Photovoltaic power prediction method, system, electronic device, and medium

Abstract

The present disclosure relates to the technical field of new energy, and provides a photovoltaic power prediction method and system, an electronic device and a medium. The method comprises: obtaining a first sub-power prediction result meeting physical consistency and explainability according to physical correlation characteristics, obtaining a second sub-power prediction result containing spatial correlation according to spatial correlation characteristics, and obtaining a third sub-power prediction result quantifying uncertainty according to uncertainty correlation characteristics; and performing calibration or fusion operation on a first power prediction result driven by pure data based on the three sub-power prediction results, so as to obtain a final photovoltaic power prediction result, thereby solving the problems of lack of physical consistency, neglect of spatial correlation, unquantifiable uncertainty and insufficient explainability in the existing photovoltaic power prediction.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of new energy, in particular to a photovoltaic power prediction method and system, an electronic device and a medium. BACKGROUND

[0002] As an important part of clean energy, photovoltaic power generation has significant intermittent, fluctuating and random characteristics, which poses a challenge to the stable operation and efficient dispatch of power grids. Achieving high-precision photovoltaic power prediction is a key technical support for improving photovoltaic power consumption capacity and ensuring the safe and economic operation of power grids. SUMMARY

[0003] The embodiments of the present disclosure provide a photovoltaic power prediction method, system, electronic device and medium, which are beneficial to solve the problems of lack of physical consistency, neglect of spatial correlation, inability to quantify uncertainty and lack of interpretability in existing photovoltaic power prediction.

[0004] According to some embodiments of the present disclosure, the first aspect of the embodiments of the present disclosure provides a photovoltaic power prediction method, the method comprising: obtaining a first feature set, the first feature set comprising basic features, physical correlation features, spatial correlation features and uncertainty correlation features, the basic features comprising electrical features, meteorological features and calendar features of a photovoltaic module, the physical correlation features comprising photovoltaic module parameters, installation information and solar trajectory information, the spatial correlation features comprising photovoltaic site coordinates, distance relationships between photovoltaic sites, power grid topology and regional attributes, and the uncertainty correlation features comprising meteorological forecast errors, historical prediction errors, data quality and time expansion features; determining a first power prediction result according to the basic features, the first power prediction result comprising a basic prediction result of photovoltaic power; determining an adjusted power prediction result according to one or more of the physical correlation features, the spatial correlation features and the uncertainty correlation features, the adjusted power prediction result comprising one or more of a first sub-power prediction result, a second sub-power prediction result and a third sub-power prediction result, the first sub-power prediction result comprising a physical constraint prediction result and a physical parameter estimation result of the photovoltaic power, the first sub-power prediction result conforming to a physical law corresponding to the physical correlation features, the second sub-power prediction result comprising a spatial correlation prediction result and a site influence weight of the photovoltaic power, the second sub-power prediction result being used to describe spatial correlation corresponding to the spatial correlation features, and the third sub-power prediction result comprising a probability interval prediction result and an uncertainty decomposition result of the photovoltaic power; and determining one or more of a point prediction result, a probability prediction interval, a confidence interval, a complete probability distribution, a physical parameter estimation result, a spatial correlation mode and an uncertainty decomposition result of the photovoltaic power according to the first power prediction result and the adjusted power prediction result.

[0005] According to some embodiments of this disclosure, a second aspect of this disclosure provides a photovoltaic power prediction system, including one or more modules for performing the photovoltaic power prediction method provided in the first aspect. For example, the system may include an acquisition module for performing the step of "acquiring a first feature set". As another example, the system may include a processing module for performing steps other than "acquiring the first feature set".

[0006] According to some embodiments of the present disclosure, a third aspect of the present disclosure provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method as described in the first aspect.

[0007] According to some embodiments of the present disclosure, a fourth aspect of the present disclosure provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the method described in the first aspect.

[0008] In some embodiments of this disclosure, determining the regulation power prediction result based on one or more of physical correlation features, spatial correlation features, and uncertainty correlation features includes: determining the first sub-power prediction result using a physical constraint prediction module based on the physical correlation features.

[0009] In some embodiments of this disclosure, the physical constraint prediction module includes a physical constraint branch. Based on physically relevant features, the physical constraint prediction module determines the prediction result of the first sub-power, including: determining the physical parameter estimation result based on the physically relevant features and the physical parameter estimation result; and determining the physical constraint prediction result based on the physically relevant features and the physical parameter estimation result.

[0010] In some embodiments of this disclosure, determining a first sub-power prediction result using a physical constraint prediction module based on physically relevant features includes: inputting the physically relevant features into the physical constraint prediction module to obtain a second power prediction result; and obtaining a first sub-power prediction result based on the first power prediction result and the second power prediction result.

[0011] In some embodiments of this disclosure, the physical constraint prediction module includes a physical constraint branch; inputting physically relevant features into the physical constraint prediction module to obtain a second power prediction result includes: determining physical parameter estimation results using the physical constraint branch based on the physically relevant features; and determining physical model prediction results based on the physically relevant features and the physical parameter estimation results.

[0012] The first sub-power prediction result is obtained based on the first power prediction result and the second power prediction result, including: obtaining a physical constraint prediction result based on the first power prediction result and the physical model prediction result.

[0013] In some embodiments of the present disclosure, the adjusted power prediction result is determined according to one or more of the physical correlation feature, the spatial correlation feature and the uncertainty correlation feature, including: determining a second sub-power prediction result by using a heterogeneous graph prediction module according to the spatial correlation feature.

[0014] In some embodiments of the present disclosure, the second sub-power prediction result is determined by using the heterogeneous graph prediction module according to the spatial correlation feature, including: obtaining updated information of the spatial dependency relationship between the photovoltaic power station, the weather station and the grid node according to the meteorological change feature and the prediction error, the prediction error being a deviation between the spatial correlation prediction result of the photovoltaic power and the observation value; updating the heterogeneous graph prediction module according to the updated information of the spatial dependency relationship between the photovoltaic power station, the weather station and the grid node; inputting the spatial correlation feature into the updated heterogeneous graph prediction module to obtain the spatial correlation prediction result of the photovoltaic power and the station influence weight.

[0015] In some embodiments of the present disclosure, the heterogeneous graph prediction module is updated according to the updated information of the spatial dependency relationship between the photovoltaic power station, the weather station and the grid node, including: updating one or more of the nodes, edges and edge weights of the heterogeneous graph prediction module according to the updated information of the spatial dependency relationship between the photovoltaic power station, the weather station and the grid node.

[0016] In some embodiments of the present disclosure, the adjusted power prediction result is determined according to one or more of the physical correlation feature, the spatial correlation feature and the uncertainty correlation feature, including: determining a third sub-power prediction result by using a probability interval prediction module according to the uncertainty correlation feature, the probability interval prediction module including one or more of a point prediction module, a conditional variational auto-encoding module, an adaptive kernel density estimation module and a quantile regression module, and being used to determine a probability interval prediction result of the photovoltaic power and an uncertainty decomposition result according to the uncertainty correlation feature.

[0017] In some embodiments of the present disclosure, the adjusted power prediction result is determined according to one or more of the physical correlation feature, the spatial correlation feature and the uncertainty correlation feature, including: determining the third sub-power prediction result by using the probability interval prediction module according to the uncertainty correlation feature, the first sub-power prediction result and the second sub-power prediction result.

[0018] In some embodiments of the present disclosure, the determination of one or more of the point prediction result, the probability prediction interval, the confidence interval, the complete probability distribution, the physical parameter estimation result, the spatial correlation pattern, and the uncertainty decomposition result of the photovoltaic power according to the first power prediction result and the adjusted power prediction result comprises: fusing the first power prediction result, the first sub-power prediction result, the second sub-power prediction result, and the third sub-power prediction result according to one of weighting, concatenation, and adaptive selection to obtain one or more of the point prediction result, the probability prediction interval, the confidence interval, the complete probability distribution, the physical parameter estimation result, the spatial correlation pattern, and the uncertainty decomposition result of the photovoltaic power.

[0019] The technical solutions provided by the embodiments of the present disclosure have at least the following advantages:

[0020] By constructing and fusing the first feature set containing four categories of features including basic, physical, spatial, and uncertainty, the data-driven rules, physical mechanism constraints, spatial correlation effects, and uncertainty information are systematically integrated into a unified framework, realizing the deep fusion and collaborative prediction of multi-dimensional and multi-source information, which fundamentally overcomes the limitations of single data-driven models and single physical models, and provides a comprehensive information basis for generating high-precision and high-reliability prediction results. By introducing the physical constraint prediction result and explicitly requiring it to "meet the physical laws corresponding to the physical related features", the prediction that may be generated by the pure data-driven model in violation of the physical common sense (such as non-zero power output at night) is effectively avoided, not only ensuring the physical rationality of the prediction result, but also improving the reliability and credibility of the prediction result, and enhancing the generalization ability and robustness of the photovoltaic power prediction model in the scene not covered by the training data (such as extreme weather and new components).

[0021] By generating the spatial correlation prediction result and the station influence weight, this method can explicitly model and quantitatively describe the mutual influence between various stations (such as photovoltaic power stations, weather stations, and power grid nodes) dispersed in different geographical locations, accurately depicting the dynamic spatial correlation characteristics of regional photovoltaic clusters, breaking through the traditional mode of predicting each photovoltaic power station as an independent individual, realizing the upgrade from "point" to "surface", and significantly improving the overall prediction accuracy of regional or cluster photovoltaic power, especially suitable for accurate prediction of short-term fluctuations caused by cloud movement and weather system changes.

