Hybrid power-oriented wind-solar combined power short-term prediction method and system

Abstract

The application discloses a wind-solar combined power short-term prediction method and system for a hybrid power supply, and relates to the technical field of new energy power prediction. The method comprises the following steps: obtaining historical operation data sets of wind-solar hybrid power supply in a target area and corresponding high-dimensional meteorological and topographic data sets; reconstructing the high-dimensional meteorological and topographic data sets in phase space, and fusing and constructing a high-dimensional phase space for describing the dynamics of regional net power; using a neural manifold learning algorithm to extract a low-dimensional manifold structure representing the dynamics of regional net power and a bidirectional mapping module from the high-dimensional phase space; predicting the future motion trajectory of phase points on the low-dimensional manifold structure to obtain a manifold trajectory prediction result in a future preset time period; and inverting to the high-dimensional physical space through the bidirectional mapping module to obtain a deterministic prediction curve of the regional net power. The technical problem of low wind-solar hybrid power supply power prediction accuracy in the prior art is solved, and the technical effect of improving the wind-solar combined power short-term prediction accuracy is achieved.

Description

Technical Field

[0001] This invention relates to the field of new energy power prediction technology, specifically to a method and system for short-term prediction of combined wind and solar power for hybrid power sources. Background Technology

[0002] With the continuous growth of installed capacity of new energy sources such as wind power and photovoltaics, wind-solar hybrid power sources are gradually becoming an important component of regional power systems. Due to the significant randomness, volatility, and intermittency of both wind and solar power output, their power variations are influenced by a combination of factors, including meteorological conditions, terrain features, and equipment operating status. Therefore, short-term power prediction for wind-solar hybrid power sources has become a key technical issue in the grid-connected operation, dispatch optimization, and safety control of new energy sources. Existing wind and solar power prediction methods are typically based on statistical or machine learning models, independently modeling wind and solar power outputs, or directly using multidimensional meteorological data and historical power data as high-dimensional inputs for regression prediction. However, these methods often struggle to effectively characterize the coupling relationship between wind and solar power outputs and the overall dynamic evolution characteristics of regional net power when facing complex and variable meteorological environments. Furthermore, direct modeling of high-dimensional meteorological and terrain features easily introduces redundant information, leading to difficulties in model training, insufficient generalization ability, and high sensitivity of prediction results to local meteorological disturbances. In addition, while some deep learning-based prediction methods have improved prediction accuracy to some extent, they usually rely on large-scale sample training and lack constraints on the intrinsic dynamic structure of power changes. This makes it difficult to guarantee the consistency and stability of prediction results in physical space, especially in short-term prediction scenarios with multiple time scales, where there are still problems with insufficient prediction accuracy and robustness. Summary of the Invention

[0003] This application provides a method and system for short-term prediction of wind and solar power combined for hybrid power sources, which solves the technical problem of low power prediction accuracy in the prior art due to the influence of multi-dimensional meteorological factors and complex nonlinear coupling relationship on the output of wind and solar hybrid power sources.

[0004] The first aspect of this application provides a short-term power forecasting method for combined wind and solar power generation for hybrid power sources, the method comprising: The process involves: acquiring historical operational datasets of wind-solar hybrid power sources within the target area and corresponding high-dimensional meteorological and topographic datasets; reconstructing the phase space of the high-dimensional meteorological and topographic datasets and fusing them to construct a high-dimensional phase space to describe the dynamics of the region's net power; using a neural manifold learning algorithm, extracting a low-dimensional manifold structure representing the dynamics of the region's net power from the high-dimensional phase space, along with a bidirectional mapping module between the low-dimensional manifold structure and the high-dimensional physical space; collecting real-time datasets, analyzing the phase point coordinates on the low-dimensional manifold structure, predicting the future trajectory of the phase points on the low-dimensional manifold structure, and obtaining the manifold trajectory prediction results within a preset future time period; and inverting the manifold trajectory prediction results to the high-dimensional physical space through the bidirectional mapping module to obtain a deterministic prediction curve for the region's net power.

