can track the pre-planned smooth target power trajectory as accurately as possible—a basic tracking objective function needs to be constructed. This is the fundamental purpose of meeting grid dispatch requirements and achieving smooth power control. Without this basic objective, any smoothness constraint will be meaningless. Specifically, first, the core control variable is defined, that is, for each partition j at time t, its control command is... This represents the adjustment amount of active power. Subsequently, the original tracking target is rewritten in L2 norm form to make it differentiable, thus facilitating processing by an efficient optimization solver. This function calculates the sum of squared errors between the sum of all partitioned adjusted power at all prediction time points and the ideal target. Minimizing this function directly results in the solver prioritizing a set of control commands. This makes the overall output curve of the entire station approximate the ideal smooth target curve in shape as closely as possible, thereby completing the basic task of suppressing power fluctuations.

[0047] Two regularization terms are constructed, namely: a time smoothing term and a space smoothing term. ;in, It is a time-smoothing regularization term. It is a moment a column vector of control commands of all partitions, is a spatial smoothing term, is an edge weight. Accordingly, to compensate for the deficiency of the basic tracking objective function, the problem of unstable control commands caused by it is solved. The individual tracking objective does not care how the commands change over time or how they are distributed in space, so additional constraints need to be introduced to regulate the control behavior. The construction of the temporal smoothing term aims to suppress the sharp changes of individual partition control commands at consecutive time steps. Its implementation is through a penalty function that calculates the sum of the squares of the differences between the control commands of each partition j at each consecutive time step t and t-1. By minimizing this term, the rate of change of the control commands can be effectively suppressed, making the control curve of each partition smoother, avoiding the command jitter caused by pursuing instantaneous accuracy, and thus reducing the frequent adjustment of devices such as inverters, enhancing the stability of the control strategy. The construction of the spatial smoothing term aims to suppress the large differences in control commands between adjacent partitions at the same time, making the control strategy consistent with the spatial continuity of physical phenomena such as cloud shadows. This term needs to consider the topological adjacency relationship between partitions, for example, using the graph Laplacian in graph theory. Specifically, first construct the power station adjacency graph, that is, consider all partitions of the power station as the vertices of the graph, if two partitions i and j are geographically adjacent, connect the edge (i,j) ∈ E between them, and construct the weighted adjacency matrix W, if i and j are adjacent, the edge weight is 1, otherwise it is zero. The graph Laplacian L is defined as L = D - W, where D is the degree matrix, that is, the diagonal matrix whose diagonal elements are the sum of the weights of all adjacent edges of the corresponding partition. For any vector x defined on the vertices of the graph (i.e., the control command vector of all partitions at time t ), satisfies: This formula calculates the square of the difference in control commands between each pair of adjacent partitions (i,j) at time t, and weights them with their adjacency weights . Therefore, the complete spatial smoothing term is the sum of this term for all time steps, that is, By minimizing this term, large control differences between adjacent partitions can be suppressed, thereby forcing a smooth control field in the entire power station space, avoiding physically unreasonable phenomena such as deep de-rating in zone A while zone B is trying to increase production, making the coordinated control strategy more coordinated and robust.

[0048] The final objective function is obtained by weighting and adjusting the temporal smoothing term, the spatial smoothing term, and the basic tracking objective function, that is: ; where, and These are hyperparameters used to control the intensity of temporal and spatial smoothing, respectively. In other words, the three objectives—basic tracking accuracy, temporal smoothness, and spatial smoothness—are mutually restrictive to some extent, requiring a mechanism to balance their relationship to adapt to different operating conditions and control preferences. By weighting and adjusting these three objective functions, the multi-objective optimization problem can be transformed into a single-objective, solvable convex optimization problem. The final objective function form is as follows: .in, and These are non-negative hyperparameters used to control the intensity of temporal and spatial smoothing, respectively. These two weighting coefficients need to be preset according to actual needs or determined through methods such as cross-validation. For example, if the power grid has extremely high requirements for the smoothness of power fluctuations, they can be appropriately increased. and When the values ​​are set to 0.1 and 0.2, the optimizer tends to sacrifice some tracking accuracy for smoother control commands. Conversely, if the primary task is to maximize tracking accuracy, the values ​​can be reduced, such as to 0.01 and 0.02. By adjusting the weights, scheduling commands that meet different optimization preferences can be flexibly generated. The resulting command curves will be smoothly changing rather than noisy broken lines, making them easier for physical devices to execute and less sensitive to small errors in the prediction model. Thus, a comprehensive optimal solution is found under multiple constraints.

