An intelligent power prediction method, device and medium based on distributed photovoltaics

Abstract

The application discloses an intelligent power prediction method and device based on distributed photovoltaics, and a medium, relates to the technical field of power prediction, and comprises the following steps: collecting distribution network operation state data, performing operation compilation on the distribution network operation state data, obtaining a source network storage data set, constructing an electrical topology dynamic graph based on the source network storage data set and real-time switch state, combining an adjacency matrix mask to constrain attention weight, and obtaining a time-space weighted feature sequence; extracting strong coupling node time sequence features from the time-space weighted feature sequence, constructing a joint fitness function according to a distribution network power balance deviation and a prediction error, and using a dynamic inertia weight particle swarm optimization algorithm to online optimize super parameter adaptive working conditions nonlinearity to obtain station-level power supply and distribution prediction; and the application realizes closed-loop cooperation of prediction, scheduling and correction, and finally improves source network storage cooperation control ability and distribution network operation stability.

Description

Technical Field

[0001] This invention relates to the field of power forecasting technology, and in particular to an intelligent power forecasting method, device and medium based on distributed photovoltaics. Background Technology

[0002] As the penetration rate of distributed photovoltaic (PV) power in distribution networks continues to increase, the randomness and intermittency of its output place higher demands on the accuracy of power prediction. Current mainstream solutions typically integrate historical operating data, micro-meteorological factors, and the electrical connections of the distribution network infrastructure to construct a spatiotemporal prediction architecture that coordinates information from multiple sources (source, grid, and energy storage). This type of method relies on graph attention modules to extract spatial topological dependencies between nodes and combines recurrent neural networks to capture the temporal evolution patterns of power. These models achieve nonlinear feature mapping through end-to-end training, effectively improving prediction accuracy under normal meteorological conditions and stable topologies, and have become an important technical foundation supporting economic dispatching of distribution networks and coordinated optimization of energy storage.

[0003] Existing graph-based spatiotemporal prediction models still suffer from insufficient online adaptability to dynamic topology reconfiguration and nonlinear fluctuations in complex distribution network environments with high proportions of distributed photovoltaic (PV) access. Traditional algorithms often rely on static adjacency matrices or pre-defined fixed graph structures, making it difficult to map the dynamic evolution of the electrical topology caused by distribution network switching actions in real time. Key hyperparameters are typically trained and fixed offline based on historical datasets; when faced with multi-level power interactions between the source, grid, and storage layers and strong nonlinear disturbances, the fixed feature weight allocation mechanism easily leads to the accumulation of prediction errors and response lag. Current technologies lack online adaptive optimization and dynamic correction capabilities for real-time operating condition drift. When distribution network safety thresholds are exceeded or inter-layer power deviations fluctuate, it is difficult to balance the robustness of global prediction with local refinement accuracy, limiting their engineering application depth in complex scheduling scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent power prediction method based on distributed photovoltaics to solve the problems of difficulty in effectively modeling the source-grid-storage coupling relationship under real-time topology changes in the distribution network and insufficient online adaptive capability of prediction parameters under complex nonlinear operating conditions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, this invention provides an intelligent power prediction method based on distributed photovoltaic (PV) systems, comprising: collecting distribution network operation status data; performing operational compilation on the distribution network operation status data to obtain a source-grid-storage dataset; constructing an electrical topology dynamic graph based on the source-grid-storage dataset and real-time switch status; constraining attention weights using an adjacency matrix mask to obtain a spatiotemporally weighted feature sequence; extracting time-series features of strongly coupled nodes from the spatiotemporally weighted feature sequence; constructing a joint fitness function based on the distribution network power balance deviation and prediction error; and using a dynamic inertial weighted particle swarm optimization algorithm to online optimize hyperparameters to adapt to the nonlinearity of the operating condition, thereby obtaining the site-level power supply and distribution prediction; identifying the sequential transmission relationship of the same-direction ramping between sites based on the site-level power supply and distribution prediction and the measured PV output in the source-grid-storage dataset, forming... The cloud-based power propagation chain is established. Predictive residuals for stations that have experienced ramp-up are extracted along this chain. Predictions for stations that have not yet experienced ramp-up are then corrected according to the propagation delay and amplitude attenuation relationship, resulting in a cloud-based power propagation correction prediction set. Inter-layer power deviations are calculated using a three-level convergence approach (source-grid-storage) to obtain deviation over-limit and deviation convergence states. A distribution network safety threshold is used to identify fluctuating operating conditions and trigger a dynamic inertial weighting strategy. Global and local refinement corrections are performed on the deviation over-limit and deviation convergence states, respectively. Gaussian mutation and elite retention mechanisms are used to iteratively fine-tune the collaborative parameters, yielding source-grid-storage collaborative prediction values. Network-side publishing and energy storage collaborative invocation are performed based on these collaborative prediction values. Error backfilling and deviation correction are performed after the corresponding source-grid-storage measured operating values ​​arrive, resulting in an updated collaborative dataset.

[0007] As a preferred embodiment of the intelligent power prediction method based on distributed photovoltaic power according to the present invention, the specific steps of collecting distribution network operation status data and compiling the distribution network operation status data to obtain a source-grid-storage dataset are as follows: Collect power distribution network operation data, distributed photovoltaic power output data, and energy storage operation data to obtain power distribution network operation status data; The distribution network operation status data is processed with unified time stamp and unified dimension to obtain standardized operation data. The standardized operation data is then compiled, verified, anomaly repaired, and collected and stored to obtain the source network storage dataset.

[0008] As a preferred embodiment of the intelligent power prediction method based on distributed photovoltaics described in this invention, the steps of constructing an electrical topology dynamic graph based on the source-grid-storage dataset and real-time switching states, and constraining the attention weights using an adjacency matrix mask to obtain a spatiotemporally weighted feature sequence are as follows: A dynamic electrical topology graph that changes over time is constructed based on the source-network-storage dataset and real-time switch status. Temporal features of nodes are extracted from the dynamic graph of electrical topology. Attention scores of unconnected nodes are masked by adjacency matrix masking. After normalizing the attention weights of adjacent nodes, feature weighting and fusion are performed to obtain a spatiotemporally weighted feature sequence.

[0009] As a preferred embodiment of the intelligent power prediction method based on distributed photovoltaics described in this invention, the steps of extracting strongly coupled node time-series features from the spatiotemporal weighted feature sequence, constructing a joint fitness function based on the power balance deviation of the distribution network and the prediction error, and using a dynamic inertial weighted particle swarm optimization algorithm to optimize hyperparameters online to adapt to the nonlinearity of the operating condition, thereby obtaining the site-level power supply and distribution prediction, are as follows: Extract the temporal features of strongly coupled nodes corresponding to the target site from the spatiotemporally weighted feature sequence, and construct the site-level input features; The station-level input features are encoded with operating conditions and matched with historical operating conditions. Combined with the real-time topology status, time-series prediction parameters adapted to the current operating conditions are generated to form a station-level prediction configuration. Using site-level prediction configuration, the power supply, power distribution, source-side injection, energy storage support, and load absorption at the site boundary are predicted simultaneously to obtain the predicted values ​​of power supply, power distribution, source-side injection, energy storage support, and load absorption at the site boundary, and to form the prediction error. A joint fitness function is constructed, and the prediction error and power balance deviation are input into the joint fitness function to calculate the fitness and obtain the fitness evaluation. Based on the fitness evaluation, the hyperparameters are optimized and updated online using the dynamic inertia weighted particle swarm optimization algorithm to obtain the site-level power supply and distribution prediction.