[0022] By outputting the probability interval prediction result and the uncertainty decomposition result, the transition from "point prediction" to "probability prediction" is completed, the complete quantitative information of prediction uncertainty is provided in the photovoltaic power prediction scheme for the first time, so that the user can not only obtain the future power point estimate value, but also understand the possible fluctuation range (confidence interval) of the prediction value and the source of the uncertainty (such as random error or model cognitive error), thereby providing important information support for risk assessment, reserve capacity configuration, robust scheduling and other risk perception type decisions of the power system, and realizing the technical leap from deterministic prediction to probabilistic prediction.

[0023] The final output physical parameter estimation result, spatial correlation mode, uncertainty decomposition result and other derived information jointly constitute a powerful model explanation capability, forming an interpretable, traceable and transparent prediction framework, so that the user can clearly understand "why the prediction result is like this", which greatly enhances the user's trust and acceptance of the photovoltaic power prediction model, and solves the pain point that the photovoltaic power prediction model is often regarded as a "black box".

[0024] In summary, based on the technical scheme provided by the embodiment of the present disclosure, through its unique "multi-source feature fusion → basic prediction and multi-path adjustment → result integration output" technical architecture, the core problems of physical inconsistency, spatial correlation neglect, uncertainty unquantifiable, and prediction result uninterpretable in the prior art are systematically solved. In particular, by essentially improving the basic time series model (deep physical constraint, dynamic graph update, uncertainty decomposition), constructing a system-level collaborative optimization architecture (multi-objective optimization, intelligent scene perception, information flow cascade enhancement), and providing an intelligent output interface carrying innovative value, a photovoltaic power prediction solution is finally achieved, which significantly surpasses the existing technical level in terms of prediction accuracy, reliability, interpretability and risk management support capability. The technical scheme provided by the embodiment of the present disclosure has great improvement in core algorithm, system architecture and application function. BRIEF DESCRIPTION OF DRAWINGS

[0025] One or more embodiments are exemplarily illustrated by pictures in the corresponding drawings, which do not constitute a limitation on the embodiments unless specifically stated. The drawings in the drawings do not constitute a proportional limitation; in order to more clearly illustrate the technical solutions in the embodiments of the present disclosure or in the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description can only be some embodiments of the present disclosure, and those skilled in the art can also obtain other drawings according to these drawings without creating any creative labor.

[0026] Figure 1A A structural schematic diagram of a photovoltaic power prediction system provided by an embodiment of the present disclosure;

[0027] Figure 1B A structure diagram of a physical constraint prediction module in Figure 1A ;

[0028] Figure 1C A structure diagram of a physical constraint prediction module in Figure 1A ;

[0029] Figure 1D A structure diagram of a heterogeneous graph prediction module in Figure 1A ;

[0030] Figure 1E A structure diagram of a probability interval prediction module in Figure 1A ;

[0031] Figure 2 A flow diagram of a photovoltaic power prediction method provided by an embodiment of the present disclosure;

[0032] Figure 3 A flow diagram of another photovoltaic power prediction method provided by an embodiment of the present disclosure;

[0033] Figure 4 A flow diagram of still another photovoltaic power prediction method provided by an embodiment of the present disclosure;

[0034] Figure 5 A flow diagram of still another photovoltaic power prediction method provided by an embodiment of the present disclosure;

[0035] Figure 6 A flow diagram of still another photovoltaic power prediction method provided by an embodiment of the present disclosure;

[0036] Figure 7 A flow diagram of still another photovoltaic power prediction method provided by an embodiment of the present disclosure;

[0037] Figure 8 A flow diagram of still another photovoltaic power prediction method provided by an embodiment of the present disclosure;

[0038] Figure 9 A flow diagram of still another photovoltaic power prediction method provided by an embodiment of the present disclosure;

[0039] Figure 10 A flow diagram of still another photovoltaic power prediction method provided by an embodiment of the present disclosure;

[0040] Figure 11 A flow diagram of still another photovoltaic power prediction method provided by an embodiment of the present disclosure;

[0041] Figure 12A flowchart of another photovoltaic power prediction method provided by an embodiment of the present disclosure is shown in FIG. 6.

[0042] Figure 13 A structural diagram of another photovoltaic power prediction system provided by an embodiment of the present disclosure is shown in FIG. 7.

[0043] Figure 14 A structural diagram of an electronic device provided by an embodiment of the present disclosure is shown in FIG. 8. DETAILED DESCRIPTION

[0044] In the description of the embodiments of the present disclosure, the technical terms "first", "second", and the like are only used to distinguish different objects, and cannot be understood as indicating or implying relative importance or implicitly indicating the number, specific order or primary and secondary relationship of the indicated technical features. In the description of the embodiments of the present disclosure, the meaning of "multiple" is two or more, unless otherwise explicitly and specifically limited. Similarly, "multiple groups" means two or more groups (including two groups), and "multiple pieces" means two or more pieces (including two pieces).

[0045] In this document, referring to "embodiments" means that the specific features, structures or characteristics described in connection with the embodiments can be included in at least one embodiment of the present disclosure. The appearance of this phrase in various places in the specification does not necessarily mean the same embodiment, nor is it an independent or alternative embodiment to other embodiments. The skilled in the art explicitly and implicitly understands that the embodiments described herein can be combined with other embodiments.

[0046] In the description of the embodiments of the present disclosure, the term "and / or" is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the existence of A, the existence of A and B, and the existence of B. In addition, the character " / " in this document generally represents a "or" relationship between the front and rear associated objects.

[0047] The terms used in the description of various embodiments described herein are only used to describe specific embodiments and are not intended to be limiting.

[0048] At present, photovoltaic power prediction technology has gradually developed from traditional physical modeling methods to data-driven methods represented by deep learning. The existing technical solutions mainly focus on the following directions.

[0049] The first is a pure data-driven prediction model, such as using a convolutional neural network (CNN), a long short-term memory (LSTM) network, or a hybrid model (such as Convolutional Neural Network-Long Short-Term Memory, CNN-LSTM) to model the time series of historical power data and weather data. Such methods can learn complex nonlinear relationships from massive data and improve the accuracy of point prediction to some extent. However, they are essentially black box models that rely entirely on data statistics and lack consideration of the inherent physical mechanisms of photovoltaic power generation. This leads to insufficient generalization ability of the model outside the training data distribution (such as extreme weather and equipment abnormalities), and even may produce prediction results that violate basic physical principles, such as predicting non-zero power at night or predicting power to rise when irradiance drops sharply, which seriously affects the reliability and practicality of the prediction results.

[0050] The second is the cluster prediction of multiple photovoltaic power stations. Existing research has attempted to introduce graph neural networks (GNN) to capture spatial correlations between sites. However, existing methods are mostly limited to constructing simple homogeneous graphs or treating graph structures as static and unchanging, failing to effectively depict the complex and dynamic interaction relationships between various heterogeneous entities such as photovoltaic power stations, weather stations, and grid nodes in real environments. For short-term and dynamic spatial correlation changes caused by weather system movement, local cloud blocking, and other factors, existing static graph models are difficult to accurately capture and adaptively respond.

[0051] The third is in the form of prediction output. Most existing technologies still focus on providing a single point prediction result, which cannot quantify the uncertainty of the prediction result. However, in actual power system dispatching and energy management decision-making, understanding the possible fluctuation range of the prediction value (i.e., the prediction interval) and its reliability is crucial for risk assessment, backup capacity configuration, and robust optimization. Although individual studies have begun to explore probabilistic prediction, they usually only use a single technique (such as quantile regression or Bayesian methods) and fail to deeply collaborate with physical mechanisms and spatial-temporal correlations, resulting in a large room for improvement in the accuracy and reliability of the probability interval.

[0052] In addition, the fusion solutions of the prior art are mostly simple model stacking or static weight combination, lacking systematic collaborative optimization and intelligent decision-making mechanism, failing to achieve the synergistic effect of "1+1+1>3", and lacking a complete and efficient solution in terms of how to systematically integrate physical constraints, dynamic spatial correlation and uncertainty into the prediction links of different time scales (day-ahead, intra-day, ultra-short-term) and realize efficient fusion and adaptive adjustment of information between links.

[0053] In summary, the existing photovoltaic power prediction technology has obvious technical gaps and improvement needs in terms of physical consistency guarantee, dynamic spatio-temporal correlation modeling, uncertainty quantification, and multi-stage prediction collaboration. Therefore, there is an urgent need in the field for a new prediction system that can systematically integrate physical knowledge, spatio-temporal dynamics, and probability information to overcome the limitations of existing technology and achieve high-precision, high-reliability, and highly interpretable photovoltaic power prediction.

[0054] The embodiments of the present disclosure provide a photovoltaic power prediction method, system, electronic device and medium, through its unique technical architecture of "multi-source feature fusion → basic prediction and multi-path adjustment → result integration output", it can systematically solve the core problems of physical inconsistency, spatial correlation neglect, uncertainty unquantifiable, and prediction results uninterpretable in the prior art. In particular, by fundamentally improving the basic time series model (deep physical constraints, dynamic graph update, uncertainty decomposition), building a system-level collaborative optimization architecture (multi-objective optimization, intelligent scene perception, information flow cascade enhancement), and providing an intelligent output interface that carries innovative value, a photovoltaic power prediction solution is ultimately achieved that significantly surpasses the existing technology in terms of prediction accuracy, reliability, interpretability, and risk management support capability. The technical solutions provided by the embodiments of the present disclosure have great improvement in core algorithms, system architecture, and application functions.