[0049] Finally, step S55 is executed. The implementation process first involves loading the model into a selected optimization solver. These solvers are highly specialized software libraries or standalone programs, such as the commercial solvers Gurobi, CPLEX, or the open-source solver GLPK. They internally incorporate efficient algorithms, such as the Simplex method or the Interior Point method, specifically designed for solving large-scale linear programming problems. The program passes the objective function (Minimize∑error(t)), all 1200 decision variables (Curtailment_MW_j(t) and Boost_MW_j(t)), and all relevant linear constraints (including boundary constraints for decision variables, total power ramp rate constraints, and auxiliary constraints for linearizing the objective function) to the solver by calling its Application Programming Interface (API). Once the model is loaded, the solver begins executing its core solving algorithm. This process can be understood as a search within a high-dimensional feasible region defined by all constraints. Each point within this feasible region represents a dispatching scheme that satisfies all physical and procedural limits. The task of the solver is to systematically find the optimal point within this vast feasible region that minimizes the objective function ∑error(t). To illustrate the work of the solver, consider a specific example: continuing with the scenario from the previous steps, at T+60 seconds in the future, the uncontrolled power trajectory shows that the total power of the plant will reach a high peak, while the smoothed target power trajectory requires a lower, steady output. There is a large positive deviation between the two. To eliminate this deviation, the solver needs to decide how to allocate the tasks of curtailment and boost. It considers the situation for all 20 sub-areas simultaneously: for sub-area Z15, the uncontrolled power prediction is 2.77 MW, and the maximum boost capacity of the inverter is 0.5 MW. The solver needs to decide the values of Curtailment_MW_Z15(T+60s) and Boost_MW_Z15(T+60s). For sub-area Z03, it may be in the cloud edge enhancement zone, with a high uncontrolled power prediction, for example, 4.73 MW. For other sub-areas, it may be in deep cloud shadow, with low power. When making its decision, the solver does not simply let the areas with high power curtail. It makes a global trade-off. For example, it may find that having Z03 take most of the curtailment task, for example, Curtailment_MW_Z03(T+60s)=2.0 MW, while having Z15 take a small amount of curtailment, for example, Curtailment_MW_Z15(T+60s)=0.5 MW, and having sub-areas in deep cloud shadow not operate at all, this combination can satisfy T+60s with the smallest error(T+60s) value, and this allocation is most favorable for satisfying the ramp rate constraints from T+50s to T+60s and from T+60s to T+70s.The solver will make this trade-off and calculation globally optimal for all 30 time points simultaneously. After a series of complex iterative calculations, if the problem has a feasible solution, the solver will return an optimal state. This state contains the specific values of all 1200 decision variables. The final output, i.e., the optimal dispatch plan, is a structured and detailed instruction set that precisely lists the specific values of the pre-de-rated power and transient over-generation power that each sub-area needs to perform at every 10-second time step in the next 300 seconds. Its data structure can be a list, where each element is a dispatch instruction, such as: {time point: T+60s, sub-area ID: Z15, pre-de-rated power: 0.5 megawatts, transient over-generation power: 0.0 megawatts}, {time point: T+60s, sub-area ID: Z03, pre-de-rated power: 2.0 megawatts, transient over-generation power: 0.0 megawatts}, {time point: T+70s, sub-area ID: Z18, pre-de-rated power: 0.0 megawatts, transient over-generation power: 0.2 megawatts}, and so on, covering all sub-areas and all time points.

[0050] In step S6, the optimal dispatch plan is decomposed into dispatch instructions and closed-loop corrections to obtain inverter power settings. That is, the optimal dispatch plan clearly defines the power adjustment tasks that each sub-area should undertake in the next few minutes. However, this plan is based on the sub-area level and cannot be directly executed by the underlying individual inverters; at the same time, there will always be deviations between prediction and reality, and there may be delays or inaccurate responses during execution. Therefore, in order to convert this macro blueprint into executable micro instructions and establish a continuous correction mechanism to ensure that the actual output of the power station can accurately and stably track the preset dispatch target, and finally realize closed-loop control from planning to reality, it is necessary to decompose the dispatch instructions and issue them, and to make closed-loop corrections.