[0010] As a preferred embodiment of the intelligent power prediction method based on distributed photovoltaics described in this invention, the specific steps for identifying the sequential transmission relationship of the same-direction ramping between sites based on the predicted power supply and distribution at the site level and the measured photovoltaic output in the source-grid-storage dataset, thus forming a cloud-shadow power propagation chain, are as follows: The predicted values ​​of source-side injection for each photovoltaic site are extracted from the power supply and distribution forecast at the site level. According to the combination of site identification and collection time scale with the actual photovoltaic output, the change in actual output between adjacent collection time scales and the prediction deviation between actual output and source-side injection predicted value are calculated. The changes in actual output with the same direction of change and continuous collection time scale are connected in sequence to form the photovoltaic output ramping section. By comparing photovoltaic power output ramping sections with consistent changes in direction across different photovoltaic sites in multiple complete ramping processes within the source-grid-storage dataset, and moving along the acquisition timescale to the photovoltaic power output ramping section with an earlier start time, the time offset at which the change sequence corresponds to the highest degree is determined as the propagation delay. The predecessor and successor sites are determined according to the order of their start times, and the amplitude attenuation relationship is determined according to the ratio of the cumulative change amplitudes. The predecessor and successor sites are connected according to the order of their start times to form the cloud shadow power output propagation chain.

[0011] As a preferred embodiment of the intelligent power prediction method based on distributed photovoltaics described in this invention, the method involves extracting the prediction residuals of stations that have already experienced climbing along the cloud shadow power propagation chain, relaying and correcting the prediction quantities of stations that have not yet experienced climbing according to the propagation delay and amplitude attenuation relationship, obtaining a set of cloud shadow relay corrected prediction quantities, and calculating the inter-layer power deviation according to the three-level convergence of source-grid-storage to obtain the deviation exceeding the limit state and the deviation convergence state. The specific steps are as follows: According to the starting time, each photovoltaic site is searched along the cloud shadow power output propagation chain. Sites that have formed a photovoltaic power output ramping section are identified as ramping sites, and subsequent sites that have not yet reached the expected ramping time are identified as non-ramping sites. The prediction deviation between the measured power output of the ramping sites and the predicted value of the source-side injection is extracted to obtain the precursor prediction residual. Read the cumulative propagation delay and amplitude attenuation relationship from the site where the climb has occurred to the site where the climb has not occurred, move the precursor prediction residual to the expected climb period and convert it into a prediction correction amount, and add it to the corresponding source side injection amount prediction value; when the closer precursor site climbs, use the corresponding precursor prediction residual to replace the original precursor prediction residual to obtain the cloud shadow relay correction prediction amount set.

[0012] As a preferred embodiment of the intelligent power prediction method based on distributed photovoltaics described in this invention, the method involves using a distribution network safety threshold to identify fluctuating operating conditions and triggering a dynamic inertial weighting strategy. Global correction and local refinement corrections are performed on the deviation exceeding limits and deviation convergence states, respectively. Furthermore, the Gaussian mutation and elite retention mechanism is used to iteratively fine-tune the collaborative parameters to obtain the source-grid-storage collaborative prediction value. The specific steps are as follows: Extract line load rate, voltage deviation, switch action density and energy storage regulation margin from the distribution network safety threshold, identify fluctuating operating conditions and trigger dynamic inertial weighting strategy; Global correction is performed on the collaborative convergence components that are identified as having excessive deviations. The cross-level collaborative parameters are adjusted as a whole by combining the collaborative quantities of the source side, the grid side, and the storage side to obtain global correction data. Local refinement correction is performed on the collaborative convergence components that are identified as being in a state of biased convergence. The active parameters are then fine-tuned locally by combining historical feasible parameters and the current topological continuity to obtain local refinement data. The source-network-storage collaborative prediction value is obtained by iteratively filtering the global correction data and the local refinement data, and by combining Gaussian mutation and elite retention mechanisms to fine-tune the collaborative parameters.

[0013] As a preferred embodiment of the intelligent power prediction method based on distributed photovoltaic power according to the present invention, the steps of publishing the prediction value based on the source-grid-storage collaborative prediction and calling the energy storage collaborative prediction, and performing error backfilling and deviation correction after the corresponding source-grid-storage measured operating value is reached to obtain an updated collaborative dataset are as follows: The source-grid-storage collaborative prediction values ​​are decomposed into collaborative power and converted into scheduling instructions to generate network-side release values ​​and energy storage collaborative call values, which are then distributed to the corresponding network side and energy storage side according to a unified scheduling identifier. The measured operating values ​​of source, grid, and storage are time-scaled, backfilled with errors, and corrected for deviations. The source-grid-storage collaborative prediction values, network-side published values, energy storage collaborative call values, and corrected measured backfill values ​​are collected and stored to obtain the updated collaborative dataset.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the intelligent power prediction method based on distributed photovoltaic as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent power prediction method based on distributed photovoltaic as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By compiling the distribution network operation status data and constructing a dynamic electrical topology diagram in conjunction with real-time switch status, and using adjacency matrix masks to constrain attention weights, the invention achieves an accurate representation of the dynamic coupling relationship between sources, grid, and storage in the distribution network, thereby improving the effectiveness of feature extraction and the reliability of the prediction basis under complex operating conditions. By constructing a joint fitness function that integrates power balance deviation and prediction error, and using a dynamic inertial weighted particle swarm optimization algorithm for online optimization, the invention achieves adaptive adjustment of prediction parameters to nonlinear operating conditions, thereby improving the accuracy and coordination of station-level power supply and distribution prediction. Furthermore, by combining hierarchical aggregation, deviation classification correction, and measured backfilling updates, the invention further realizes closed-loop coordination of prediction, scheduling, and correction, ultimately improving the source-grid-storage coordinated control capability and the operational stability of the distribution network. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a smart power prediction method based on distributed photovoltaics.

[0019] Figure 2 A flowchart for obtaining the source network storage dataset.

[0020] Figure 3 A flowchart for obtaining the predicted power supply and distribution quantities at the site level.

[0021] Figure 4 A flowchart for obtaining the source-grid-storage collaborative prediction value. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a smart power prediction method based on distributed photovoltaic power, including the following steps: S1. Collect distribution network operation status data, organize the distribution network operation status data to obtain source network storage dataset, and construct an electrical topology dynamic graph based on the source network storage dataset and real-time switch status. Combine the adjacency matrix mask to constrain the attention weights to obtain a spatiotemporal weighted feature sequence.

[0026] S1.1 Collect distribution network operation data, distributed photovoltaic power output data and energy storage operation data to obtain distribution network operation status data.

[0027] Furthermore, by using distribution automation terminals, distributed photovoltaic monitoring units, and energy storage management systems, distribution network operation data, distributed photovoltaic output data, and energy storage operation data are collected respectively. The distribution network operation data, distributed photovoltaic output data, and energy storage operation data are then time-stamp aligned, format unified, and measurement point matched. The processed multi-source data are then integrated to obtain distribution network operation status data.

[0028] S1.2. Perform unified time stamping and unified dimension processing on the distribution network operation status data to obtain standardized operation data. Then, perform compilation, verification, anomaly repair, and collection and storage on the standardized operation data to obtain the source network storage dataset.

[0029] Furthermore, the distribution network operation status data is time-aligned according to a unified sampling period, and the power, voltage, current, and state of charge data are uniformly converted to obtain standardized operation data. Then, missing data, isolated jump point data, communication delay duplicate data, and data with abnormal dimensions in the standardized operation data are identified, and anomalies are corrected by interpolation completion, sliding window smoothing, timestamp deduplication, and unit conversion correction. The data is classified and collected according to source side data, network side data, and storage side data to obtain source-network-storage dataset.