[0005] A second aspect of this application provides a short-term power forecasting system for hybrid power sources combining wind and solar power, the system comprising: Data Acquisition Component: Acquires historical operational datasets of wind-solar hybrid power sources within the target area and corresponding high-dimensional meteorological and topographic datasets; Phase Space Reconstruction Component: Reconstructs the phase space of the high-dimensional meteorological and topographic datasets and fuses them to construct a high-dimensional phase space describing the dynamics of the region's net power; Manifold Structure Extraction Component: Utilizes a neural manifold learning algorithm, combined with the historical operational datasets, to extract a low-dimensional manifold structure representing the dynamics of the region's net power from the high-dimensional phase space, along with a bidirectional mapping module between the low-dimensional manifold structure and the high-dimensional physical space; Trajectory Prediction Component: Collects real-time datasets, analyzes the phase point coordinates on the low-dimensional manifold structure, predicts the future motion trajectory of the phase points on the low-dimensional manifold structure, and obtains the manifold trajectory prediction results within a preset future time period; Result Inversion Component: Inverts the manifold trajectory prediction results to the high-dimensional physical space through the bidirectional mapping module to obtain a deterministic prediction curve of the region's net power.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, historical operational datasets of the wind-solar hybrid power source within the target area and corresponding high-dimensional meteorological and topographic datasets are acquired. Next, the high-dimensional meteorological and topographic datasets are reconstructed into a phase space, and a high-dimensional phase space describing the dynamics of the region's net power is constructed. Further, using a neural manifold learning algorithm, a low-dimensional manifold structure representing the dynamics of the region's net power is extracted from the high-dimensional phase space, along with a bidirectional mapping module between the low-dimensional manifold structure and the high-dimensional physical space, based on the historical operational dataset. Then, real-time datasets are collected, and the phase point coordinates on the low-dimensional manifold structure are analyzed. The future trajectories of the phase points are predicted on the low-dimensional manifold structure, yielding manifold trajectory prediction results for a preset future time period. Finally, the manifold trajectory prediction results are inverted to the high-dimensional physical space through the bidirectional mapping module to obtain a deterministic prediction curve of the region's net power. This approach solves the technical problem of low power prediction accuracy in existing technologies due to the influence of multidimensional meteorological factors and complex nonlinear coupling relationships on the output of wind-solar hybrid power sources, achieving the technical effect of improving the short-term prediction accuracy of combined wind and solar power. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A schematic flowchart of a short-term power prediction method for hybrid power sources based on wind and solar power provided in an embodiment of this application; Figure 2 A schematic diagram of the structure of a short-term power prediction system for hybrid power sources provided in this application embodiment.

[0009] Figure labeling: Data acquisition component 11, phase space reconstruction component 12, manifold structure extraction component 13, trajectory prediction component 14, result inversion component 15. Detailed Implementation

[0010] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0011] Example 1, as Figure 1 As shown, this application provides a short-term power prediction method for combined wind and solar power for hybrid power sources, wherein the method includes: Obtain the historical operational dataset of the wind-solar hybrid power source within the target area and the corresponding high-dimensional meteorological and topographic dataset.

[0012] In this embodiment, a spatial boundary of a target area is determined, the target area covering at least one wind farm and at least one photovoltaic power station, and data on the wind-solar hybrid power sources within the area is collected at a uniform time scale. By accessing a power station monitoring system, energy management system, or dispatch data platform, operational data of wind turbines and photovoltaic arrays within the target area during historical operating cycles are collected. This historical operational data includes at least: active power output value, reactive power output value, unit operating status identifier, grid connection status identifier, and sampling timestamp; wherein, the historical operating cycle is a continuous multi-day or multi-month period, and the sampling time interval is set to a minute-level or higher time resolution according to dispatch requirements.

[0013] Simultaneously, for the target area, a high-dimensional meteorological and topographic dataset corresponding to the historical operating cycle is acquired. This high-dimensional meteorological and topographic dataset is obtained by connecting to a meteorological data service platform or a local meteorological observation system, and includes, but is not limited to, multi-dimensional meteorological elements such as wind speed, wind direction, air temperature, air pressure, humidity, solar irradiance, cloud cover, and precipitation. Furthermore, combined with the topographic and geographical information of the target area, parameter data characterizing the topographic features are acquired or constructed. These topographic features include at least altitude, slope, aspect, surface roughness, and distribution of land cover types.

[0014] The high-dimensional meteorological and topographic dataset is reconstructed into a phase space and then fused to construct a high-dimensional phase space for describing the dynamics of regional net power.

[0015] Phase space reconstruction processing is performed on high-dimensional meteorological and topographic datasets to construct a high-dimensional phase space capable of characterizing the dynamic evolution of regional net power. Specifically, for each type of meteorological or topographic feature variable in the high-dimensional meteorological and topographic dataset, a multivariate time-series vector is constructed in a unified time series format, where each time step corresponds to an observation state containing multidimensional meteorological elements and topographic parameters. Based on the multivariate time-series vector, a time-delay embedding method is used to reconstruct the phase space of the high-dimensional meteorological and topographic data. For any time step, a delay embedding vector is constructed based on the meteorological and topographic observation values ​​at the current time and several delay times before it, where the delay time interval and embedding dimension are preset according to the sampling frequency and the length of historical data or determined by empirical rules. Through the time-delay embedding processing, the original multidimensional time series is mapped to a set of state points located in a high-dimensional space, thereby forming a phase space point set characterizing the dynamic evolution of meteorology and topography.