[0051] In one feasible embodiment of the present application, the implementation process of step S6 is as follows: at the beginning of each control period, the processing program first extracts the instructions to be executed at the current time from the optimal dispatch plan. Taking the time T+60s as an example, the program finds from the plan that the instruction for sub-area Z15 is {pre-de-rated power: 0.5 megawatts, transient over-generation power: 0.0 megawatts}.

[0052] Next is the decomposition of the instruction. The 0.5 MW derating task of sub-area Z15 needs to be reasonably assigned to all inverters inside it. A fair and efficient assignment strategy is proportional allocation. The handler will collect the current maximum power point tracking (MPPT) power of all inverters inside Z15 in real time, i.e. the maximum power they can naturally generate under the current light. Let’s say the sum of MPPT power of all inverters inside Z15 is 2.8 MW. For an inverter i inside Z15, if its current MPPT power is 480 kW, then the amount of derating task assigned to it is: (0.5 MW / 2.8 MW) x 480 kW ≈ 85.7 kW. Therefore, the final power setpoint of this inverter at T+60 seconds is calculated as: 480 kW - 85.7 kW = 394.3 kW. This calculation process is performed for every inverter inside Z15, and the calculated power setpoints are then issued to the corresponding inverters through the communication network.

[0053] Finally, the closed-loop correction, which is the key to ensure the accuracy of the planning implementation. After the instruction is issued, the processing program will immediately obtain the actual total output power P_actual of the entire station through the high-speed data acquisition channel. At the same time, it will query the ideal and regulated total power target P_intended from the optimal scheduling plan, that is, the current value of Adjusted_Total_Power(t). There will inevitably be a deviation Error=P_intended-P_actual between the two. This deviation may be caused by small errors in light prediction, inherent delays in the communication network, or non-ideal response of the inverter. In order to dynamically eliminate this deviation, a proportional-integral (PI) controller is used to calculate a global correction power ΔP_correction. The calculation of the correction power combines two mechanisms: first, a proportional term (Kp*Error) proportional to the current deviation Error is calculated, which provides an immediate and fast response. The larger the deviation, the greater the correction. Second, an integral term (Ki*∫Errordt) that accumulates historical deviations is calculated, which aims to eliminate small steady-state errors that cannot be completely corrected by the proportional term. The sum of these two components is the total correction power ΔP_correction. In the next control period (for example, T+62 seconds), this calculated correction power ΔP_correction will be an additional and dynamic adjustment item, proportionally and fairly distributed to all inverters in the field that are performing regulation tasks, and superimposed on their original power set value calculated based on the optimal scheduling plan. This continuous measurement-comparison-calculation-correction high-frequency cycle ensures that the actual total output of the power station can closely track the smooth trajectory planned by the optimal scheduling plan at all times, thereby achieving precise and robust closed-loop control effect.

[0054] In summary, the intelligent photovoltaic grid-connected regulation method based on the embodiments of the present application is illustrated, which converts the photovoltaic power station into a large distributed cloud shadow sensor network, abandoning expensive external observation equipment and instead deeply mining the power data of the inverter inside the station. Specifically, by aggregating and normalizing real-time power data streams, a power map that can intuitively reflect the position and intensity of cloud shadows is constructed. Based on time series analysis of the map, the moving track of the power valley (i.e. cloud shadow) is detected and tracked, thereby establishing a cloud motion vector field model and achieving accurate prediction of future power impact. This active prediction mechanism solves the control lag problem of traditional solutions. Finally, through global spatio-temporal optimization scheduling, the predicted centralized power impact is distributed in advance and smoothly to the inverters in different areas of the station, effectively resolving the spatial inconsistency problem and achieving a transition from passive lag compensation to active prediction and collaborative regulation, greatly improving the grid-friendliness.