[0030] S1.3 Construct a dynamic electrical topology diagram that changes over time based on the source network storage dataset and real-time switch status.

[0031] It should be noted that real-time switch status refers to the actual opening and closing status information of various switching devices in the distribution network at the current moment, which is used to characterize whether the corresponding lines or nodes are in a conductive connection state. The source-network-storage dataset refers to a comprehensive data set formed by organizing source-side data, network-side data, and storage-side data according to unified rules. Furthermore, the source-side nodes, network-side nodes, and storage-side nodes and their connection information are extracted from the source-network-storage dataset. Combined with the real-time switch status, the actual conduction relationship at each moment is determined. When the switch is in the closed state, the corresponding branch edge is activated to indicate the conduction connection between nodes. When the switch is in the open state, the corresponding branch edge is deactivated or shielded to indicate the disconnection between nodes. The node-branch topology structure at each moment is updated in chronological order to form a dynamic electrical topology diagram that changes with time.

[0032] S1.4 Extract the temporal features of nodes from the electrical topology dynamic graph, mask the attention scores of unconnected nodes according to the adjacency matrix mask, and perform feature weighted fusion after normalizing the attention weights of adjacent nodes to obtain the spatiotemporally weighted feature sequence.

[0033] Furthermore, from the dynamic electrical topology graph, the power, voltage, current and connection status change information of each node at continuous time points are extracted in chronological order to form node temporal features. Based on the node connection relationship at each time point, an adjacency matrix mask is generated, and the calculation range of attention weights is constrained so that nodes are only assigned weights with associated nodes that have electrical connection relationships. The node temporal features are then weighted and fused to obtain a spatiotemporally weighted feature sequence.

[0034] It should be noted that a matrix mask is a constraint matrix constructed based on whether there is an electrical connection between nodes. It is used to limit the calculation range of attention weights, so that nodes only perform feature weighting on nodes connected to them.

[0035] S2. Extract the time-series features of strongly coupled nodes from the spatiotemporal weighted feature sequence, construct a joint fitness function based on the power balance deviation of the distribution network and the prediction error, and use the dynamic inertial weighted particle swarm optimization algorithm to optimize hyperparameters online to adapt to the nonlinearity of the operating condition, thereby obtaining the site-level power supply and distribution prediction.

[0036] S2.1 Extract the temporal features of strongly coupled nodes corresponding to the target site from the spatiotemporal weighted feature sequence, and construct the site-level input features.

[0037] Furthermore, based on the spatiotemporal weighted feature sequence, the node position corresponding to the target site is located, and according to the weight between the target site and each associated node in the spatiotemporal weighted feature sequence, strongly coupled nodes with a high degree of association with the target site are selected. Then, the temporal features corresponding to the strongly coupled nodes at continuous time points are extracted and combined with the temporal features corresponding to the target site at continuous time points to obtain the site-level input features.

[0038] It should be noted that the spatiotemporal weighted feature sequence refers to the feature sequence formed by weighting the power, voltage, current and connection status changes of each node at continuous time based on the electrical topology dynamic diagram, reflecting the temporal change characteristics of the nodes and the electrical correlation between the nodes; A strongly coupled node refers to a node in the electrical topology dynamic diagram that has a direct or indirect electrical connection with the target site and whose normalized association weight with the target site reaches the strong coupling judgment threshold. The strong coupling judgment threshold is taken as the upper quartile value of the historical association weight between the target site and each associated node, with an exemplary value range of 0.15 to 0.30, and 0.20 in this embodiment. If there is no node that reaches the strong coupling judgment threshold, the node with the largest normalized association weight is selected.

[0039] S2.2. Perform working condition concept encoding and historical working condition matching on the station-level input features, and combine the real-time topology status to generate time-series prediction parameters that are adapted to the current working conditions, thus forming a station-level prediction configuration.

[0040] It should be noted that real-time topology status refers to the actual node connection relationships and branch conduction status of the distribution network at the current moment; Furthermore, operating condition features such as load level, photovoltaic output level, energy storage state of charge, node voltage, current changes, and real-time topology status are extracted from the site-level input features, and these operating condition features are encoded into a current operating condition concept vector. The current operating condition concept vector is then matched with historical operating condition concept vectors in the historical operating condition database to select historical operating conditions that meet preset similarity conditions, and the historical time series prediction parameters and parameter change patterns corresponding to these historical operating conditions are extracted. By combining the node connection relationships and branch conduction status in the real-time topology, the parameter combinations and adaptation relationships of historical time series prediction parameters are adjusted to obtain time series prediction parameters that are adapted to the current operating conditions. Then, according to the prediction start time, prediction time range and parameter action order corresponding to the target site, the time series prediction parameters are divided into time periods, arranged in order and bound accordingly to form a site-level prediction configuration that includes the parameter content of each prediction time period and the order of parameter use.

[0041] It should be noted that the similarity is determined based on the cosine similarity between the current operating condition concept vector and the historical operating condition concept vector. The operating condition concept vector includes load level, photovoltaic output level, energy storage state of charge, node voltage, current change, and real-time topology status. The exemplary range of the historical operating condition matching threshold is 0.80 to 0.95, and this embodiment uses 0.85. When the similarity reaches the historical operating condition matching threshold, it is determined as a matching operating condition. When there are multiple matching operating conditions, the historical operating condition with the highest similarity is selected. The prediction time range is determined based on the prediction task type and the unified sampling period. The exemplary time range corresponding to ultra-short-term prediction is 15 to 60 minutes, and the exemplary time range corresponding to short-term prediction is 1 to 6 hours. This embodiment uses ultra-short-term prediction with a unified sampling period of 5 minutes and a prediction time range of 60 minutes, which corresponds to 12 consecutive prediction time scales.

[0042] S2.3. Using site-level prediction configuration, the power supply, power distribution, source-side injection, energy storage support, and load absorption at the site boundary are predicted simultaneously to obtain the predicted values ​​of the power supply, power distribution, source-side injection, energy storage support, and load absorption at the site boundary, and to form the prediction error.

[0043] It should be noted that site-level prediction configuration refers to the configuration content organized according to the prediction start time, prediction time range and parameter application order corresponding to the target site. Site-level prediction configuration includes the time series prediction parameters corresponding to each prediction period, the order of parameter use, and the correspondence between the parameters and the prediction time series of the target site. Furthermore, based on the site-level prediction configuration, according to the time-series prediction parameters corresponding to each prediction period, the site boundary power supply, site boundary power distribution, source-side injection, energy storage support and load absorption of the target site in the corresponding period are calculated synchronously, and the predicted values ​​of site boundary power supply, site boundary power distribution, source-side injection, energy storage support and load absorption are obtained respectively. The predicted values ​​of power supply, power distribution, source-side injection, energy storage support, and load absorption at the site boundary are compared with the actual values ​​of power supply, power distribution, source-side injection, energy storage support, and load absorption at the corresponding times. The differences are calculated to form the prediction errors of power supply, power distribution, source-side injection, energy storage support, and load absorption at the site boundary. These errors are then summarized to form the prediction error.

[0044] The expression for calculating the power prediction vector is as follows: ; in, Indicates the target site During the forecast period The multi-output power prediction vector; Indicates the target site During the forecast period Corresponding to the first The output coefficient matrix at each backtracking time step; Indicates the target site During the forecast period The synchronous coupling correction matrix; Represents the identity matrix, used to maintain the baseline mapping relationship of the original prediction results; Indicates the target site At any moment The standardized input feature vector; Indicates the target site During the forecast period The bias vector; This indicates the length of the backtracking time series, which is the total number of historical time series steps used for the current prediction; Indicates the first step in the backtracking sequence. A historical step-size index.