[0016] Based on this, the regional net power data at the corresponding time step in the phase space point set and the historical operation dataset are fused. The regional net power is the superposition result of wind power and photovoltaic power within the target region. Specifically, the fusion method involves introducing the net power value or its derived features at the corresponding time step as an additional dimension into each phase space state point. This ensures that the reconstructed phase space state points simultaneously contain meteorological and topographical dynamic information and net power response information. Through this process, a high-dimensional phase space is constructed to describe the dynamics of regional net power. Each state point in this high-dimensional phase space corresponds to the comprehensive operating state of the target region at a certain time step, forming a dynamic trajectory in the phase space that reflects the evolution of the net power of the wind-solar hybrid power source as meteorological and topographical conditions change. This provides a unified state space foundation for subsequent low-dimensional manifold structure extraction and dynamic modeling.

[0017] Using a neural manifold learning algorithm, and combining the historical running dataset, a low-dimensional manifold structure representing the net power dynamics of the region is extracted from the high-dimensional phase space, along with a bidirectional mapping module between the low-dimensional manifold structure and the high-dimensional physical space.

[0018] After constructing a high-dimensional phase space to describe the dynamics of the region's net power, a low-dimensional manifold structure characterizing the dynamics of the region's net power is extracted based on the high-dimensional phase space state points and the corresponding historical operating datasets using a neural manifold learning algorithm. Specifically, the state points in the high-dimensional phase space are used as input samples to construct a manifold learning neural network. This network includes at least an encoder and a decoder, where the encoder maps the high-dimensional phase space state points to low-dimensional manifold coordinates, and the decoder reconstructs the high-dimensional phase space state from the low-dimensional manifold coordinates.

[0019] During training, historical datasets are used as supervision or constraint information to optimize the parameters of the manifold learning neural network. Training objectives include: minimizing the reconstruction error between the reconstructed state output by the decoder and the original high-dimensional phase space state, ensuring the low-dimensional manifold structure's ability to preserve the dynamic information of the region's net power; and simultaneously, imposing constraints on the continuity and smoothness of the low-dimensional manifold coordinates over time, enabling the low-dimensional manifold to reflect the actual dynamic evolution of the region's net power. Through this training, the encoder learns to form a stable mapping relationship from the high-dimensional phase space to the low-dimensional manifold structure.

[0020] After the manifold learning neural network training converges, the trained encoder and decoder are jointly encapsulated to form a bidirectional mapping module between the low-dimensional manifold structure and the high-dimensional physical space. The encoder constitutes a forward mapping from the high-dimensional phase space to the low-dimensional manifold structure, used to project real-time or historical states onto low-dimensional manifold coordinates; the decoder constitutes a reverse mapping from the low-dimensional manifold structure to the high-dimensional physical space, used to invert the prediction results on the low-dimensional manifold into the corresponding high-dimensional physical state, thereby realizing the dynamic bidirectional transformation of the region's net power between the low-dimensional manifold space and the high-dimensional physical space.

[0021] Furthermore, utilizing a neural manifold learning algorithm, and combining the historical running dataset, a low-dimensional manifold structure representing the net power dynamics of the region is extracted from the high-dimensional phase space, along with a bidirectional mapping module between the low-dimensional manifold structure and the high-dimensional physical space, comprising: Topological data analysis is performed on the historical time-series data corresponding to the high-dimensional phase space to identify potential dimensions; a manifold learning neural network containing an encoder and a decoder is constructed; based on the potential dimensions, the manifold learning neural network is trained using the historical operational dataset and the high-dimensional meteorological and topographic dataset, so that the encoder learns to map the high-dimensional phase space data to low-dimensional manifold coordinates, and the decoder learns to reconstruct the dynamic state of regional net power from the low-dimensional manifold coordinates; the manifold learning neural network trained to convergence is encapsulated to obtain the bidirectional mapping module.

[0022] Preferably, the state points arranged in time order in the high-dimensional phase space are regarded as multivariate time-series point cloud data. By performing topological data analysis on the point cloud data, the topological structure features presented under different embedding dimensions are extracted, thereby identifying potential dimensions that can stably characterize the dynamic characteristics of the net power of the region, which are used to limit the intrinsic dimension of subsequent manifold learning.

[0023] After identifying the latent dimension, a manifold learning neural network comprising an encoder and a decoder is constructed. The encoder maps the high-dimensional phase space state to the low-dimensional manifold coordinate space corresponding to the latent dimension, while the decoder reverses the mapping of the low-dimensional manifold coordinates and reconstructs the dynamic state of the region's net power in the high-dimensional phase space. The encoder and decoder are implemented using a multi-layer neural network structure, with their input and output dimensions matching the high-dimensional phase space dimension and the latent dimension, respectively.