[0055] The embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings. However, those skilled in the art can understand that in the embodiments of the present disclosure, many technical details are presented to help the reader better understand the present disclosure. However, the technical solutions claimed by the present disclosure can be implemented even without these technical details and various changes and modifications based on the following embodiments.

[0056] First, the photovoltaic power prediction system provided by the embodiments of the present disclosure will be described. Figures 1A-1E The photovoltaic power prediction system provided by the embodiments of the present disclosure will be described.

[0057] Figure 1AA structural schematic diagram of a photovoltaic power prediction system provided by an embodiment of the present disclosure. The system can be referred to as a Physics-Graph-Uncertainty Quantification for Pixel and Vertex Features (PG-UQ-PV) prediction system. As shown in Figure 1A the system includes a data input layer, a data preprocessing and feature extraction module, a core prediction module, such as a basic prediction module, a physical constraint prediction module, a heterogeneous graph prediction module, a probability interval prediction module, and a multi-stage rolling prediction controller, an integration and decision module, a cloud microservice and energy management system (EMS) interface module, an application layer, etc.

[0058] The data input layer is responsible for inputting data, such as photovoltaic power station monitoring data, meteorological observation data, power grid topology data, and static parameter data.

[0059] The data preprocessing and feature extraction module is responsible for data cleaning, data filling, normalization, feature construction, etc. on the input raw data, and extracts features to construct a rich feature set, including basic features (such as electrical features, meteorological features, calendar features), and physical related features (such as component parameters), spatial related features (such as site distance, topology relationship), and uncertainty related features.

[0060] The basic prediction module is used to perform pure data-driven photovoltaic power prediction based on the basic features, and obtain a basic prediction result. The basic prediction result can be used as a benchmark for physical consistency prediction, spatio-temporal correlation prediction, and probability interval prediction of the photovoltaic power prediction system provided by the present disclosure. The final output prediction result of the photovoltaic power prediction system can be considered as a prediction result calibrated or adjusted based on the basic prediction result using the prediction results output by the physical constraint prediction module, the heterogeneous graph prediction module, and the probability interval prediction module. For example, the prediction results output by the physical constraint prediction module, the heterogeneous graph prediction module, and the probability interval prediction module can be used in a cascaded or parallel manner to calibrate or adjust the basic prediction result.

[0061] The basic prediction module can be constructed based on a deep learning network, such as a Multi-Scale Convolutional Neural Network Bidirectional Long Short-Term Memory (MS-CNN-BiLSTM) network, and can include a multi-scale CNN layer, a BiLSTM layer, a full connection layer, and the like, and is configured to obtain a first power prediction result based on basic features. In the specific embodiments of the present disclosure, the first power prediction result includes a basic prediction result of photovoltaic power.

[0062] It should be noted that the basic prediction module can exist as a module together with the physical constraint prediction module, the heterogeneous graph prediction module, and the probability interval prediction module, or can exist as a sub-module of the physical constraint prediction module, and the embodiments of the present disclosure do not limit this.

[0063] The physical constraint prediction module, i.e., a Physics-Informed Neural Network for Pixel and Vertex Features (PINN-PV) module, can also be referred to as a physics-informed neural network module, a physics-constrained neural network module, a PINN-PV network, or a PINN-PV module, and is one of the core innovative modules provided by the embodiments of the present disclosure. The physical constraint prediction module is constructed based on a deep learning network, embeds a set of physical equations of photovoltaic power generation as a physical constraint, and generates a power prediction result that conforms to physical laws by optimizing a hybrid loss function that includes data loss and physical loss. Unlike existing technologies that only use data-driven models or simply introduce physical features, the physical constraint prediction module provided by the embodiments of the present disclosure is deeply embedded with physical constraints, such as a physical law of photovoltaic power generation (such as a solar irradiance model, a component attenuation model, and an inverter efficiency curve model) in the form of a partial differential equation or an algebraic equation, which is deeply embedded into a loss function of a photovoltaic power prediction model to form a hybrid loss function. This forces the photovoltaic power prediction model to strictly follow physical laws in the training process, so that it can still make reasonable inferences based on physical laws in the case of sparse data or extreme working conditions, significantly improving the reliability, credibility, and generalization ability of the photovoltaic power prediction model, and enhancing the physical interpretability of the prediction result.

[0064] In some embodiments of the present disclosure, as Figure 1BAs shown, the physical constraint prediction module can include a basic timing branch and a physical constraint branch. The basic timing branch can be constructed by juxtaposing the basic prediction module as a submodule of the physical constraint prediction module with the physical constraint branch. The physical constraint branch can include a parameter estimation layer, a physical model calculation layer, etc., for obtaining physical parameter estimation results and physical model prediction results based on physical related features. Then, the basic prediction results and the physical model prediction results are fused, such as calculating the weighted sum of the two, to obtain the first sub-power prediction results. In the specific embodiments of the present disclosure, the first sub-power prediction results include the fused physical constraint prediction results, to ensure that the photovoltaic power prediction results conform to the physical law, thereby improving the physical consistency.

[0065] In addition, as Figure 1C As shown, in the training process of the physical constraint prediction module, the data error and the physical residual can also be calculated based on the basic prediction results, the first sub-power prediction results, and the photovoltaic power observation value calculation data, and the hybrid loss function value of the photovoltaic power is calculated based on the data error and the physical residual, to update the weights of the basic timing branch and / or the physical constraint branch, thereby improving the prediction accuracy and the physical consistency.

[0066] In the specific embodiments of the present disclosure, the heterogeneous graph prediction module is a heterogeneous graph neural network (HGNN-PV) module for pixel and vertex features. The heterogeneous graph prediction module is used to construct a graph structure containing heterogeneous nodes such as photovoltaic power stations, weather stations, and grid nodes, so as to learn the spatio-temporal dependency between nodes by using graph neural network technology, and realize collaborative prediction of regional photovoltaic station group power, such as obtaining the second sub-power prediction results. Unlike the existing technology which regards the graph structure as static and unchangeable, the heterogeneous graph prediction module provided in the embodiments of the present disclosure introduces a dynamic graph updating mechanism. The structure of the graph (connection and weight between nodes) can be dynamically adjusted according to the latest real-time data (such as weather changes, prediction errors, and power station operating states). This mechanism can accurately capture short-term and dynamic spatio-temporal dependencies caused by local cloud movement, weather system changes, etc., so that the photovoltaic power prediction model can better adapt to the strong randomness and volatility of photovoltaic power generation, and significantly improve the collaborative prediction accuracy of regional photovoltaic groups.

[0067] In some embodiments of the present disclosure, as Figure 1DAs shown, the heterogeneous graph prediction module can perform a node feature initialization operation on the input features (spatial position, topological relationship, station attribute, meteorological data) to obtain the node features of each node (such as a photovoltaic node, a meteorological node, and a power grid node), and determine the dependency relationship between each node based on the node features to construct the edges between the nodes, and the weight of the edge is used to describe the influence size between the nodes. In the embodiments of the present disclosure, the dependency relationship between each node is also defined as an edge relationship, which includes a spatial proximity relationship, an electrical connection relationship, a functional association relationship, etc. Then, based on the node features and the edge relationship, a heterogeneous graph neural network layer is constructed and time series modeling is performed (such as a time convolution network and a runaway attention mechanism). In the embodiments of the present disclosure, the heterogeneous graph neural network layer includes multiple relational graph convolutional network (RGCN) layers, and the time series modeling can be performed by using the time convolution and the runaway attention mechanism. Subsequently, hierarchical prediction is performed to obtain the prediction results of each prediction stage, such as power grid level prediction results, regional level prediction results, cluster level prediction results, and station level prediction results. In addition, the heterogeneous graph neural network can also be dynamically updated according to the uncertainty features such as meteorological changes, prediction errors, and data quality, such as calculating a loss function value to adjust the edge weight and optimize the graph topology structure, so as to improve the prediction accuracy of the heterogeneous graph prediction module.

[0068] The probabilistic interval prediction module, i.e., a probabilistic interval for pixel and vertex features (PI-PV) prediction module, is configured to generate a complete probability distribution or confidence interval of the prediction result based on the output of the basic prediction module (e.g., the first power prediction result) or directly according to the first feature set, and to quantitatively decompose the sources of uncertainty. In the specific embodiments of the present disclosure, the probabilistic prediction technology can include one or more of the following: conditional variational autoencoder (CVAE), adaptive kernel density estimation (AKDE), quantile regression (QR), etc. The innovation of the probabilistic interval prediction module provided in the embodiments of the present disclosure lies in the following aspects: first, the quality of the input information source is significantly improved. The probabilistic interval prediction module receives the enhanced prediction result output by the physically consistent calibration module and the heterogeneous graph prediction module, rather than the original data or the point prediction result of a single prediction model, which lays a solid foundation for generating more accurate and reliable probabilistic interval prediction results. Second, the probabilistic interval prediction module not only generates a probability interval, but also can decompose the sources of uncertainty (e.g., distinguish the uncertainty caused by weather forecast errors, photovoltaic power prediction model errors, and equipment state changes), which provides a deeper decision basis for power grid dispatching. Third, the probabilistic interval prediction module provided in the embodiments of the present disclosure expands the intelligent alarm and risk assessment functions (e.g., high-uncertainty risk early warning, out-of-limit probability alarm, etc.) based on probabilistic prediction, and realizes a functional leap from providing probability information to supporting active risk management.