[0055] Figure 4 This is a block diagram of an intelligent photovoltaic grid-connected control system according to an embodiment of this application. Figure 4 As shown, the intelligent photovoltaic grid-connected control system 100 according to an embodiment of this application includes: a power data stream aggregation module 110, used to aggregate the original power data stream of the inverter based on the power plant partition map to obtain the total power of each sub-region; a power data normalization module 120, used to normalize the total power of each sub-region to obtain a normalized power map composed of the normalized power index of each sub-region; a cloud motion vector field generation module 130, used to perform cloud shadow event detection and motion vector modeling on the time series of the normalized power map based on the power plant partition map to obtain a cloud motion vector field; an impact prediction module 140, used to perform downstream region power impact prediction based on the cloud motion vector field and the normalized power map of the current time step to obtain the impact prediction of all sub-regions; an optimal scheduling module 150, used to optimize the spatiotemporal transfer strategy under the global power target for the impact prediction of all sub-regions to obtain the optimal scheduling plan; and an inverter power setting module 160, used to decompose and issue scheduling instructions and perform closed-loop correction on the optimal scheduling plan to obtain the inverter power setting value.

[0056] Here, those skilled in the art will understand that the specific operations of each step in the above-described intelligent photovoltaic grid-connected control system have been referenced above. Figures 1 to 3 The intelligent photovoltaic grid-connected control method has been described in detail, and therefore, its repeated description will be omitted.

Claims

1. A smart photovoltaic grid-connected regulation method, characterized in that, The method comprises the following steps: aggregating original power data streams of inverters based on a power plant zoning map to obtain total power of each sub-region; normalizing the total power of each sub-region to obtain a normalized power map composed of normalized power indicators of each sub-region; detecting cloud shadow events and modeling motion vectors based on a time sequence of the normalized power map to obtain a cloud motion vector field based on the power plant zoning map; performing downstream regional power impact prediction based on the cloud motion vector field and the normalized power map at a current time step to obtain overall sub-region impact prediction; optimizing a spatiotemporal transfer strategy under a global power target based on the overall sub-region impact prediction to obtain an optimal scheduling plan; decomposing and issuing scheduling instructions and performing closed-loop correction on the optimal scheduling plan to obtain inverter power setpoints; wherein the optimization of the spatiotemporal transfer strategy under the global power target based on the overall sub-region impact prediction to obtain the optimal scheduling plan comprises: aggregating baseline power trajectories at each time point in the overall sub-region impact prediction based on a clear sky power model to obtain uncontrolled power trajectories; smoothing target trajectories based on the uncontrolled power trajectories to obtain smoothed target power trajectories; taking pre-reduction power and transient power increase of each sub-region at each future time point as decision variables; constructing an objective function based on the smoothed target power trajectories, the uncontrolled power trajectories, the pre-reduction power and the transient power increase, and constructing constraint conditions based on inverter safety margins and total power change limits; calling a solver to solve the objective function and the constraint conditions to obtain the optimal scheduling plan, which includes optimal values of the pre-reduction power and the transient power increase of each sub-region at each future time point.

2. The intelligent photovoltaic grid-connected regulation method according to claim 1, characterized in that, The normalization of the total power of each sub-region to obtain the normalized power map composed of normalized power indicators of each sub-region comprises: inputting the ID and time of each sub-region into a clear sky power model to obtain theoretical clear sky power of each sub-region; normalizing the total power of each sub-region based on the theoretical clear sky power of each sub-region to obtain the normalized power map. 3.The intelligent photovoltaic grid-connected regulation method according to claim 1, characterized in that, The cloud shadow event detection and motion vector modeling based on the time sequence of the normalized power map to obtain the cloud motion vector field based on the power plant zoning map comprises: detecting power drop events based on a sliding window and a threshold from the time sequence of the normalized power map to obtain a detected event list; performing event pairing and local motion vector calculation based on spatiotemporal correlation based on the power plant zoning map to obtain an original motion vector set from the detected event list; generating a continuous vector field based on spatial interpolation from the original motion vector set to obtain the cloud motion vector field.