[0024] Based on the aforementioned potential dimension, the manifold learning neural network is trained using the historical operational dataset and the high-dimensional meteorological and topographic dataset. During training, the high-dimensional phase space state points are used as encoder input, and the difference between the regional net power dynamic state reconstructed by the decoder and the corresponding state in the original high-dimensional phase space is used as the optimization objective. By iteratively updating the network parameters, the encoder gradually learns to form a mapping relationship from the high-dimensional phase space to the low-dimensional manifold coordinates, while the decoder learns the inverse mapping relationship from the low-dimensional manifold coordinates to reconstruct the regional net power dynamic state, until the network training reaches the preset convergence condition.

[0025] After training is completed and convergence is achieved, the trained encoder and decoder are encapsulated as a whole to form a bidirectional mapping module between the low-dimensional manifold structure and the high-dimensional physical space. The bidirectional mapping module is used to realize the mapping from the high-dimensional phase space state to the low-dimensional manifold coordinates in the subsequent prediction process, as well as the inversion from the low-dimensional manifold prediction result to the dynamic state of the net power of the region in the high-dimensional physical space.

[0026] Furthermore, topological data analysis is performed on the historical time-series data corresponding to the high-dimensional phase space to identify potential dimensions, including: Based on the multivariate time-series data collected in the high-dimensional phase space, an analog point cloud dataset in the high-dimensional phase space is constructed using a time-delay embedding method. A persistent graph representing topological features is generated from the analog point cloud dataset using a continuous cohomology calculation method. The topological features include at least the birth time, death time, and persistence of different dimensional topological structures. The topological features of the persistent graph are analyzed, and the cumulative persistence is calculated. The dimension corresponding to the previous topological structure when the cumulative persistence first decreases by a preset value is taken as the latent dimension.

[0027] Preferably, an analog point cloud dataset is constructed based on multivariate temporal state data arranged in chronological order in a high-dimensional phase space, using a time-delay embedding method. For each time step, the high-dimensional phase space states at the current time and several previous delayed times are combined to form an embedding vector, thereby obtaining a set of analog point cloud data samples for topology analysis in the high-dimensional phase space. After obtaining the analog point cloud dataset, a persistent cohomology calculation is performed on the analog point cloud dataset to generate a persistent graph for characterizing topological features. The persistent cohomology calculation tracks and analyzes the topological structures in different dimensions of the point cloud data by gradually increasing the neighborhood scale. The topological structures in different dimensions include at least connected components, holes, and high-dimensional cavities. The birth time, death time, and persistence represented by the difference between the birth time and death time are recorded in the persistent graph for each topological structure. Subsequently, the topological features in the persistent graph are analyzed, and the cumulative persistence of the topological structures under different topological dimensions is calculated. The cumulative persistence is the sum or weighted sum of the persistence of all topological structures under the same topological dimension. By comparing the trend of cumulative persistence with the change of topological dimension, it is determined that when the cumulative persistence first decreases by more than a preset value, the topological dimension of the previous topological structure is used as the potential dimension to characterize the intrinsic dimension of the region's net power dynamics.

[0028] Furthermore, after identifying potential dimensions, it also includes: Based on the potential dimensions, a candidate dimension range is constructed; for each candidate dimension within the candidate dimension range, the following steps are performed sequentially: Step a: A small sample set is constructed from the historical running dataset and the high-dimensional meteorological and topographic dataset according to a preset ratio, and the encoder and decoder are trained on a small scale using each candidate dimension as an intrinsic dimension to obtain the corresponding encoder-decoder model; Step b: Forward prediction is performed using the trained encoder-decoder model, and the root mean square error of the regional net power prediction result is calculated; Step c: The candidate dimension that minimizes the root mean square error is selected as the optimized potential dimension.

[0029] After identifying the potential dimensions through topological data analysis, to improve the accuracy of the low-dimensional manifold structure in representing the dynamics of regional net power, the method further includes an optimization and verification step for the potential dimensions. Specifically, a candidate dimension range is constructed centered on the potential dimensions, and the candidate dimension range is [potential dimension - 1, potential dimension + 1], which is used to cover possible dimensional deviations near the potential dimensions.

[0030] Within the candidate dimension range, each candidate dimension is selected sequentially, and the following steps are performed for each candidate dimension: First, a small sample set is constructed from the historical operational dataset and the high-dimensional meteorological and topographic dataset according to a preset ratio. This preset ratio limits the scale of data used for training, reducing computational overhead and avoiding repeated training on the full dataset. Using the current candidate dimension as the intrinsic dimension of the low-dimensional manifold, a small-scale training is performed on the manifold learning neural network composed of the encoder and decoder to obtain the encoder-decoder model corresponding to that candidate dimension. Subsequently, the trained encoder-decoder model is used to perform forward prediction processing, i.e., the encoder maps the high-dimensional phase space state to low-dimensional manifold coordinates, and the decoder inverts the low-dimensional manifold coordinates into a regional net power prediction result. The prediction result is compared with the actual regional net power data at the corresponding time step, and the root mean square error (RMSE) of the regional net power prediction result is calculated to quantify the predictive performance of the model under that candidate dimension. Finally, the RMSEs corresponding to each candidate dimension are compared and analyzed, and the candidate dimension that minimizes the RMSE is selected as the optimized potential dimension.