[0069] In some embodiments of the present disclosure, as Figure 1EAs shown, the probability interval prediction module can extract the basic features and uncertainty-related features, and then output the complete probability distribution of the photovoltaic power based on the basic features and uncertainty-related features, and optionally, the first sub-power prediction result output by the physical constraint prediction module and the second sub-power prediction result output by the heterogeneous graph prediction module, and by using a selected probability prediction method. Then, based on the complete probability distribution, the confidence interval of the complete probability distribution is calculated or quantile extraction is performed to decompose the uncertainty sources to obtain the uncertainty quantification result, and the uncertainty quantification result, the confidence interval calculation result, the quantile extraction result and the complete probability distribution are integrated to obtain the final uncertainty prediction result, and the prediction quality is evaluated to obtain the interval coverage rate, the interval width and the reliability index. In the specific embodiments of the present disclosure, the selected probability prediction method includes but is not limited to one of conditional variational autoencoder, adaptive kernel density estimation, quantile regression method. The uncertainty quantification result includes but is not limited to random uncertainty, cognitive uncertainty, or total uncertainty, etc. The uncertainty prediction result includes point prediction result, interval prediction result, complete probability distribution, uncertainty quantification result, etc.

[0070] The multi-stage rolling prediction controller is used to coordinate the application strategies and information interaction of the basic prediction module, the physical constraint prediction module, the heterogeneous graph prediction module and the probability interval prediction module in the three prediction stages of day-ahead prediction (such as 24 hours in the future), intra-day prediction (such as updating every 4 hours), and ultra-short-term rolling prediction (such as updating every 5 minutes), to realize the optimization of the prediction tasks in each stage. The embodiment of the present disclosure constructs a physical-temporal fusion prediction architecture based on multi-objective collaborative optimization. This is not simply a model stacking, but a multi-objective optimization framework is introduced in the training stage to optimize the prediction accuracy, physical consistency, interval coverage rate and other multiple objectives at the same time, so that each module evolves towards the global optimal target. In addition, an intelligent "scene perception and model selection" capability (which can be integrated in the controller or other modules) is designed, which can dynamically adjust the fusion strategy according to the real-time data characteristics, realizing intelligent decision and organic fusion.

[0071] The integration and decision module is configured to fuse the prediction results from the basic prediction module, the physical constraint prediction module, the heterogeneous graph prediction module, and the probability interval prediction module, and generate a final photovoltaic power prediction result according to a preset fusion strategy. The preset fusion strategy includes a weighted fusion strategy, a cascaded fusion strategy, and an adaptive selection fusion strategy. In the specific embodiments of the present disclosure, one of the weighted fusion strategy, the cascaded fusion strategy, and the adaptive selection fusion strategy can be selected to fuse the prediction results from the basic prediction module, the physical constraint prediction module, the heterogeneous graph prediction module, and the probability interval prediction module. There is an explicit information transmission and enhancement relationship between the modules. For example, the first sub-power prediction result output by the physical constraint prediction module can be used as a more reliable input of the heterogeneous graph prediction module to incorporate physical consistency, the second sub-power prediction result output by the heterogeneous graph prediction module to the probability interval prediction module is further incorporated with spatial and temporal correlation, and the final probability interval prediction module quantifies the uncertainty on this basis. This is a cascading process of layer-by-layer progression and information value-added, which realizes the synergistic effect of “1+1+1>3”.

[0072] The cloud microservice and EMS interface is configured to encapsulate the entire photovoltaic power prediction system as an extensible cloud microservice, and provide an enhanced Application Programming Interface (API), such as a point prediction API, a probability interval API, a physical interpretation API, a spatial correlation API, and an uncertainty API, to support the release of prediction results, such as point prediction results, probability interval results, physical interpretation information, spatial correlation patterns, and uncertainty information, to an EMS or other application system. The interface carries and outputs the unique value brought by the physical constraint prediction module, the heterogeneous graph prediction module, and the probability interval prediction module. It not only transmits data, but also outputs multi-dimensional information including physical interpretation information, spatial correlation patterns, probability prediction results, and uncertainty decomposition results, to support intelligent alarm and risk assessment functions. This enables the EMS to make more refined risk-aware scheduling decisions, and embodies the comprehensive innovation of the photovoltaic power prediction system provided by the embodiments of the present disclosure from the core algorithm to the application function.

[0073] The data flow and processing process of the photovoltaic power prediction system includes:

[0074] Data input: The photovoltaic power prediction system receives multi-source data from photovoltaic power plant monitoring systems, weather service providers, grid dispatching centers, and the like, including high-frequency (such as 30 seconds) photovoltaic power, voltage, current, and other electrical quantities, irradiance, temperature, humidity, wind speed, and other weather observation and forecast data, photovoltaic module types, installation angles, rated power, and other static parameters, geographic locations of each site, grid topology connection relationships, and time stamps, calendar information, and the like.

[0075] Data processing: Data first enters the data preprocessing and feature extraction module to obtain the first feature set. Then, depending on the prediction stage and configuration, it is sent to the basic prediction module, physical constraint prediction module, heterogeneous graph prediction module, and probability interval prediction module for parallel or serial computation. A multi-stage rolling controller coordinates the operation and data interaction of each module, and the integration and decision module integrates the prediction results output by each module to obtain the final prediction result.

[0076] Output Results: The photovoltaic power prediction system ultimately outputs multi-dimensional prediction results, including one or more of the following: point prediction results of photovoltaic power in the next H time steps (H varies according to the prediction stage), probability prediction intervals for each time step, complete probability distribution of the prediction results, physical parameter estimation results, site influence weights, spatial correlation patterns, and uncertainty decomposition results (random uncertainty, cognitive uncertainty, and total uncertainty).

[0077] The following is combined Figures 2-12 This document provides a detailed explanation of the specific implementation of the photovoltaic power prediction method provided in the embodiments of this disclosure.

[0078] Figure 2 This is a schematic flowchart illustrating a photovoltaic power prediction method provided in an embodiment of this disclosure. Figure 2 As shown, the method includes:

[0079] S101, Obtain the first feature set.

[0080] In some embodiments of this disclosure, the first feature set includes basic features, physically related features, spatially related features, and uncertainty-related features. The basic features include the electrical, meteorological, and calendar characteristics of the photovoltaic module; the physically related features include photovoltaic module parameters, installation information, and solar trajectory information; the spatially related features include photovoltaic site coordinates, distance relationships between photovoltaic sites, grid topology, and regional attributes; and the uncertainty-related features include weather forecast errors, historical forecast errors, data quality, and time-spread characteristics.

[0081] In some embodiments of this disclosure, photovoltaic module parameters may include module type, rated power, area, standard efficiency, temperature coefficient, etc., and installation information may include: installation tilt angle, orientation angle, geographical latitude, etc., and solar trajectory information may include theoretical solar trajectories calculated based on time and geographical location, such as declination angle, hour angle, incident angle, etc.

[0082] In some embodiments of this disclosure, the site coordinates may include the latitude and longitude, geodetic coordinates, etc. of each photovoltaic power station and meteorological station; the distance between sites may include the geographical distance between photovoltaic power stations; the power grid topology may include the connection relationship between photovoltaic power stations and power grid nodes, line parameters, etc.; and the regional attributes may include the climate type, altitude, etc. of the region to which the photovoltaic site belongs.

[0083] In some embodiments of this disclosure, weather forecast error may include the deviation between historical weather forecast values ​​and measured values; historical prediction error may include the historical prediction error of the prediction model itself; data quality indicators may include the missing rate of the original data, the proportion of outliers, etc.; and time feature extension may include more refined time coding, such as whether it is during the sunrise / sunset transition period, or whether it is during a special weather event (such as cloudy, rainy, etc.).

[0084] In some embodiments of this disclosure, the acquisition of the first feature set can be accomplished by a data preprocessing and feature extraction module. This module can not only perform conventional cleaning, filling, normalization, alignment and other operations, but also construct the features required for training and prediction of the physical constraint prediction module, heterogeneous graph prediction module and probability interval prediction module.

[0085] S102, determine the first power prediction result based on the basic characteristics.

[0086] In some embodiments of this disclosure, the first power prediction result includes the basic prediction result of photovoltaic power, which refers to the prediction result obtained based on the basic characteristics of the photovoltaic module itself, such as the electrical characteristics, meteorological characteristics and calendar characteristics of the photovoltaic module and the basic prediction module. It is a purely data-driven prediction result, and can also be regarded as a benchmark prediction result. For specific implementation, please refer to the existing solutions. The embodiments of this disclosure will not be described in detail.

[0087] S103. Determine the regulation power prediction result based on one or more of the physical correlation characteristics, spatial correlation characteristics, and uncertainty correlation characteristics.

[0088] In some embodiments of this disclosure, the power prediction result includes one or more of a first sub-power prediction result, a second sub-power prediction result, and a third sub-power prediction result. The first sub-power prediction result includes the physical constraint prediction result and the physical parameter estimation result of photovoltaic power. The first sub-power prediction result conforms to the physical laws corresponding to the physical correlation characteristics. The second sub-power prediction result includes the spatial correlation prediction result of photovoltaic power and the site influence weight. The second sub-power prediction result is used to describe the spatial correlation corresponding to the spatial correlation characteristics. The third sub-power prediction result includes the probability interval prediction result and the uncertainty decomposition result of photovoltaic power.

[0089] The following is combined Figures 3-11 The process of processing the first sub-power prediction results, the second sub-power prediction results, and the third sub-power prediction results is explained in detail.

[0090] In some embodiments of this disclosure, Figure 3 This is a schematic flowchart illustrating another photovoltaic power prediction method provided in an embodiment of this disclosure. (In conjunction with...) Figure 2 ,likeFigure 3 As shown, Figure 2 S103 in the process determines the regulation power prediction result based on one or more of the following: physical correlation characteristics, spatial correlation characteristics, and uncertainty correlation characteristics. This may include... Figure 3 The following steps are involved:

[0091] S1031, Based on the physical characteristics, the prediction result of the first sub-power is determined using the physical constraint prediction module.