[0045] S2.4 Construct a joint fitness function, input the prediction error and power balance deviation into the joint fitness function to calculate the fitness, obtain the fitness evaluation, and use the dynamic inertial weighted particle swarm optimization algorithm to optimize and update the hyperparameters online based on the fitness evaluation to obtain the site-level power supply and distribution prediction.

[0046] It should be noted that the power balance deviation refers to the degree of deviation between the predicted power supply, source-side injection, and energy storage support at the site boundary and the predicted power distribution and load absorption at the site boundary. The power balance deviation is determined by calculating the prediction error corresponding to each predicted quantity and by the difference between the sum of the predicted power supply, source side injection and energy storage support at the site boundary and the sum of the predicted power distribution and load absorption at the site boundary. After normalizing the prediction error and power balance deviation, the prediction error and power balance deviation are weighted and fused according to preset weights to construct a joint fitness function that characterizes the prediction accuracy and power balance consistency. Furthermore, based on the magnitude of each prediction error in the prediction error, error evaluation is performed on the predicted values ​​of power supply, power distribution, source-side injection, energy storage support, and load absorption at the site boundary. At the same time, deviation evaluation is performed on the consistency of power income and expenditure based on the power balance deviation. Then, the error evaluation results and deviation evaluation results are substituted into the joint fitness function for calculation to obtain the fitness evaluation corresponding to each hyperparameter.

[0047] Based on fitness evaluation, the hyperparameters to be updated are used as particle search objects. The current position and historical best position of each particle are compared according to fitness evaluation, and the particle search position is iteratively updated to obtain a hyperparameter combination with better fitness evaluation. Then, the updated hyperparameter combination is used to determine the prediction results corresponding to the power supply, power distribution, source-side injection, energy storage support and load absorption at the site boundary, so as to obtain the site-level power supply and distribution prediction.

[0048] It should be noted that the preset weights include prediction error weights and power balance deviation weights, the sum of which is 1; valid samples that have completed the backfilling of measured values ​​are selected from the historical operating condition database, and the joint fitness is calculated using different weight combinations, and the weight combination with the smallest corresponding comprehensive prediction error is selected as the preset weight; the exemplary value range of the prediction error weight is 0.6 to 0.8, and the exemplary value range of the power balance deviation weight is 0.2 to 0.4, and in this embodiment, they are 0.7 and 0.3 respectively; The operating parameters of the dynamic inertia weighted particle swarm optimization algorithm are determined based on the optimization results of historical effective samples and the allowable computation time for a single prediction. The exemplary range of the number of particles is 20 to 50, the exemplary range of the inertia weight is 0.4 to 0.9, the exemplary range of the maximum number of iterations is 50 to 200, and random perturbation is performed on the particle positions when the fitness optimum has not been updated for 5 to 10 consecutive iterations. The particle number is set to 30, the inertia weight is dynamically changed between 0.4 and 0.9, the maximum number of iterations is 100, and random perturbation is triggered when there has been no update for 8 consecutive iterations.

[0049] S3. Based on the predicted power supply and distribution at the site level and the measured photovoltaic output in the source-grid-storage dataset, identify the sequential transmission relationship of the same-direction ramping between sites and form a cloud-shadow output propagation chain. Extract the prediction residuals of the sites that have already ramped along the cloud-shadow output propagation chain, and correct the predictions of the sites that have not yet ramped according to the propagation delay and amplitude attenuation relationship to obtain the cloud-shadow relay correction prediction set. Calculate the inter-layer power deviation according to the three-level convergence of source-grid-storage to obtain the deviation over-limit state and deviation convergence state.

[0050] S3.1 Extract the predicted value of the source-side injection of each photovoltaic site from the site-level power supply and distribution prediction. According to the site identification and the corresponding combination of the acquisition time and the photovoltaic measured output, calculate the change in measured output between adjacent acquisition time and the prediction deviation between the measured output and the predicted value of the source-side injection. Connect the measured output changes with the same direction of change and continuous acquisition time in sequence to form the photovoltaic output ramp-up section.

[0051] Furthermore, the site identifier, prediction period, and predicted source-side injection value of each photovoltaic site are read from the site-level power supply and distribution forecast. Using the site identifier as the search key, the measured photovoltaic output with the same acquisition time point within the prediction period is retrieved from the source-grid-storage dataset. The predicted source-side injection value is matched one-to-one with the measured photovoltaic output of the same site and acquisition time point, and arranged according to the acquisition time point to form a site output comparison sequence. Adjacent data are read sequentially, and the measured photovoltaic output of the previous acquisition time point is subtracted from the measured photovoltaic output of the later acquisition time point to obtain the change in measured output. The predicted source-side injection value is subtracted from the measured photovoltaic output of the same acquisition time point to obtain the prediction deviation.

[0052] The measured output change is compared with the natural fluctuation boundary. If it is greater than the natural fluctuation boundary, it is determined to be a positive change; if it is less than the natural fluctuation boundary (negative value), it is determined to be a negative change; if it is within the natural fluctuation boundary, it is determined to be stable. Measured output changes with the same direction of change, continuous acquisition time scale, and not yet determined to be stable are connected sequentially until a change in direction of change, a stable state, or an interruption in acquisition time scale occurs. The connected section is determined as the photovoltaic output ramp-up section, and the start and end times of the section, the direction of change, the cumulative change amplitude, and the prediction deviation corresponding to each acquisition time scale are recorded.

[0053] It should be noted that the natural fluctuation boundary is determined based on the site identification, rated capacity, switch status, data collection timescale, and measured photovoltaic output of the same photovoltaic site. A historical period with unchanged switch status and continuous data within the last 7–30 days is selected. The absolute value of the difference in measured output between adjacent data collection timescales is calculated to form a historical variation set, and the upper quartile of this set is taken as the natural fluctuation boundary. The sampling period is 5–15 minutes, and the natural fluctuation boundary is 2%–6% of the site's rated capacity. For example, if the rated capacity of the photovoltaic site is 10MW, the sampling period is 5 minutes, and the upper quartile of the historical variation set is 0.4MW, then the natural fluctuation boundary is 0.4MW. The measured output variation is within... The range of 0.4 MW to 0.4 MW is defined as stable; the range greater than 0.4 MW is defined as a positive change; and the range less than 0.4 MW is defined as a positive change. The change was identified as negative at 0.4MW.

[0054] S3.2. Compare the photovoltaic power output ramping sections with consistent change direction at different photovoltaic sites in the source-grid-storage data center during multiple complete ramping processes. Move along the collection time marker to the photovoltaic power output ramping section with the earlier start time. Determine the time offset when the change sequence corresponds to the highest degree as the propagation delay. Determine the predecessor site and successor site according to the order of the start time. Determine the amplitude attenuation relationship according to the ratio of the cumulative change amplitude. Connect the predecessor site and successor site according to the order of the start time to form the cloud shadow power output propagation chain.

[0055] Furthermore, based on the photovoltaic site combination, multiple complete ramping processes of photovoltaic output ramping sections are read. From each complete ramping process, two ramping sections with the same direction of change are selected, and the section with the earlier start time is determined as the section to be moved. Taking the difference in start time between the two sections as the center, within a range not exceeding the duration of the shorter section, the section to be moved is moved backward step by step according to a unified sampling period. After each move, the measured output change sequence with the same collection time of the two sections is extracted, the mean of the corresponding sequence is deducted, and the normalized correlation is calculated. The time offset with the largest normalized correlation and reaching the correlation judgment boundary is recorded as the propagation delay sample. At the same time, the ratio of the cumulative change amplitude of the section with the later start time to the cumulative change amplitude of the section with the earlier start time is recorded as the amplitude attenuation sample.