[0031] Furthermore, when training the manifold learning neural network, the training objectives include minimizing the reconstruction error while constraining the trajectory smoothness and dynamic predictability on the low-dimensional manifold.

[0032] When training the manifold learning neural network, in addition to minimizing the reconstruction error as the basic training objective, a joint constraint on the trajectory smoothness and dynamic predictability on the low-dimensional manifold is introduced to ensure that the learned low-dimensional manifold structure can not only accurately reconstruct the dynamic state of the net power of the region, but also has good temporal continuity and dynamic consistency.

[0033] Specifically, during training, the sequence of low-dimensional manifold coordinates output by the encoder is regarded as a state trajectory evolving over time. Constraints are placed on the magnitude of changes in the low-dimensional manifold coordinates between adjacent time steps. By limiting the first-order or higher-order time differences of the low-dimensional manifold trajectory, the trajectory changes on the low-dimensional manifold remain continuous and smooth, thus avoiding drastic jumps that do not conform to the actual power evolution law. Simultaneously, dynamic predictability constraints are introduced, incorporating the evolution relationship between the low-dimensional manifold coordinates at the current time step and subsequent time steps into the training objective. By utilizing the low-dimensional manifold coordinates within adjacent time steps or short time windows, the low-dimensional representation learned by the network is constrained to have a stable and predictable evolution trend over time, enabling a relatively accurate prediction of the low-dimensional manifold state at future moments based on the current low-dimensional manifold coordinates.

[0034] By jointly optimizing the goal of minimizing reconstruction error with constraints on the smoothness of low-dimensional manifold trajectory and the predictability of dynamics, the trained manifold learning neural network can form a structurally stable and temporally consistent low-dimensional manifold representation while maintaining the integrity of information in the high-dimensional phase space. This provides a reliable foundation for subsequent power evolution trajectory prediction in the low-dimensional manifold space.

[0035] Real-time datasets are collected, the coordinates of phase points on the low-dimensional manifold structure are analyzed, and the future motion trajectories of the phase points on the low-dimensional manifold structure are predicted to obtain the manifold trajectory prediction results within a preset time period.

[0036] After constructing the low-dimensional manifold structure and its bidirectional mapping module with the high-dimensional physical space, the real-time prediction stage begins. Specifically, a real-time dataset is collected, which includes at least real-time operational data of wind power and photovoltaic equipment within the target area, as well as meteorological and topographical data corresponding to the current moment. The real-time operational data includes real-time active power, equipment operating status, and timestamp information, while the meteorological and topographical data includes real-time or near-real-time meteorological elements such as wind speed, wind direction, irradiance, and temperature, along with corresponding topographical parameters.

[0037] After acquiring the real-time dataset, the real-time data is first time-aligned, scale-normalized, and feature-organized in a manner consistent with historical data processing. Then, the encoder in the bidirectional mapping module maps the real-time high-dimensional phase space state to a low-dimensional manifold structure, thereby obtaining the low-dimensional manifold phase point coordinates corresponding to the current moment. These low-dimensional manifold phase point coordinates characterize the comprehensive dynamic state of the target region under the current operating conditions. Subsequently, the future trajectory of the phase points is predicted on the low-dimensional manifold structure. Specifically, based on the low-dimensional manifold dynamics model obtained during the historical training phase, the current low-dimensional manifold phase point coordinates are used as the initial state input. The state evolution in the low-dimensional manifold space is advanced over time to deduce the low-dimensional manifold coordinate sequence corresponding to each time step within a preset future time period, thus forming the manifold trajectory prediction result. The preset time period is set as a short-term prediction window at the minute or hour level according to scheduling or prediction requirements.

[0038] Furthermore, real-time datasets are collected, the coordinates of phase points on the low-dimensional manifold structure are analyzed, and the future trajectories of the phase points on the low-dimensional manifold structure are predicted to obtain manifold trajectory prediction results within a preset future time period, including: A neural differential equation network is constructed based on the low-dimensional manifold structure. This network is used to learn the vector field on the manifold. Historical manifold coordinate time series based on the low-dimensional manifold structure are collected. The neural differential equation network is trained with the goal of predicting the coordinates of the next time step until the manifold simulation converges. The current manifold coordinates obtained by mapping the real-time dataset through the encoder in the bidirectional mapping module are used as initial conditions. The neural differential equation network is then used to deduce the manifold coordinates for a future preset time period, generating the manifold trajectory prediction result.