[0092] In some embodiments of this disclosure, the physical constraint prediction module includes a neural network that has learned the physical knowledge of photovoltaic power generation, such as physical laws, through training. Specifically, the physical laws embodying photovoltaic power generation can be integrated into the photovoltaic power prediction model in the form of a system of physical equations. For example, the physical laws can be incorporated into the calculation of the hybrid loss function to train the photovoltaic power prediction model. The specific process is as follows:

[0093] First, we construct a system of physical equations.

[0094] Specifically, based on the principles of photovoltaic power generation, physical or semi-physical models can be established to describe the relationship between power output and environmental factors and equipment characteristics. Core formulas include: a solar irradiance model, a photovoltaic conversion model, and a system loss model.

[0095] In some embodiments of this disclosure, the solar irradiance model is used to calculate the theoretical tilt surface irradiance at a given time and location, taking into account factors such as geographical location, time, component installation angle, etc., and involving solar declination angle, hour angle, incident angle, etc.

[0096] It should be noted that, in order to further improve accuracy, atmospheric scattering, reflection and cloud cover can be taken into account in practical applications to obtain the actual irradiance.

[0097] In some embodiments of this disclosure, the photovoltaic conversion model is used to describe the relationship between irradiance, temperature and output power. The module temperature can be calculated first, and then the conversion efficiency can be calculated based on the module temperature and irradiance to obtain the theoretical power generation.

[0098] In some embodiments of this disclosure, a system loss model is used to take into account inverter efficiency and other system losses to obtain the final physical model predicted power.

[0099] Then, determine the loss function.

[0100] Physical residual calculation: Calculate the basic prediction results output by the basic prediction module. Or the prediction results after fusion Substitute the predicted result of the first sub-power into the physical equation (or its differential form) and calculate the degree of deviation from the physical law, i.e., the physical residual. .in, The predicted value can be used as the basis for the prediction result. Or the prediction results after fusion , For observed photovoltaic power, estimated physical parameters can be used. Calculated.

[0101] Hybrid loss function: The optimization objective consists of two parts: data loss. and physical loss Among them, data loss Forecast value of photovoltaic power Compared with observed values Deviations between them, such as MSE, physical loss Physical residuals Norms, such as the L2 norm.

[0102] ;

[0103] in, The total loss is [amount], and the physical loss is [amount]. Data loss is , and These are weighting coefficients used to balance data-driven and physics-driven approaches; the weights can be adjusted. and To balance data fit and physical consistency.

[0104] Next, a physical constraint prediction module is constructed.

[0105] For example, one or more parallel physical parameter estimation branches can be added to the existing MS-CNN-BiLSTM to estimate unknown or time-varying physical parameters in the photovoltaic power prediction model. The base network branches are still responsible for learning complex patterns from the data; that is, the network architecture of the physical constraint prediction module integrates the original MS-CNN-BiLSTM and physical constraints.

[0106] The prediction results output by the physical constraint prediction module (such as the first sub-power prediction result) can be an adaptive fusion result of the prediction results output by the basic time series branch (such as the first power prediction result) and the physical model prediction results based on the physical constraint branch output. In other words, the prediction results output by the physical constraint prediction module can also be regarded as the result after calibrating, adjusting or correcting the basic prediction results using the physical model prediction results, so that the prediction results output by the physical constraint prediction module conform to physical laws, thereby improving the physical consistency of photovoltaic power prediction results.

[0107] It should be noted that the aforementioned basic timing branch and physical constraint branch can be set up in parallel, such as setting them up as a basic prediction module and a physical constraint prediction module respectively, or they can be integrated into a single module, such as... Figure 1B The physical constraint prediction module shown in this disclosure is not limited to any particular embodiment. Examples are given below.

[0108] In some embodiments of this disclosure, the physical constraint prediction module may include only the physical constraint branch. Figure 4 This is a schematic flowchart illustrating yet another photovoltaic power prediction method provided in this disclosure. (In conjunction with...) Figure 3 ,like Figure 4 As shown, Figure 3 In step S1031, based on physical correlation characteristics, the physical constraint prediction module determines the prediction result of the first sub-power, which may include... Figure 4 The following steps are involved:

[0109] S10311, Based on the physical characteristics, the physical parameter estimation results are determined by using physical constraint branches.

[0110] S10312, determine the physical constraint prediction results based on physical correlation characteristics and physical parameter estimation results.

[0111] In some embodiments of this disclosure, when the basic prediction module and the physical constraint prediction module are set up in parallel, a theoretical prediction result (physical model prediction result) that conforms to the physical laws can be determined based on physical related characteristics and physical laws. This theoretical prediction result is used to perform physical consistency calibration on the basic prediction result obtained by the basic prediction module in subsequent processes.

[0112] In some embodiments of this disclosure, the physical constraint prediction module may include a basic temporal branch and a physical constraint branch. Figure 5 This is a schematic flowchart illustrating yet another photovoltaic power prediction method provided in this disclosure. (In conjunction with...) Figure 3 ,like Figure 5 As shown, Figure 3 In step S1031, based on physical correlation characteristics, the physical constraint prediction module determines the prediction result of the first sub-power, which may include... Figure 5 The following steps are involved:

[0113] S10313, input the physical-related features into the physical constraint prediction module to obtain the second power prediction result.

[0114] S10314, based on the first power prediction result and the second power prediction result, the first sub-power prediction result is obtained.

[0115] In some embodiments of this disclosure, a theoretical prediction result conforming to physical laws (second power prediction result, physical model prediction result) can be obtained first using the physically relevant features and physical constraint prediction module. This result is then fused with the first power prediction result output by the basic prediction module to obtain a first sub-power prediction result (physical constraint prediction result). Specifically, this first sub-power prediction result is the prediction result obtained after the physical model prediction result has undergone preliminary physical law calibration of the basic prediction result. Figure 6 As shown.

[0116] In some embodiments of this disclosure, Figure 6 This is a schematic flowchart illustrating yet another photovoltaic power prediction method provided in this disclosure. (In conjunction with...) Figure 5 ,like Figure 6 As shown, the physics constraint prediction module includes a physics constraint branch. Figure 5 In step S10313, the physically relevant features are input into the physical constraint prediction module to obtain the second power prediction result, which may include... Figure 6 The following steps are involved:

[0117] S10313A, based on physical correlation characteristics, uses physical constraint branches to determine the physical parameter estimation results.

[0118] S10313B, determine the prediction results of the physical model based on the physical correlation characteristics and the estimation results of physical parameters.

[0119] Accordingly, in S10314, based on the first power prediction result and the second power prediction result, the first sub-power prediction result is obtained, including:

[0120] S10314A, based on the first power prediction result and the physical model prediction result, obtains the physical constraint prediction result.

[0121] As can be seen, the calibration of the first power prediction result using the physical model prediction result can be completed within the physical constraint prediction module or in a subsequent process, and this disclosure embodiment does not limit it.

[0122] In some embodiments of this disclosure, Figure 7 This is a schematic flowchart illustrating yet another photovoltaic power prediction method provided in this disclosure. (In conjunction with...) Figure 2 ,like Figure 7 As shown, Figure 2 S103 in the process determines the regulation power prediction result based on one or more of the following: physical correlation characteristics, spatial correlation characteristics, and uncertainty correlation characteristics. This may include... Figure 7 The following steps are involved:

[0123] S1032, Based on spatial correlation characteristics, the heterogeneous graph prediction module is used to determine the prediction result of the second sub-power.

[0124] In some embodiments of this disclosure, the heterogeneous graph prediction module can be constructed based on a heterogeneous graph neural network. The heterogeneous graph neural network defines the spatial relationships between photovoltaic power plants in multiple regions and is used to mine and utilize the spatiotemporal correlation information between multiple sites. The specific process is as follows:

[0125] First, construct a heterogeneous graph. This heterogeneous graph can include:

[0126] Node (V): Includes different types of entities such as photovoltaic power plants, weather stations, and power grid nodes.

[0127] Edge (E): Defines the relationships between different types of nodes, such as spatial proximity edges, meteorological influence edges, and electrical connection edges. Among them, spatial proximity edges are used to describe the spatial proximity relationship between photovoltaic power stations (based on distance), meteorological influence edges are used to describe the influence relationship between meteorological stations and photovoltaic stations (based on distance or meteorological similarity), and electrical connection edges are used to describe the electrical connection relationship between photovoltaic stations and grid nodes.

[0128] Feature Matrix: Assigning features to nodes and edges constructs a node feature matrix containing the attributes of each node mentioned above, and an edge feature matrix containing the attributes of each edge mentioned above. Among them, node features include photovoltaic power station features, meteorological station features, power grid node features, such as historical power sequences, equipment parameters, and geographical location; edge features can include spatially adjacent edge features, meteorological influence edge features, and electrical connection edge features, such as distance, connection strength, and line impedance.

[0129] Heterogeneous graph neural networks can include: heterogeneous graph convolutional / attention layers, spatiotemporal joint modeling, and hierarchical prediction. The heterogeneous graph convolutional / attention layers employ graph convolutional network (GNN) layers adapted to heterogeneous graphs, such as variants of Relational Graph Convolutional Network (RGCN), Heterogeneous Graph Attention Network (HAN), and Graph Attention Network (GAT), to aggregate neighbor information. These layers can distinguish different node types and relation types, learning information propagation patterns under specific relations. Spatiotemporal joint modeling combines time series models, such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Temporal Convolutional Network (TCN), with GNN layers to form a spatiotemporal graph neural network, simultaneously capturing temporal dynamics and spatial dependencies. Hierarchical prediction can include designing a multi-layered GNN structure to progressively aggregate information from the bottom-level station nodes to the cluster, regional, and even grid levels, achieving multi-level collaborative prediction. The hierarchical prediction framework realizes a hierarchical prediction system from grid-level prediction to regional-level prediction to cluster-level prediction to station-level prediction, obtaining the final spatial correlation prediction results.