[0056] The normalized corresponding degree expressions for the two photovoltaic output ramp-up sections at the current time offset are as follows: ; in, This indicates the degree of correspondence between the two photovoltaic power output ramp-up sections at the current time offset; This represents the sequence number of the valid sampling points, with values ​​increasing sequentially from 1 to... ; This represents the total number of valid sampling points used in both photovoltaic power output ramp-up sections, taking the smaller value between the number of sampling points in the two sections; This indicates the photovoltaic power output ramp-up section with an earlier start time. One measured change in output force; This indicates the movement of another photovoltaic power generation ramp section. After each sampling step and The corresponding measured change in output force; This represents the number of candidate time steps, measured in units of a uniform sampling period. This indicates the sampling point number after the completion of the second photovoltaic power output ramp-up section.

[0057] The propagation delay samples, amplitude attenuation samples, and start time relationships of the same site combination during multiple complete ramp processes are collected. When the number of times the start time relationship is consistent is greater than the number of times it is opposite, the site with the earlier start time in most cases is identified as the predecessor site, and the other site is identified as the successor site. The median values ​​of the propagation delay samples and amplitude attenuation samples are taken as the propagation delay and amplitude attenuation relationship between the corresponding sites. The predecessor sites and successor sites are connected according to the propagation delay from small to large. When the same site corresponds to multiple connection relationships, the connection relationship with the largest normalized correlation median value is retained to form the cloud shadow output propagation chain.

[0058] It should be noted that the normalized correlation is used to characterize the consistency of the change patterns of two measured output change sequences under corresponding time offsets; the correlation determination boundary is taken as the lower quartile of the maximum normalized correlation of the same site combination in multiple complete climbing processes; the amplitude decay relationship is determined by the median of multiple cumulative change amplitude ratios, which is used to characterize the transmission ratio of output change amplitude between sites during cloud shadow propagation.

[0059] S3.3. Search each photovoltaic site along the cloud shadow power output propagation chain according to the starting time. The sites that have formed a photovoltaic power output ramping section are identified as ramping sites, and the subsequent sites that have not yet reached the expected ramping time are identified as non-ramping sites. Extract the prediction deviation between the measured power output of the ramping sites and the predicted value of the source side injection to obtain the precursor prediction residual.

[0060] Furthermore, starting from the head of the cloud-based power output propagation chain, the measured photovoltaic output of each photovoltaic site is read sequentially, and photovoltaic output ramp-up segments with the same direction of change recorded in the cloud-based power output propagation chain are retrieved. Taking the starting time of the first retrieved photovoltaic output ramp-up segment as the propagation starting point, the propagation delay between the starting point and adjacent sites along the way is accumulated sequentially to obtain the expected ramp-up time of each subsequent site. When a subsequent site has formed a photovoltaic output ramp-up segment with the same direction and the starting time is within the expected ramp-up time range, it is determined as a site where ramp-up has occurred; when a subsequent site has not yet formed a corresponding photovoltaic output ramp-up segment and the current acquisition time is earlier than the end of the expected ramp-up time range, it is determined as a site where ramp-up has not occurred; when the current acquisition time exceeds the end of the expected ramp-up time range and no corresponding photovoltaic output ramp-up segment has been formed, the subsequent retrieval of the corresponding connection relationship is stopped.

[0061] Read the data collection time and measured photovoltaic output within the photovoltaic output ramping section corresponding to each ramping site. Retrieve the source-side injection prediction value corresponding to the same site identifier and data collection time from the site-level power supply and distribution prediction. Subtract the source-side injection prediction value from the measured photovoltaic output of each data collection time. Arrange the difference obtained according to the time interval relative to the start time of the photovoltaic output ramping section. Correlate it with the ramping sites, subsequent ramping sites, and corresponding propagation delays to obtain the precursor prediction residual.

[0062] It should be noted that the expected climbing time range includes the range start point, the expected climbing time, and the range end point. Historical effective propagation delay samples are determined based on the start time difference of corresponding climbing segments during multiple complete climbing processes for the same precursor and successor sites, retaining propagation delay samples with consistent change directions and normalized correlation reaching the association determination boundary. The median value of the historical effective propagation delay samples is taken as the expected propagation delay. The propagation start point is added to the expected propagation delay to obtain the expected climbing time. The range start point and range end point are determined based on the lower quartile and upper quartile values ​​of the historical effective propagation delay samples, respectively. For example, when the historical effective propagation delay samples are 6 min, 8 min, 9 min, 11 min, 12 min, 14 min, and 16 min, the median value is 11 min, the lower quartile value is 8 min, and the upper quartile value is 14 min. When the propagation start point is 10:00, the expected climbing time is 10:11, and the expected climbing time range is 10:08–10:14.

[0063] S3.4 Read the cumulative propagation delay and amplitude attenuation relationship from the site where the climb has occurred to the site where the climb has not occurred, move the precursor prediction residual to the expected climb period and convert it into a prediction correction amount, and add it to the corresponding source side injection amount prediction value; when the closer precursor site climbs, use the corresponding precursor prediction residual to replace the original precursor prediction residual to obtain the cloud shadow relay correction prediction amount set.

[0064] Furthermore, the propagation delay and amplitude attenuation coefficient of each segment between the climbing station and the non-climbing station are read along the cloud shadow output propagation chain. The propagation delay of each segment is added together to obtain the cumulative propagation delay, and the amplitude attenuation coefficient of each segment is multiplied together to obtain the cumulative amplitude attenuation coefficient. The climbing start time of the climbing station is subtracted from the acquisition time of the predecessor prediction residual to obtain the time interval of each predecessor prediction residual relative to the climbing start time. The time interval is added to the expected climbing time of the non-climbing station to determine the target correction time scale corresponding to each predecessor prediction residual. Then, the predecessor prediction residual is multiplied by the cumulative amplitude attenuation coefficient to obtain the prediction correction amount corresponding to each target correction time scale, and then superimposed on the source-side injection amount prediction value with the same target correction time scale.

[0065] The expression for the cloud shadow relay correction prediction value is as follows: ; in, This indicates the predicted value after correction at subsequent sites; Indicates the identifier of the successor site; This represents the original predicted value; This indicates the measured output of the preceding station; Indicates the preceding station identifier; This represents the predicted value for the preceding site; Indicates the time of data collection; Indicates the cumulative transmission delay; This represents the cumulative amplitude transfer coefficient.

[0066] It should be noted that, It is obtained by multiplying the median values ​​of the cumulative output change amplitude ratios of each segment from the preceding station to the successor station in the cloud shadow output propagation chain.

[0067] When a precursor site closer to a site that has not yet experienced a ramp-up forms a photovoltaic power ramp-up section in the cloud-shadow power output propagation chain, the precursor prediction residual of the closer precursor site is extracted, and the cumulative propagation delay and cumulative amplitude attenuation coefficient from the closer precursor site to the site that has not yet experienced a ramp-up are recalculated to generate a new prediction correction. The new prediction correction replaces the original prediction correction corresponding to the correction time scale of the overlapping target, while the correction time scale of the non-overlapping target continues to retain the original prediction correction until a site that has not yet experienced a ramp-up forms a photovoltaic power output ramp-up section. The corrected source-side injection prediction value is combined with the corresponding site boundary power supply prediction value, site boundary power distribution prediction value, and energy storage support prediction value to obtain the cloud-shadow relay correction prediction set.