[0039] A neural differential equation network (NDE) is established based on the low-dimensional manifold structure to characterize the dynamic evolution of the manifold. This NDE network learns the vector field relationships in the low-dimensional manifold space to describe the continuous dynamic process of the manifold coordinates changing with time. Specifically, based on the low-dimensional manifold structure, a NDE network model is constructed. Its inputs are the low-dimensional manifold coordinates and their corresponding time variables, and its output is the rate of change of the low-dimensional manifold coordinates with respect to time, which represents the instantaneous evolution direction and velocity in the manifold space. In this way, the discrete-time evolution process on the low-dimensional manifold is modeled as a continuous-time differential equation.

[0040] During the training phase, historical manifold coordinate time series obtained based on the low-dimensional manifold structure are collected, and the neural differential equation network is trained with the prediction of the low-dimensional manifold coordinates for the next time step as the training objective. During training, the neural differential equations are time-progressed using numerical integration methods, and the error between the integration results and the actual historical manifold coordinates is calculated. The network parameters are then updated in reverse based on the error until the manifold dynamics simulation reaches the preset convergence condition.

[0041] In the real-time prediction phase, the real-time dataset is mapped to low-dimensional manifold coordinates corresponding to the current moment through the encoder in the bidirectional mapping module, and these low-dimensional manifold coordinates are used as initial conditions input into the trained neural differential equation network. By performing continuous-time integration on the neural differential equation, the sequence of low-dimensional manifold coordinates corresponding to each time point within a preset future time period is obtained, thereby generating the manifold trajectory prediction result.

[0042] Furthermore, the neural differential equation network takes manifold coordinates and time as input and outputs the derivative of the manifold coordinates with respect to time.

[0043] The neural differential equation network is used to characterize the continuous time dynamic relationship in a low-dimensional manifold space. The network input and output are defined as follows: The input of the neural differential equation network includes the low-dimensional manifold coordinates corresponding to the current time and the time variable. The low-dimensional manifold coordinates are used to characterize the comprehensive operating state of the target region at this time, and the time variable is used to identify the evolution position of the manifold coordinates on the time axis or the time step information.

[0044] The output of the neural differential equation network is the rate of change of the low-dimensional manifold coordinates with respect to time, i.e., the derivative of the manifold coordinates with respect to time, which represents the instantaneous evolution trend of the state along the time direction at the current manifold position. Through this output, the neural differential equation network can characterize the vector field distribution in the low-dimensional manifold space, thereby describing the continuous evolution behavior of the region's net power dynamics on the low-dimensional manifold.

[0045] In the actual prediction process, the time derivative of the manifold coordinates output by the neural differential equation network is used as the right-hand side of the differential equation. The differential equation is then advanced over time using numerical integration methods, thereby realizing the process of deriving the manifold coordinates at future moments from the current manifold coordinates.

[0046] The manifold trajectory prediction results are inverted to a high-dimensional physical space through the bidirectional mapping module to obtain a deterministic prediction curve of the regional net power.

[0047] After obtaining the low-dimensional manifold trajectory prediction results for a predetermined future time period, the low-dimensional manifold trajectory is inverted to a high-dimensional physical space through the bidirectional mapping module to generate a deterministic prediction result of the regional net power. Specifically, the low-dimensional manifold coordinates of each time step in the manifold trajectory prediction results are sequentially input into the decoder in the bidirectional mapping module. Using the inverse mapping relationship learned by the decoder, the low-dimensional manifold coordinates are reconstructed into the high-dimensional phase space state of the corresponding time step.

[0048] During the inversion process, the high-dimensional phase space state output by the decoder includes at least physical quantity features that are dynamically related to the net power of the region. By analyzing the dimension corresponding to the power output in the high-dimensional phase space state, the predicted net power of the region for each future time step is extracted. The predicted net power of the region is the combined power result of wind power and photovoltaic power in the target region, and it maintains a one-to-one correspondence with the low-dimensional manifold trajectory in the time dimension.

[0049] By arranging the predicted net power values ​​of the region obtained at each time step in chronological order, a deterministic prediction curve of the regional net power covering a preset future period is formed. This deterministic prediction curve characterizes the evolution trend of the regional wind-solar hybrid power net power over time under current operating conditions and meteorological conditions, and can be directly used as input for subsequent grid dispatching, power balance, and risk analysis.

[0050] Furthermore, it also includes: The system connects to a meteorological forecasting platform and collects a set of continuous time-series numerical weather forecasts to generate multiple future meteorological scenarios. It then uses the low-dimensional manifold structure and the bidirectional mapping module to obtain the net power prediction trajectories corresponding to each of the multiple future meteorological scenarios, generating a set of prediction trajectories. Continuous cohomology analysis is performed on the point cloud structure formed by the prediction trajectory set in high-dimensional phase space to extract topological evolution features over time, including connectivity, holes, and boundary curvature, resulting in a topological feature sequence. Based on the topological feature sequence, the system predicts the topological evolution of the net power feasible region in future time periods and calculates emerging risk indicators. When any emerging risk indicator exceeds a preset threshold, a risk warning signal for power grid dispatch is generated.