[0130] In addition, heterogeneous maps can be updated in real time based on meteorological changes, prediction errors, and data quality assessment results, and predictions can be made based on the updated heterogeneous maps to improve the real-time performance and accuracy of prediction results.

[0131] In some embodiments of this disclosure, Figure 8 This is a schematic flowchart illustrating yet another photovoltaic power prediction method provided in this disclosure. (In conjunction with...) Figure 7 ,like Figure 8 As shown, Figure 7 In S1032, based on spatial correlation characteristics, the heterogeneous graph prediction module determines the prediction result of the second sub-power, which may include... Figure 8 The following steps are involved:

[0132] S10321, based on meteorological change characteristics and prediction errors, obtain updated information on the spatial dependence between photovoltaic power stations, meteorological stations, and power grid nodes.

[0133] The prediction error is the deviation between the spatial correlation prediction results of photovoltaic power and the observed values.

[0134] S10322, Update the heterogeneous graph prediction module based on the updated information of the spatial dependency relationship between photovoltaic power plants, meteorological stations, and power grid nodes.

[0135] S10323, input the spatial correlation features into the updated heterogeneous graph prediction module to obtain the spatial correlation prediction results of photovoltaic power and the site influence weight.

[0136] In some embodiments of this disclosure, the dynamic graph update mechanism adjusts the graph structure based on real-time meteorological change characteristics and prediction errors. Specifically, this may include: obtaining update information for edges and / or nodes based on real-time meteorological change characteristics and prediction errors, and updating or adjusting the graph structure based on this update information, such as adding or deleting nodes, adding, deleting, or changing the connection relationships (edges) between nodes, and changing the weights of edges, etc. Specific implementations include... Figure 9 As shown.

[0137] In some embodiments of this disclosure, Figure 9 This is a schematic flowchart illustrating yet another photovoltaic power prediction method provided in this disclosure. (In conjunction with...) Figure 8 ,like Figure 9 As shown, Figure 8 In step S10322, the heterogeneous graph neural network is updated based on the updated information of the spatial dependencies between photovoltaic power plants, meteorological stations, and grid nodes. This can include... Figure 9 The following steps are involved:

[0138] S10322A updates one or more of the nodes, edges, and edge weights of the heterogeneous graph prediction module based on the updated information of the spatial dependency relationship between photovoltaic power plants, meteorological stations, and power grid nodes.

[0139] In some embodiments of this disclosure, Figure 10 This is a schematic flowchart illustrating yet another photovoltaic power prediction method provided in this disclosure. (In conjunction with...) Figure 2 ,like Figure 10 As shown, Figure 2 S103 in the process determines the regulation power prediction result based on one or more of the following: physical correlation characteristics, spatial correlation characteristics, and uncertainty correlation characteristics. This may include... Figure 10 The following steps are involved:

[0140] S1033, Based on the uncertainty-related characteristics, the probability interval prediction module is used to determine the prediction result of the third sub-power.

[0141] In some embodiments of this disclosure, the probability interval prediction module includes a probability interval prediction network, which includes one or more of a point prediction module, a conditional variational autoencoder module, an adaptive kernel density estimation module, and a quantile regression module, for determining the probability interval prediction result and uncertainty decomposition result of photovoltaic power based on uncertainty-related characteristics.

[0142] In some embodiments of this disclosure, Figure 11 This is a schematic flowchart illustrating yet another photovoltaic power prediction method provided in this disclosure. (In conjunction with...) Figure 2 ,like Figure 11 As shown, Figure 2 S103 in the process determines the regulation power prediction result based on one or more of the following: physical correlation characteristics, spatial correlation characteristics, and uncertainty correlation characteristics. This may include... Figure 11 The following steps are involved:

[0143] S1034, Based on the uncertainty-related characteristics, the first sub-power prediction result, and the second sub-power prediction result, the probability interval prediction module is used to determine the third sub-power prediction result.

[0144] In some embodiments of this disclosure, the probability interval prediction module can be constructed based on probability distribution modeling techniques. These techniques may include: methods based on model output, direct probability prediction methods, quantile regression, etc.

[0145] The specific implementation of the model output-based method involves assuming that the prediction error of the base point prediction model (such as the output of the PINN-PV module or HGNN-PV module) follows a certain distribution (such as a Gaussian distribution), and then estimating the parameters of this distribution (such as the mean and variance). The variance can be further decomposed into random uncertainty (inherent noise in the data) and cognitive uncertainty (model uncertainty). Cognitive uncertainty can be estimated using ensemble methods (such as Dropout as Bayesian Approximation, Deep Ensembles).

[0146] Direct probabilistic prediction methods can be implemented using one of the following: Conditional Variational Autoencoders (CVAs), Adaptive Kernel Density Estimation (RDE), or Quantile Regression. Specifically, CVAs learn a latent representation of the conditional probability distribution from input features to the output prediction result; RDE constructs a nonparametric conditional probability density function based on historical samples or model prediction residuals; and quantile regression directly trains the model to predict different quantile levels. The predicted values ​​are typically generated using pinball loss as the loss function. Then, the pre-constructed probability interval prediction module is used to generate the prediction interval, which includes:

[0147] Based on the obtained complete probability distribution or quantile prediction, construct a prediction interval. The interval width should be adaptively adjustable to reflect the current level of uncertainty in the prediction.

[0148] Next, uncertainty decomposition is performed. For example, by using a model output-based method, random uncertainty (usually the variance term is directly output by the model) and cognitive uncertainty (the variance of the predicted mean is obtained through model ensemble or Monte Carlo sampling) can be estimated separately.

[0149] It should be noted that the probability interval prediction module can be set up in parallel with the physical constraint prediction module and the heterogeneous graph prediction module, or it can be cascaded. For example, the first sub-power prediction result output by the physical constraint prediction module and the second sub-power prediction result output by the heterogeneous graph prediction module can be used as the input of the probability interval prediction module. The probability interval prediction module can determine the third sub-power prediction result based on the uncertainty-related characteristics, the first sub-power prediction result, and the second sub-power prediction result, as detailed below. Figure 11 As shown.

[0150] S104. Based on the first power prediction result and the adjusted power prediction result, determine one or more of the following: point prediction result of photovoltaic power, probability prediction interval, confidence interval, complete probability distribution, physical parameter estimation result, spatial correlation mode, and uncertainty decomposition result.

[0151] In some embodiments of this disclosure, the first power prediction result, the first sub-power prediction result, the second sub-power prediction result, and the third sub-power prediction result are fused according to one of the following methods: weighted, cascaded, or adaptive selection, to obtain the final prediction result. The following is in conjunction with... Figure 12 Please provide an explanation.

[0152] In some embodiments of this disclosure, Figure 12 This is a schematic flowchart illustrating yet another photovoltaic power prediction method provided in this disclosure. (In conjunction with...) Figure 2 ,like Figure 12 As shown, Figure 2 In step S104, based on the first power prediction result and the adjusted power prediction result, one or more of the following are determined: point prediction result of photovoltaic power, probability prediction interval, confidence interval, complete probability distribution, physical parameter estimation result, spatial correlation mode, and uncertainty decomposition result. These may include... Figure 12 The following steps are involved:

[0153] S1041, the first power prediction result, the first sub-power prediction result, the second sub-power prediction result and the third sub-power prediction result are fused according to one of the following methods: weighting, cascading, and adaptive selection, to obtain one or more of the following: photovoltaic power point prediction result, probability prediction interval, confidence interval, complete probability distribution, physical parameter estimation result, spatial correlation mode, and uncertainty decomposition result.

[0154] In some embodiments of this disclosure, a fusion strategy among weighted integration, cascaded fusion, and adaptive selection can be used to fuse the first power prediction result, the first sub-power prediction result, the second sub-power prediction result, and the third sub-power prediction result.

[0155] In some embodiments of this disclosure, weighted integration can dynamically adjust the weights of each module based on the accuracy of historical predictions to further improve prediction accuracy. For example, the weighting weights can be inversely proportional to the proportion of each module's historical prediction error in the total error.

[0156] In some embodiments of this disclosure, the modules can also be applied in a specific order. For example, the physical constraint prediction module can be used first to obtain a baseline prediction result that satisfies physical consistency, the heterogeneous graph prediction module can be used for spatial adjustment, and finally the probability interval prediction module can be used to generate the final prediction result.

[0157] In some embodiments of this disclosure, the most suitable module or fusion strategy can also be dynamically selected based on real-time evaluation indicators (such as data sufficiency, weather extremes, and forecast uncertainty level).

[0158] In some embodiments of this disclosure, intermediate results or feature representations can be shared between modules. For example, the physical parameter estimation results of the physical constraint prediction module can be used as node features of the heterogeneous graph prediction module, and the spatial aggregation features of the heterogeneous graph prediction module can be input to the probability interval prediction module.