[0068] S3.5 Map the cloud shadow relay correction prediction set according to the source-network-storage three-level mapping, and calculate the inter-layer power deviation by combining inter-layer transmission loss and topology continuity; when the inter-layer power deviation exceeds the allowable deviation range, it is determined to be the deviation over-limit state, and when it continuously decreases and returns to the allowable deviation range, it is determined to be the deviation convergence state.

[0069] Furthermore, the corrected source-side injection volume prediction value, site boundary power supply volume prediction value, site boundary power distribution volume prediction value, and energy storage support volume prediction value are read from the cloud shadow relay correction prediction volume. Based on the corresponding site identifier and electrical topology dynamic diagram, the source-side node, grid-side node, and energy storage node to which each prediction value belongs are determined, and the corresponding feeder boundary is traced along the conduction branch between nodes to form a three-level mapping relationship of source-grid-energy storage. When there is a broken branch in the mapping path, it is divided into mutually independent continuous topology segments with the broken position as the boundary.

[0070] The predicted values ​​of source-side injection, site boundary power supply, and energy storage support are summarized in each continuous topological segment to obtain the segment input power. The predicted values ​​of site boundary power distribution and inter-layer transmission loss are summarized to obtain the segment output power. The absolute value of the difference between the segment input power and the segment output power is determined as the inter-layer power deviation. When the inter-layer power deviation exceeds the allowable deviation range, it is determined as a deviation over-limit state. When it is within the allowable deviation range and the inter-layer power deviation of each adjacent layer does not increase or decreases at least once since the most recent deviation over-limit time, it is determined as a deviation convergence state.

[0071] The expression for the normalized inter-layer power deviation is as follows: ; in, This represents the normalized interlayer power deviation in topologically continuous segments; Indicates a topologically continuous segment; Indicates the current calculation time; This represents the segment input power obtained by summing the source-side injection, the site boundary power supply, and the energy storage support. This represents the segment output power obtained by summing the power distribution at the site boundary; Indicates the inter-layer transmission loss within the segment; This represents the power regularization quantity, which can be taken as the smallest positive power increment that the power acquisition device can recognize.

[0072] It should be noted that, Determined by the minimum resolution of the power acquisition device; The source-network-storage mapping path is obtained by dividing it into disconnected branches; It is determined by the unified acquisition timescale corresponding to each predicted quantity involved in the power deviation calculation.

[0073] S4. Use the distribution network safety threshold to identify fluctuating operating conditions and trigger the dynamic inertia weight strategy. Perform global correction and local refinement correction for deviation over-limit state and deviation convergence state respectively. Use Gaussian mutation and elite retention mechanism to iteratively fine-tune the collaborative parameters to obtain the source-grid-storage collaborative prediction value.

[0074] S4.1 Extract line load rate, voltage deviation, switch action density and energy storage regulation margin from the distribution network safety threshold, identify fluctuating operating conditions and trigger dynamic inertia weight strategy.

[0075] Furthermore, the line load rate, voltage deviation, switch action density, and energy storage regulation margin corresponding to the current moment are extracted from the source-grid-storage dataset and real-time topology status. The line load rate, voltage deviation, switch action density, and energy storage regulation margin are compared with the corresponding discrimination ranges in the distribution network safety threshold to obtain the over-limit situation and fluctuation situation of each operating status. Based on the over-limit situation and fluctuation situation, it is determined whether the current moment belongs to the fluctuating condition; when the fluctuation degree of at least one of the line load rate, voltage deviation, switching action density or energy storage regulation margin reaches the trigger condition, the dynamic inertia weight strategy is executed to dynamically adjust the inertia weight of the particle swarm optimization algorithm to form a weight state that is adapted to the current fluctuating condition.

[0076] The expression for calculating the voltage deviation is as follows: ; in, Indicates the current time The voltage deviation is used to characterize the degree of deviation of the actual node voltage from the reference voltage; Indicates voltage; Represents a node The corresponding voltage; Indicates the corresponding distribution network node identifier; Represents a node The reference voltage value or the rated reference voltage value, Indicates a reference value; This represents the voltage regularization term, used to avoid anomalies in the denominator when the reference voltage is involved in normalization, and to improve calculation stability.

[0077] It should be noted that the dynamic inertia weight strategy refers to a processing strategy that dynamically adjusts the inertia weight in the dynamic inertia weight particle swarm optimization algorithm according to changes in fluctuating operating conditions, so that the dynamic inertia weight particle swarm optimization algorithm can match different search intensities under different fluctuating operating conditions.

[0078] S4.2 Perform global correction on the collaborative convergence components that are identified as having excessive deviations. Combine the collaborative quantities of the source side, network side, and storage side to adjust the cross-level collaborative parameters as a whole to obtain global correction data.

[0079] Furthermore, when performing global correction on collaborative aggregation components that are identified as having excessive deviations, the source-side collaborative injection prediction, grid-side collaborative power supply prediction, grid-side collaborative power distribution prediction, and storage-side collaborative support prediction are first extracted from the collaborative aggregation components. Based on the mapping relationship between the source-side, grid-side, and storage-side levels, the hierarchical position and correlation relationship corresponding to the cross-level collaborative parameters are determined.

[0080] Based on the predicted quantities of source-side collaborative injection, grid-side collaborative power supply, grid-side collaborative distribution, and energy storage collaborative support, the cross-level collaborative parameters are adjusted in a coordinated manner to synchronously correct the collaborative relationships between the source-side, grid-side, and energy storage levels. The adjusted cross-level collaborative parameters are then collected and organized with the corresponding predicted quantities of source-side collaborative injection, grid-side collaborative power supply, grid-side collaborative distribution, and energy storage collaborative support to obtain global correction data.

[0081] It should be noted that the collaborative aggregation component refers to the processing part formed after the source-side collaborative injection prediction, grid-side collaborative power supply prediction, grid-side collaborative power distribution prediction, and storage-side collaborative support prediction are aggregated and collected step by step. Cross-level coordination parameters refer to adjustment parameters used to characterize the power coordination relationship between the source-side, grid-side, and storage-side levels. These parameters include source-side injection allocation weights, grid-side power transmission correction coefficients, storage-side support sharing ratios, and inter-level power balance correction coefficients.

[0082] S4.3. Perform local refinement correction on the collaborative convergence components that are identified as having a deviation convergence state. Combine historical feasible parameters and current topological continuity to locally fine-tune the active parameters and obtain local refinement data.

[0083] Furthermore, when performing local refinement correction on the collaborative convergence components identified as being in a deviation convergence state, the source-side collaborative injection prediction, grid-side collaborative power supply prediction, grid-side collaborative power distribution prediction, and storage-side collaborative support prediction are first extracted from the collaborative convergence components. Combined with historical feasible parameters and the current topology continuity, the location of active parameters that need to be locally adjusted and the adjustment range of active parameters are determined.

[0084] Based on the parameter value relationships corresponding to historical feasible parameters, the active parameters are adjusted within a small range. In conjunction with the hierarchical connection continuity corresponding to the current topology continuity, the adjustment range and order of the active parameters are limited, so that the active parameters can be locally fine-tuned under the existing hierarchical correlation. Then, the locally fine-tuned active parameters are collected and organized with the corresponding source-side collaborative injection prediction, grid-side collaborative power supply prediction, grid-side collaborative power distribution prediction, and storage-side collaborative support prediction to obtain locally refined data.

[0085] It should be noted that active parameters refer to cross-level collaborative parameters that still have a direct impact on the inter-layer power deviation between the source-side collaborative injection prediction, grid-side collaborative power supply prediction, grid-side collaborative power distribution prediction, and storage-side collaborative support prediction, even when the deviation converges. The active parameters are determined based on the sensitivity of each cross-level collaborative parameter to inter-layer power deviation, the magnitude of parameter changes in the most recent iteration, and the contribution of deviation convergence.