[0051] Building upon the deterministic prediction of regional net power, to assess potential power fluctuation risks during future operation, the process also includes net power prediction trajectory analysis and emergence risk warning based on multiple meteorological scenarios. Specifically, by connecting to a meteorological forecasting platform, a continuous time-series numerical weather forecast set covering a preset future period is collected. This numerical weather forecast set includes multiple meteorological forecast results with different initial disturbances or parameter configurations, used to generate multiple future meteorological scenarios. For each future meteorological scenario, the corresponding meteorological data is input into the low-dimensional manifold structure and the bidirectional mapping module. Following the aforementioned prediction process, the manifold trajectory under each scenario is deduced in the low-dimensional manifold space, and the manifold trajectory is inverted to a high-dimensional physical space to obtain the regional net power prediction trajectory corresponding to each future meteorological scenario, thus forming a set of net power prediction trajectories.

[0052] After obtaining the predicted trajectory set, the predicted trajectory set is represented in a high-dimensional phase space as a point cloud structure that evolves over time. Continuous cohomology analysis is performed on the point cloud structure to extract topological evolution features that reflect the dynamic evolution characteristics of net power. The topological evolution features include at least changes in the connectivity of the point cloud structure, the generation and disappearance of hole structures, and changes in boundary curvature. These topological evolution features are then organized in chronological order to form a topological feature sequence.

[0053] Based on topological feature sequences, the topological evolution trend of the feasible net power domain in a future time period is analyzed to identify potential emerging behaviors such as structural abrupt changes, splits, or contractions. One or more emerging risk indicators are then calculated to quantify potential instability risks during the dynamic evolution of net power. When any of these emerging risk indicators exceeds a preset threshold, a risk warning signal for power grid dispatch is generated to prompt the dispatch system to take proactive measures to adjust power balancing, reserve configuration, or operational strategies.

[0054] Furthermore, the emergence risk indicators include at least one of phase space trajectory divergence, topological mutation early warning index, and attractor steady-state switching probability.

[0055] Phase space trajectory divergence is used to characterize the degree of dispersion of the predicted trajectory of regional net power in high-dimensional phase space under different future meteorological scenarios. Specifically, within the same prediction time step, the distance statistics between phase space state points corresponding to different predicted trajectories are calculated. The distance statistics can be in the form of average distance, variance, or maximum distance. By analyzing the changing trend of the distance statistics over time, the phase space trajectory divergence is obtained, which is used to characterize the uncertainty and sensitivity of future power evolution.

[0056] The topological mutation warning index is used to characterize the risk of significant changes in the topology of the net power feasible domain. Specifically, based on the topological feature sequence obtained from continuous coherence analysis, the birth, death, and persistence changes of topological structures under different topological dimensions are monitored. When abrupt changes in the number of topological structures, significant changes in persistence, or rapid generation or disappearance of high-dimensional topological features occur within adjacent time windows, the corresponding topological mutation warning index is calculated to reflect the structural transformation risks that may occur during the dynamic evolution of net power.

[0057] The attractor steady-state switching probability is used to characterize the likelihood of a region's net power dynamically transitioning from one stable operating mode to another. Specifically, by clustering or performing stability analysis on the attractor regions that the predicted trajectory approaches in phase space, multiple potential steady-state attractors are identified. The frequency of different predicted trajectories transitioning from one attractor region to another within the prediction period is statistically analyzed. Based on this frequency, the attractor steady-state switching probability is calculated to characterize the risk level of system operating mode switching.

[0058] In summary, the embodiments of this application have at least the following technical effects: First, historical operational datasets of the wind-solar hybrid power source within the target area and corresponding high-dimensional meteorological and topographic datasets are acquired. Next, the high-dimensional meteorological and topographic datasets are reconstructed into a phase space, and a high-dimensional phase space describing the dynamics of the region's net power is constructed. Further, using a neural manifold learning algorithm, a low-dimensional manifold structure representing the dynamics of the region's net power is extracted from the high-dimensional phase space, along with a bidirectional mapping module between the low-dimensional manifold structure and the high-dimensional physical space, based on the historical operational dataset. Then, real-time datasets are collected, and the phase point coordinates on the low-dimensional manifold structure are analyzed. The future trajectories of the phase points are predicted on the low-dimensional manifold structure, yielding manifold trajectory prediction results for a preset future time period. Finally, the manifold trajectory prediction results are inverted to the high-dimensional physical space through the bidirectional mapping module to obtain a deterministic prediction curve of the region's net power. This approach solves the technical problem of low power prediction accuracy in existing technologies due to the influence of multidimensional meteorological factors and complex nonlinear coupling relationships on the output of wind-solar hybrid power sources, achieving the technical effect of improving the short-term prediction accuracy of combined wind and solar power.