4. The intelligent photovoltaic grid-connected regulation method according to claim 3, characterized in that, The power drop event detection based on a sliding window and a threshold from the time sequence of the normalized power map to obtain a detected event list comprises: extracting a normalized power index of a previous time step and a normalized power index of a current time step of a first sub-region from the time sequence of the normalized power map; determining whether an absolute value between the normalized power index of the current time step and the normalized power index of the previous time step is greater than a preset drop threshold to obtain a first determination result; determining whether the normalized power index of the current time step exceeds a preset state threshold to obtain a second determination result; in response to the first determination result and the second determination result being true, determining that the first sub-region has a drop event.

5. The intelligent photovoltaic grid-connected regulation method according to claim 3, characterized in that, performing spatial interpolation-based continuous vector field generation on the original motion vector set to obtain a cloud motion vector field, including: extracting a first grid point from a power grid definition; calculating distances between each original motion vector in the original motion vector set and the first grid point to obtain a distance set; based on the distance set, determining an original vector weight set; based on the original vector weight set, calculating a weighted sum of the original motion vector set to obtain a motion vector of the first grid point.

6. The intelligent photovoltaic grid-connected regulating method according to claim 1, wherein, based on the cloud motion vector field and the normalized power map of the current time step, performing downstream region power impact prediction to obtain overall sub-region impact prediction, including: based on the cloud motion vector field, performing prediction time domain-based reverse space-time trajectory tracing on each sub-region to obtain a source point request list; based on the normalized power map of the current time step, performing source point power sampling and prediction sequence construction on each source point request in the source point request list to obtain the overall sub-region impact prediction. 7.The intelligent photovoltaic grid-connected regulation method according to claim 6, characterized in that, based on the cloud motion vector field, performing prediction time domain-based reverse space-time trajectory tracing on each sub-region to obtain a source point request list, including: obtaining a central geographic coordinate of a first sub-region; obtaining a running vector of the first sub-region from the cloud motion vector field; based on the running vector of the first sub-region, performing reverse tracing calculation on the central geographic coordinate of the first sub-region to obtain a first source point request, the first source point request including an ID of the first sub-region, a source point coordinate, and a future time point. 8.The intelligent photovoltaic grid-connected regulation method according to claim 7, characterized in that, based on the normalized power map of the current time step, performing source point power sampling and prediction sequence construction on each source point request in the source point request list to obtain the overall sub-region impact prediction, including: extracting a first source point request from the source point request list; extracting a source point coordinate from the first source point request; based on the normalized power map of the current time step, performing spatial interpolation sampling on the source point coordinate to obtain a normalized power index prediction value of the source point coordinate at the current time step.

9. An intelligent photovoltaic grid-connected regulation system, characterized in that, including: a power data stream aggregation module configured to aggregate inverter original power data streams based on a power plant partition map to obtain total power of each sub-region; a power data normalization module configured to normalize the total power of each sub-region to obtain a normalized power map composed of normalized power indexes of each sub-region; a cloud motion vector field generation module configured to perform cloud shadow event detection and motion vector modeling on a time sequence of the normalized power map based on the power plant partition map to obtain a cloud motion vector field; an impact prediction module configured to perform downstream region power impact prediction based on the cloud motion vector field and the normalized power map of the current time step to obtain overall sub-region impact prediction; An optimal scheduling module is configured to perform time-space transfer strategy optimization under a global power target on the overall sub-area impact prediction to obtain an optimal scheduling plan; An inverter power setting module is configured to perform scheduling instruction decomposition and closed-loop correction on the optimal scheduling plan to obtain inverter power setting values; The optimal scheduling module is configured to perform time-space transfer strategy optimization under a global power target on the overall sub-area impact prediction to obtain an optimal scheduling plan, including: A clear-sky power model is used to aggregate baseline power trajectories at each time point in the overall sub-area impact prediction to obtain an uncontrolled power trajectory; The uncontrolled power trajectory is smoothed to obtain a smoothed target power trajectory; The pre-reduction power and transient power increase of each sub-area at each future time point are used as decision variables; Based on the smoothed target power trajectory, the uncontrolled power trajectory, the pre-reduction power, and the transient power increase, a target function is constructed, and constraint conditions are constructed based on inverter safety margins and total power change limits; A solver is called to solve the target function and the constraint conditions to obtain the optimal scheduling plan, which includes optimal values of the pre-reduction power and the transient power increase of each sub-area at each future time point.

Images