[0159] In some embodiments of this disclosure, the physical constraint prediction module, heterogeneous graph prediction module, and probability interval prediction module can be integrated into the existing three-stage rolling prediction framework to accommodate both long-term and short-term prediction needs. Depending on the length of the prediction period, the prediction stages can include the following three types:

[0160] The current forecast phase (future 24 hours) includes inputs such as historical data, weather forecasts for the next 24 hours, and calendar features. The core modules include the physical constraint forecast module (which generates physical scenarios using forecasts) and the heterogeneous graph forecast module (which models large-scale spatial relationships). The outputs include the point forecast baseline curve for the next 24 hours, regional / group forecasts, and preliminary forecast intervals.

[0161] Intraday correction stage (e.g., update the remaining periods of the day every 4 hours): The input includes the measured data of the day, updated short-term weather forecasts, and day-ahead prediction errors. The core modules include a physical constraint prediction module (dynamically adjusting physical parameters), a heterogeneous graph prediction module (updating the graph structure and weights), and a probabilistic interval prediction module (correcting the complete probability distribution based on the measurements). The output includes the corrected point prediction and interval prediction for the remaining periods of the day.

[0162] Ultra-short-term rolling prediction stage (predict the next 1 hour every 5 minutes): The input includes the high-frequency measured data in the last 1 hour and the latest sensor readings. The core modules include a probabilistic interval prediction module (updating the probability prediction at high frequency), a simplified physical constraint prediction module (fast physical response), and a local heterogeneous graph prediction module (capturing short-term spatial propagation). The output includes high-resolution point prediction and interval prediction for the next 1 hour for real-time control.

[0163] It can be seen that in the above three prediction stages, the day-ahead prediction stage provides the benchmark prediction values for the intraday correction stage and the ultra-short-term rolling prediction stage. The intraday correction stage and the ultra-short-term rolling prediction stage are both generated after further correction or calibration directly or indirectly based on the prediction benchmarks provided by the day-ahead prediction stage.

[0164] In some embodiments of the present disclosure, a cloud microservice architecture can also be constructed. Each module (data processing, basic prediction module, physical constraint prediction module, heterogeneous graph prediction module, probabilistic interval prediction module, integrated decision-making, rolling control) is encapsulated as an independent and scalable microservice, and is managed and called through an API gateway to support distributed computing (especially the heterogeneous graph prediction module) and real-time stream processing. Among them, the cloud microservice architecture includes adding an EMS interface for EMS linkage interaction.

[0165] In some embodiments of the present disclosure, based on the original API, an EMS interface can be added for one or more of the following: obtaining probability prediction results (intervals at a specified confidence level), querying physical interpretation information (estimated key physical parameters), obtaining spatial association information (influence graph between stations), and obtaining uncertainty analysis results (decomposed uncertainty sources).

[0166] Correspondingly, based on the added EMS interface, one or more of the following linkages can be performed with the EMS: The EMS uses the enhanced prediction information for more optimized scheduling decisions, performs robust energy storage charging and discharging plans and grid trading plans based on the point prediction results and interval prediction results, adjusts the regional control strategy according to the physical interpretability information and spatial association patterns, uses the uncertainty information for risk assessment and reserve capacity optimization, and performs refined real-time power balance control based on the ultra-short-term probability prediction.

[0167] The photovoltaic power prediction method provided in this disclosure has at least the following advantages:

[0168] By constructing and integrating a first feature set encompassing four major categories of characteristics—fundamental, physical, spatial, and uncertain—this system systematically incorporates data-driven patterns, physical mechanism constraints, spatial correlation effects, and uncertainty information into a unified framework. This achieves deep fusion and collaborative prediction of multi-dimensional, multi-source information, fundamentally overcoming the limitations of single data-driven and single physical models. It provides a comprehensive information foundation for generating high-precision, high-reliability predictions. By introducing physical constraints on prediction results and explicitly requiring them to "conform to the physical laws corresponding to the relevant physical characteristics," it effectively avoids predictions that violate physical common sense (such as non-zero power output at night) that might occur with purely data-driven models. This not only ensures the physical rationality of the prediction results and improves their reliability and credibility, but also enhances the generalization ability and robustness of the photovoltaic power prediction model in scenarios not covered by training data (such as extreme weather and new types of components).

[0169] By generating spatial correlation prediction results and site influence weights, this method can explicitly model and quantify the mutual influence between various geographically dispersed sites (such as photovoltaic power plants, weather stations, and grid nodes), accurately characterizing the dynamic spatial correlation characteristics of regional photovoltaic clusters. It breaks through the traditional model of treating each photovoltaic power plant as an independent entity for prediction, realizing the prediction upgrade from "point" to "area", significantly improving the overall prediction accuracy of regional or cluster-level photovoltaic power, and is especially suitable for accurate prediction of short-term fluctuations caused by cloud movement, weather system changes, etc.

[0170] By outputting probability interval prediction results and uncertainty decomposition results, it has achieved a leap from "point prediction" to "probabilistic prediction". For the first time, it has provided complete quantitative information on prediction uncertainty in photovoltaic power prediction schemes. This allows users not only to obtain future power point estimates, but also to understand the possible fluctuation range (confidence interval) of the predicted values ​​and the sources of uncertainty (such as random errors or model cognitive errors). This provides crucial information support for risk-aware decisions such as power system risk assessment, reserve capacity configuration, and robust dispatch, and realizes a technological leap from deterministic prediction to probabilistic prediction.

[0171] The final output of physical parameter estimation results, spatial correlation patterns, uncertainty decomposition results, and other derived information together constitute a powerful model interpretation capability, forming an interpretable and traceable transparent prediction framework. This allows users to clearly understand "why the prediction results are as they are," which greatly enhances users' trust and acceptance of photovoltaic power prediction models and solves the pain point that photovoltaic power prediction models are often regarded as "black boxes."

[0172] In summary, based on the technical solution provided in this disclosure, through its unique technical architecture of "multi-source feature fusion → basic prediction and multi-path adjustment → result integration and output," it systematically solves the core problems existing in the prior art, such as physical inconsistency, neglect of spatial correlation, unquantifiable uncertainty, and uninterpretable prediction results. In particular, by fundamentally improving the basic time-series model (deep physical constraints, dynamic graph updates, uncertainty decomposition), constructing a system-level collaborative optimization architecture (multi-objective optimization, intelligent scene perception, information flow cascading enhancement), and providing an intelligent output interface that carries innovative value, it ultimately achieves a photovoltaic power prediction solution that significantly surpasses the level of existing technologies in terms of prediction accuracy, reliability, interpretability, and risk management support capabilities. The technical solution provided in this disclosure has significant improvements in core algorithms, system architecture, and application functions.

[0173] For example, the multi-dimensional prediction results (including point prediction results, probability interval results, physical parameter estimation results, site influence weights, uncertainty decomposition results, etc.) obtained based on the photovoltaic power prediction method provided in the embodiments of this disclosure can be further used for the collaborative optimization scheduling of energy storage systems to achieve efficient operation of photovoltaic-energy storage integrated systems.

[0174] In some embodiments of this disclosure, long-term energy storage planning can be optimized based on the physical constraint prediction results output by the physical constraint prediction module. That is, the physical constraint prediction results output by the physical constraint prediction module can provide a reliable basis for the long-term capacity planning of the energy storage system. By embedding the photovoltaic power generation physical equations, the long-term power prediction results provided by the physical constraint prediction module conform to physical laws and can be used to guide the energy capacity and power capacity configuration of the energy storage system. The energy storage capacity configuration optimization model aims to minimize total investment and operating costs, and determines the optimal energy storage configuration scheme under conditions such as load balance, state of charge (SOC) constraints, and power limitations.

[0175] Furthermore, the key physical parameters estimated by the physical constraint prediction module (such as photoelectric conversion efficiency, temperature coefficient, and irradiance response coefficient) reflect the actual operating characteristics of the photovoltaic system and can be used to formulate energy storage scheduling strategies based on physical parameters. For example, when the actual photoelectric conversion efficiency is higher than the reference value, the energy storage charging power can be appropriately increased; conversely, charging can be reduced, thereby achieving energy storage scheduling based on dynamic adjustments of physical parameters.

[0176] In some embodiments of this disclosure, regional energy storage coordination strategies can be optimized based on the spatial correlation prediction results output by the heterogeneous graph prediction module. Specifically, the heterogeneous graph prediction module learns the spatial correlations between photovoltaic power plants, weather stations, and grid nodes to generate site influence weights. These weights can be used to guide the coordinated operation of multiple energy storage systems within the region. The regional energy storage coordination optimization model aims to minimize the total grid power purchase cost and energy storage loss cost, achieving power coordination among energy storage systems while satisfying conditions such as load balance at each node and total grid power limitations.

[0177] Furthermore, the site impact weights output by the heterogeneous graph prediction module can be dynamically adjusted based on real-time meteorological conditions and grid status, and the energy storage coordination strategy is also updated accordingly. The dynamic coordination power calculation comprehensively considers the meteorological similarity distance between nodes, the electrical distance of the grid, and the difference between the SOC state of each energy storage system and the target SOC, thereby realizing intelligent coordination of regional energy storage.

[0178] In some embodiments of this disclosure, the real-time energy storage response can be optimized based on the probability interval prediction results output by the probability interval prediction module. Specifically, the complete probability distribution and confidence interval of the photovoltaic power prediction results provided by the probability interval prediction module can be used to guide the real-time response adjustment strategy of the energy storage system. When the actual photovoltaic power exceeds the upper and lower bounds of the prediction interval, the energy storage system performs power correction based on the emergency response coefficient to address the uncertainty caused by power fluctuations.

[0179] Furthermore, by utilizing the uncertainty decomposition results (random uncertainty and cognitive uncertainty) provided by the probability interval prediction module, the system can dynamically adjust the reserved capacity of the energy storage system. The adaptive adjustment of uncertainty weights takes into account the intensity of weather changes and model uncertainty indicators, enabling the energy storage system to better cope with uncertainties from different sources.