[0086] S4.4 Iteratively filter the global correction data and local refinement data, and fine-tune the collaborative parameters by combining Gaussian mutation and elite retention mechanisms to obtain the source-network-storage collaborative prediction value.

[0087] It should be noted that the Gaussian mutation and elite preservation mechanism refers to the process of retaining the better combination of cooperative parameters during the adjustment of cooperative parameters, and applying Gaussian mutation to the cooperative parameters to be adjusted to generate new combinations of cooperative parameters and perform iterative comparison. Furthermore, when iteratively filtering global correction data and local refined data, cross-level collaborative parameters in global correction data and active parameters in local refined data are first extracted to form different combinations of collaborative parameters. Each combination of collaborative parameters is then applied to the current site-level prediction results to update the source-side collaborative injection prediction, grid-side collaborative power supply prediction, grid-side collaborative distribution prediction, and storage-side collaborative support prediction. Based on the updated collaborative predictions, the prediction error and inter-layer power deviation of the current round are recalculated. Different combinations of collaborative parameters are compared round by round according to the prediction error and inter-layer power deviation, and the combinations of collaborative parameters with smaller prediction errors and lower inter-layer power deviations are retained to form iterative filtering information. Based on iterative screening information, Gaussian mutation is performed on the retained collaborative parameter combinations to make minor adjustments to the collaborative parameter combinations around their existing values. At the same time, an elite retention mechanism is implemented on the better collaborative parameter combinations obtained in each round of screening so that the better collaborative parameter combinations can continue to participate in subsequent comparisons. After multiple rounds of fine-tuning, the better collaborative parameter combinations corresponding to the source-side collaborative injection prediction, grid-side collaborative power supply prediction, grid-side collaborative distribution prediction, and storage-side collaborative support prediction are extracted, and the source-grid-storage collaborative prediction values ​​are determined accordingly.

[0088] S5. Based on the source-grid-storage collaborative prediction value, publish it on the network side and call it in a collaborative manner. After the corresponding source-grid-storage measured operating value is reached, backfill the error and correct the deviation to obtain the updated collaborative dataset.

[0089] S5.1 Perform collaborative power decomposition and scheduling instruction conversion on the source-grid-storage collaborative prediction value to generate network-side release value and energy storage collaborative call value, and distribute them to the corresponding network side and energy storage side according to the unified scheduling identifier.

[0090] Furthermore, the source-grid-storage collaborative prediction values ​​are decomposed into collaborative power according to the correspondence between the source-side, grid-side, and storage-side levels to obtain power allocation results for the network side and energy storage side. The power allocation results are then converted into corresponding scheduling instructions to generate network-side release values ​​and energy storage collaborative call values. The network-side release values ​​and energy storage collaborative call values ​​are then matched with their corresponding objects and written into a unified scheduling identifier to form the corresponding network-side and energy storage-side distribution content, which is then distributed to the corresponding network-side and energy storage-side.

[0091] It should be noted that the unified scheduling identifier refers to the unified identification information used to associate the source-network-storage collaborative prediction value, the network-side published value, and the energy storage collaborative call value. The unified scheduling identifier is used to characterize the corresponding scheduling time, scheduling object, and scheduling category.

[0092] S5.2. Time-scale alignment, error backfilling, and deviation correction are performed on the measured operating values ​​of source, grid, and storage. The source-grid-storage collaborative prediction values, network-side published values, energy storage collaborative call values, and corrected measured backfill values ​​are collected and stored to obtain the updated collaborative dataset.

[0093] Furthermore, the measured operating values ​​of the source-grid-storage system are time-aligned according to the scheduling time corresponding to the unified scheduling identifier. The measured operating values ​​of the source-grid-storage system are then compared with the corresponding source-grid-storage collaborative prediction values, network-side published values, and energy storage collaborative call values ​​to obtain error data under each correspondence. The error data is written into the corresponding prediction record to complete error backfilling. Based on the error data, the source-grid-storage collaborative prediction values ​​are corrected for deviations to obtain deviation-corrected prediction values. Finally, the source-grid-storage collaborative prediction values, network-side published values, energy storage collaborative call values, measured operating values ​​of the source-grid-storage system, error data, and deviation-corrected prediction values ​​are collected and stored to obtain an updated collaborative dataset.

[0094] This embodiment also provides a computer device applicable to the intelligent power prediction method based on distributed photovoltaics, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the intelligent power prediction method based on distributed photovoltaics as proposed in the above embodiment.

[0095] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0096] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the intelligent power prediction method based on distributed photovoltaics as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0097] In summary, this invention achieves accurate characterization of the dynamic coupling relationship between power sources, grid, and energy storage in the distribution network by: compiling distribution network operation status data and constructing a dynamic electrical topology diagram based on real-time switch status; constraining attention weights using an adjacency matrix mask; and thus improving the effectiveness of feature extraction and the reliability of prediction under complex operating conditions. Furthermore, by constructing a joint fitness function that integrates power balance deviation and prediction error, and using a dynamic inertial weighted particle swarm optimization algorithm for online optimization, the invention enables adaptive adjustment of prediction parameters to nonlinear operating conditions, thereby improving the accuracy and coordination of station-level power supply and distribution prediction. Finally, by combining hierarchical aggregation, deviation classification correction, and measured backfilling updates, the invention further realizes closed-loop coordination of prediction, scheduling, and correction, ultimately improving the collaborative control capability of power sources, grid, and energy storage, and the operational stability of the distribution network.

[0098] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A smart power prediction method based on distributed photovoltaic power, characterized in that, include: Data on the operation status of the distribution network is collected, and the data is processed to obtain a source-grid-storage dataset. An electrical topology dynamic graph is constructed based on the source-grid-storage dataset and the real-time switch status. The attention weights are constrained by the adjacency matrix mask to obtain a spatiotemporal weighted feature sequence. The temporal features of strongly coupled nodes are extracted from the spatiotemporal weighted feature sequence. A joint fitness function is constructed based on the power balance deviation of the distribution network and the prediction error. The hyperparameters are optimized online using the dynamic inertial weighted particle swarm optimization algorithm to adapt to the nonlinearity of the operating condition, and the power supply and distribution prediction at the site level is obtained. Based on the predicted power supply and distribution at the site level and the measured photovoltaic output in the source-grid-storage dataset, the sequential transmission relationship of the same-direction ramping between sites is identified, forming a cloud-shadow output propagation chain. The prediction residuals of the sites that have already ramped are extracted along the cloud-shadow output propagation chain, and the predictions of the sites that have not yet ramped are corrected in relay according to the propagation delay and amplitude attenuation relationship, resulting in a cloud-shadow relay corrected prediction set. The inter-layer power deviation is calculated according to the three-level convergence of source-grid-storage, and the deviation over-limit state and deviation convergence state are obtained. The distribution network safety threshold is used to identify fluctuating operating conditions and trigger a dynamic inertial weight strategy. Global correction and local refinement correction are performed on the deviation over-limit state and the deviation convergence state, respectively. The Gaussian mutation and elite retention mechanism are used to iteratively fine-tune the collaborative parameters to obtain the source-grid-storage collaborative prediction value. Based on the source-grid-storage collaborative prediction values, the network side publishes and the energy storage collaborative calls are carried out. After the corresponding source-grid-storage measured operating values ​​are reached, error backfilling and deviation correction are performed to obtain the updated collaborative dataset.