[0180] Furthermore, based on the complete probability distribution output by the probability interval prediction module, the system can also calculate the risks of power shortage and power surplus, and formulate risk-aware energy storage scheduling strategies accordingly. When the risk of power shortage is high, the energy storage system prepares to discharge; when the risk of power surplus is high, the energy storage system prepares to charge, thereby achieving risk-adaptive operation of the photovoltaic-energy storage system.

[0181] In summary, through the above three-level energy storage optimization strategies, this system achieves intelligent scheduling of multi-dimensional energy storage systems based on physical constraints, spatial correlation, and probabilistic prediction, significantly improving the operating efficiency and reliability of the photovoltaic-storage integrated system.

[0182] In some embodiments of this disclosure, Figure 13 This is a schematic diagram of another photovoltaic power prediction system provided in an embodiment of this disclosure. Figure 13 As shown, the system includes:

[0183] The acquisition module 1301 is used to acquire a first feature set, which includes basic features, physical related features, spatial related features, and uncertainty related features. The basic features include electrical features, meteorological features, and calendar features of photovoltaic modules. The physical related features include photovoltaic module parameters, installation information, and solar trajectory information. The spatial related features include photovoltaic site coordinates, distance relationships between photovoltaic sites, power grid topology, and regional attributes. The uncertainty related features include weather forecast errors, historical forecast errors, data quality, and time extension features.

[0184] Processing module 1302 is used to determine a first power prediction result based on basic characteristics, the first power prediction result including the basic prediction result of photovoltaic power;

[0185] The processing module 1302 is further configured to determine the regulation power prediction result based on one or more of the physical correlation characteristics, spatial correlation characteristics, and uncertainty correlation characteristics. The regulation power prediction result includes one or more of the first sub-power prediction result, the second sub-power prediction result, and the third sub-power prediction result. The first sub-power prediction result includes the physical constraint prediction result and the physical parameter estimation result of the photovoltaic power. The first sub-power prediction result conforms to the physical laws corresponding to the physical correlation characteristics. The second sub-power prediction result includes the spatial correlation prediction result of the photovoltaic power and the site influence weight. The second sub-power prediction result is used to describe the spatial correlation corresponding to the spatial correlation characteristics. The third sub-power prediction result includes the probability interval prediction result and the uncertainty decomposition result of the photovoltaic power.

[0186] The processing module 1302 is also used to determine one or more of the following based on the first power prediction result and the adjusted power prediction result: point prediction result of photovoltaic power, probability prediction interval, confidence interval, complete probability distribution, physical parameter estimation result, spatial correlation mode, and uncertainty decomposition result.

[0187] Figure 14 This is a schematic diagram of the structure of an electronic device provided in one embodiment of this disclosure. Figure 14 As shown, the electronic device includes: at least one processor 1401; and a memory 1402 communicatively connected to at least one processor 1401; wherein the memory 1402 stores instructions executable by at least one processor 1401, the instructions being executed by at least one processor 1401 to enable at least one processor 1401 to perform the method described in the above method embodiment.

[0188] The memory 1402 and processor 1401 can be connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors 1401 and memory 1402. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface can provide an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 1401 is transmitted over a wireless medium via an antenna, which further receives data and transmits it to processor 1401.

[0189] Processor 1401 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interface, voltage regulation, power management, and other control functions. Memory 1402 can be used to store data used by processor 1401 during operation.

[0190] In some embodiments, the electronic device may be an energy storage device, such as an energy storage cabinet. This energy storage device can be part of a larger energy storage system, or it can be a stand-alone energy storage device, such as a small household energy storage device. It can also be a sub-device or energy storage battery component or module in other devices, such as a power battery pack installed in an electric vehicle. In other embodiments, the electronic device may also be a charging chip or chip system installed within the aforementioned devices.

[0191] One embodiment of this disclosure provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the method described in the first aspect.

[0192] Those skilled in the art will understand that the above embodiments are specific examples of implementing this disclosure, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of this disclosure. Any person skilled in the art can make various alterations and modifications without departing from the spirit and scope of this disclosure; therefore, the scope of protection of this disclosure should be determined by the scope defined in the claims.

Claims

1. A photovoltaic power prediction method, characterized in that, The method includes: Obtain a first feature set, which includes basic features, physical-related features, spatial-related features, and uncertainty-related features; A first power prediction result is determined based on the aforementioned basic characteristics, and the first power prediction result includes the basic prediction result of photovoltaic power. The first sub-power prediction result is determined based on the physical correlation characteristics; the second sub-power prediction result is determined based on the spatial correlation characteristics; the third sub-power prediction result is determined using the probability interval prediction module based on the uncertainty correlation characteristics, the first sub-power prediction result, and the second sub-power prediction result; the third sub-power prediction result is used as the adjustment power prediction result. Based on the first power prediction result and the adjusted power prediction result, determine one or more of the following: point prediction result of photovoltaic power, probability prediction interval, confidence interval, complete probability distribution, and uncertainty decomposition result.

2. The method according to claim 1, characterized in that, The step of determining the first sub-power prediction result based on the physical correlation features includes: Based on the physical correlation characteristics, the first sub-power prediction result is determined using the physical constraint prediction module.

3. The method according to claim 2, characterized in that, The physical constraint prediction module includes a physical constraint branch. The step of determining the first sub-power prediction result using the physical constraint prediction module based on the physical correlation features includes: Based on the physical relevant characteristics, the physical constraint branches are used to determine the physical parameter estimation results; The physical constraint prediction result is determined based on the physical relevant characteristics and the physical parameter estimation results; The physical constraint prediction result is used as the first sub-power prediction result.

4. The method according to claim 2, characterized in that, The step of determining the first sub-power prediction result using the physical constraint prediction module based on the physical correlation characteristics includes: The physical correlation features are input into the physical constraint prediction module to obtain the second power prediction result; Based on the first power prediction result and the second power prediction result, the first sub-power prediction result is obtained.

5. The method according to claim 4, characterized in that, The physical constraint prediction module includes a physical constraint branch; the step of inputting the physically relevant features into the physical constraint prediction module to obtain the second power prediction result includes: Based on the physical relevant characteristics, the physical constraint branches are used to determine the physical parameter estimation results; The physical model prediction results are determined based on the physical relevant characteristics and the physical parameter estimation results. The prediction result of the physical model is used as the second power prediction result.

6. The method according to claim 1, characterized in that, The step of determining the second sub-power prediction result based on the spatial correlation features includes: Based on the spatial correlation characteristics, the heterogeneous graph prediction module is used to determine the prediction result of the second sub-power.

7. The method according to claim 6, characterized in that, The step of determining the second sub-power prediction result using the heterogeneous graph prediction module based on the spatial correlation features includes: Based on meteorological change characteristics and prediction errors, updated information on the spatial dependence between photovoltaic power plants, meteorological stations, and power grid nodes is obtained. The prediction error is the deviation between the spatial correlation prediction result of photovoltaic power and the observed value. The heterogeneous graph prediction module is updated based on the updated information of the spatial dependencies between photovoltaic power plants, weather stations, and power grid nodes. The spatial correlation features are input into the updated heterogeneous graph prediction module to obtain the spatial correlation prediction results of photovoltaic power and the site influence weight.

8. The method according to claim 7, characterized in that, The step of updating the heterogeneous graph prediction module based on the updated information of the spatial dependencies between the photovoltaic power station, meteorological station, and power grid nodes includes: Based on the updated information of the spatial dependencies between photovoltaic power plants, weather stations, and power grid nodes, update one or more of the nodes, edges, and edge weights of the heterogeneous graph prediction module.

9. The method according to claim 1, characterized in that, The probability interval prediction module includes one or more of the following: a point prediction module, a conditional variational autoencoder module, an adaptive kernel density estimation module, and a quantile regression module. It is used to determine the probability interval prediction result and uncertainty decomposition result of photovoltaic power based on the uncertainty-related characteristics.

10. The method according to any one of claims 1-9, characterized in that, The step of determining one or more of the following based on the first power prediction result and the adjusted power prediction result: point prediction result, probability prediction interval, confidence interval, complete probability distribution, and uncertainty decomposition result of photovoltaic power, including: The first power prediction result, the first sub-power prediction result, the second sub-power prediction result, and the third sub-power prediction result are fused according to one of the following methods: weighting, cascading, or adaptive selection, to obtain one or more of the following photovoltaic power prediction results: point prediction result, probability prediction interval, confidence interval, complete probability distribution, and uncertainty decomposition result.

11. A photovoltaic power prediction system, characterized in that, include: The acquisition module is used to acquire a first feature set, which includes basic features, physical related features, spatial related features, and uncertainty related features; The processing module is used to determine a first power prediction result based on the basic characteristics, wherein the first power prediction result includes the basic prediction result of photovoltaic power; The processing module is further configured to determine a first sub-power prediction result based on the physical correlation characteristics, determine a second sub-power prediction result based on the spatial correlation characteristics, determine a third sub-power prediction result using a probability interval prediction module based on the uncertainty correlation characteristics, the first sub-power prediction result, and the second sub-power prediction result, and use the third sub-power prediction result as an adjustment power prediction result. The processing module is further configured to determine one or more of the following based on the first power prediction result and the adjusted power prediction result: point prediction result, probability prediction interval, confidence interval, complete probability distribution, and uncertainty decomposition result of the photovoltaic power.

12. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method as described in any one of claims 1-10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the method of any one of claims 1-10.

Images