2. The intelligent power prediction method based on distributed photovoltaic power as described in claim 1, characterized in that, The process of collecting distribution network operation status data and then organizing this data to obtain the source-network-storage dataset is as follows: Collect power distribution network operation data, distributed photovoltaic power output data, and energy storage operation data to obtain power distribution network operation status data; The distribution network operation status data is processed with unified time stamp and unified dimension to obtain standardized operation data. The standardized operation data is then compiled, verified, anomaly repaired, and collected and stored to obtain the source network storage dataset.

3. The intelligent power prediction method based on distributed photovoltaic power as described in claim 2, characterized in that, The process involves constructing a dynamic electrical topology graph based on the source-network-storage dataset and real-time switch states, and then constraining the attention weights using an adjacency matrix mask to obtain a spatiotemporally weighted feature sequence. The specific steps are as follows: A dynamic electrical topology graph that changes over time is constructed based on the source-network-storage dataset and real-time switch status. Temporal features of nodes are extracted from the dynamic graph of electrical topology. Attention scores of unconnected nodes are masked by adjacency matrix masking. After normalizing the attention weights of adjacent nodes, feature weighting and fusion are performed to obtain a spatiotemporally weighted feature sequence.

4. The intelligent power prediction method based on distributed photovoltaic power as described in claim 3, characterized in that, The process involves extracting time-series features of strongly coupled nodes from the spatiotemporal weighted feature sequence, constructing a joint fitness function based on the power balance deviation and prediction error of the distribution network, and using a dynamic inertial weighted particle swarm optimization algorithm to optimize hyperparameters online to adapt to the nonlinearity of the operating conditions, thereby obtaining the site-level power supply and distribution prediction. The specific steps are as follows: Extract the temporal features of strongly coupled nodes corresponding to the target site from the spatiotemporally weighted feature sequence, and construct the site-level input features; The station-level input features are encoded with operating conditions and matched with historical operating conditions. Combined with the real-time topology status, time-series prediction parameters adapted to the current operating conditions are generated to form a station-level prediction configuration. Using site-level prediction configuration, the power supply, power distribution, source-side injection, energy storage support, and load absorption at the site boundary are predicted simultaneously to obtain the predicted values ​​of power supply, power distribution, source-side injection, energy storage support, and load absorption at the site boundary, and to form the prediction error. A joint fitness function is constructed, and the prediction error and power balance deviation are input into the joint fitness function to calculate the fitness and obtain the fitness evaluation. Based on the fitness evaluation, the hyperparameters are optimized and updated online using the dynamic inertia weighted particle swarm optimization algorithm to obtain the site-level power supply and distribution prediction.

5. The intelligent power prediction method based on distributed photovoltaic power as described in claim 4, characterized in that, The process of identifying the sequential transmission relationship of unidirectional ramp-up between sites based on the predicted power supply and distribution at the site level and the measured photovoltaic output in the source-grid-storage data set, forming a cloud-shadow power output propagation chain, is as follows: The predicted values ​​of source-side injection for each photovoltaic site are extracted from the power supply and distribution forecast at the site level. According to the combination of site identification and collection time scale with the actual photovoltaic output, the change in actual output between adjacent collection time scales and the prediction deviation between actual output and source-side injection predicted value are calculated. The changes in actual output with the same direction of change and continuous collection time scale are connected in sequence to form the photovoltaic output ramping section. By comparing photovoltaic power output ramping sections with consistent changes in direction across different photovoltaic sites in multiple complete ramping processes within the source-grid-storage dataset, and moving along the acquisition timescale to the photovoltaic power output ramping section with an earlier start time, the time offset at which the change sequence corresponds to the highest degree is determined as the propagation delay. The predecessor and successor sites are determined according to the order of their start times, and the amplitude attenuation relationship is determined according to the ratio of the cumulative change amplitudes. The predecessor and successor sites are connected according to the order of their start times to form the cloud shadow power output propagation chain.

6. The intelligent power prediction method based on distributed photovoltaic power as described in claim 5, characterized in that, The prediction residuals of stations that have experienced ramp-up are extracted along the cloud shadow power propagation chain. The prediction quantities of stations that have not experienced ramp-up are then corrected in relay according to the propagation delay and amplitude attenuation relationship to obtain the cloud shadow relay corrected prediction quantity set. The inter-layer power deviation is then calculated according to the source-grid-storage three-level convergence to obtain the deviation over-limit state and deviation convergence state. The specific steps are as follows: According to the starting time, each photovoltaic site is searched along the cloud shadow power output propagation chain. Sites that have formed a photovoltaic power output ramping section are identified as ramping sites, and subsequent sites that have not yet reached the expected ramping time are identified as non-ramping sites. The prediction deviation between the measured power output of the ramping sites and the predicted value of the source-side injection is extracted to obtain the precursor prediction residual. Read the cumulative propagation delay and amplitude attenuation relationship from the site where the climb has occurred to the site where the climb has not occurred, move the precursor prediction residual to the expected climb period and convert it into a prediction correction amount, and add it to the corresponding source side injection amount prediction value; when the closer precursor site climbs, use the corresponding precursor prediction residual to replace the original precursor prediction residual to obtain the cloud shadow relay correction prediction amount set. The cloud shadow relay correction prediction set is mapped in three levels: source-network-storage. The inter-layer power deviation is calculated by combining inter-layer transmission loss and topology continuity. When the inter-layer power deviation exceeds the allowable deviation range, it is determined to be a deviation over-limit state. When it continuously decreases and returns to the allowable deviation range, it is determined to be a deviation convergence state.

7. The intelligent power prediction method based on distributed photovoltaic power as described in claim 6, characterized in that, The process involves using a distribution network safety threshold to identify fluctuating operating conditions and triggering a dynamic inertial weighting strategy. Global correction and local refinement corrections are performed for deviation exceeding limits and deviation convergence states, respectively. Furthermore, Gaussian mutation and elite retention mechanisms are used to iteratively fine-tune the collaborative parameters to obtain the source-grid-storage collaborative prediction value. The specific steps are as follows: Extract line load rate, voltage deviation, switch action density and energy storage regulation margin from the distribution network safety threshold, identify fluctuating operating conditions and trigger dynamic inertial weighting strategy; Global correction is performed on the collaborative convergence components that are identified as having excessive deviations. The cross-level collaborative parameters are adjusted as a whole by combining the collaborative quantities of the source side, the grid side, and the storage side to obtain global correction data. Local refinement correction is performed on the collaborative convergence components that are identified as being in a state of biased convergence. The active parameters are then fine-tuned locally by combining historical feasible parameters and the current topological continuity to obtain local refinement data. The source-network-storage collaborative prediction value is obtained by iteratively filtering the global correction data and the local refinement data, and by combining Gaussian mutation and elite retention mechanisms to fine-tune the collaborative parameters.

8. The intelligent power prediction method based on distributed photovoltaic power as described in claim 7, characterized in that, The process of publishing network-side data and invoking energy storage collaboratively based on the source-grid-storage collaborative prediction values, and performing error backfilling and deviation correction after the corresponding measured operating values ​​of the source-grid-storage system are reached, to obtain an updated collaborative dataset, is as follows: The source-grid-storage collaborative prediction values ​​are decomposed into collaborative power and converted into scheduling instructions to generate network-side release values ​​and energy storage collaborative call values, which are then distributed to the corresponding network side and energy storage side according to a unified scheduling identifier. The measured operating values ​​of source, grid, and storage are time-scaled, backfilled with errors, and corrected for deviations. The source-grid-storage collaborative prediction values, network-side published values, energy storage collaborative call values, and corrected measured backfill values ​​are collected and stored to obtain the updated collaborative dataset.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent power prediction method based on distributed photovoltaics as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent power prediction method based on distributed photovoltaics as described in any one of claims 1 to 8.

Images