[0059] Example 2, based on the same inventive concept as the short-term power prediction method for hybrid power sources in the foregoing examples, such as... Figure 2 As shown, this application provides a short-term power forecasting system for combined wind and solar power for hybrid power sources, wherein the system includes: Data Acquisition Component 11: Acquires historical operational datasets of wind-solar hybrid power sources within the target area and corresponding high-dimensional meteorological and topographic datasets; Phase Space Reconstruction Component 12: Reconstructs the phase space of the high-dimensional meteorological and topographic datasets and fuses them to construct a high-dimensional phase space for describing the dynamics of the region's net power; Manifold Structure Extraction Component 13: Utilizes a neural manifold learning algorithm, combined with the historical operational datasets, to extract a low-dimensional manifold structure representing the dynamics of the region's net power from the high-dimensional phase space, as well as a bidirectional mapping module between the low-dimensional manifold structure and the high-dimensional physical space; Trajectory Prediction Component 14: Collects real-time datasets, analyzes the phase point coordinates on the low-dimensional manifold structure, predicts the future trajectory of the phase points on the low-dimensional manifold structure, and obtains the manifold trajectory prediction results within a preset future time period; Result Inversion Component 15: Inverts the manifold trajectory prediction results to the high-dimensional physical space through the bidirectional mapping module to obtain a deterministic prediction curve of the region's net power.

[0060] Furthermore, the manifold structure extraction component 13 is used to perform the following method: Topological data analysis is performed on the historical time-series data corresponding to the high-dimensional phase space to identify potential dimensions; a manifold learning neural network containing an encoder and a decoder is constructed; based on the potential dimensions, the manifold learning neural network is trained using the historical operational dataset and the high-dimensional meteorological and topographic dataset, so that the encoder learns to map the high-dimensional phase space data to low-dimensional manifold coordinates, and the decoder learns to reconstruct the dynamic state of regional net power from the low-dimensional manifold coordinates; the manifold learning neural network trained to convergence is encapsulated to obtain the bidirectional mapping module.

[0061] Furthermore, the manifold structure extraction component 13 is used to perform the following method: Based on the multivariate time-series data collected in the high-dimensional phase space, an analog point cloud dataset in the high-dimensional phase space is constructed using a time-delay embedding method. A persistent graph representing topological features is generated from the analog point cloud dataset using a continuous cohomology calculation method. The topological features include at least the birth time, death time, and persistence of different dimensional topological structures. The topological features of the persistent graph are analyzed, and the cumulative persistence is calculated. The dimension corresponding to the previous topological structure when the cumulative persistence first decreases by a preset value is taken as the latent dimension.

[0062] Furthermore, the manifold structure extraction component 13 is used to perform the following method: Based on the potential dimensions, a candidate dimension range is constructed; for each candidate dimension within the candidate dimension range, the following steps are performed sequentially: Step a: A small sample set is constructed from the historical running dataset and the high-dimensional meteorological and topographic dataset according to a preset ratio, and the encoder and decoder are trained on a small scale using each candidate dimension as an intrinsic dimension to obtain the corresponding encoder-decoder model; Step b: Forward prediction is performed using the trained encoder-decoder model, and the root mean square error of the regional net power prediction result is calculated; Step c: The candidate dimension that minimizes the root mean square error is selected as the optimized potential dimension.

[0063] Furthermore, the manifold structure extraction component 13 is used to perform the following method: When training the manifold learning neural network, the training objectives include minimizing the reconstruction error while constraining the trajectory smoothness and dynamic predictability on the low-dimensional manifold.

[0064] Furthermore, the trajectory prediction component 14 is used to perform the following method: A neural differential equation network is constructed based on the low-dimensional manifold structure. This network is used to learn the vector field on the manifold. Historical manifold coordinate time series based on the low-dimensional manifold structure are collected. The neural differential equation network is trained with the goal of predicting the coordinates of the next time step until the manifold simulation converges. The current manifold coordinates obtained by mapping the real-time dataset through the encoder in the bidirectional mapping module are used as initial conditions. The neural differential equation network is then used to deduce the manifold coordinates for a future preset time period, generating the manifold trajectory prediction result.

[0065] Furthermore, the trajectory prediction component 14 is used to perform the following method: The neural differential equation network takes manifold coordinates and time as input and outputs the derivative of the manifold coordinates with respect to time.

[0066] Furthermore, the result inversion component 15 is used to perform the following method: The system connects to a meteorological forecasting platform and collects a set of continuous time-series numerical weather forecasts to generate multiple future meteorological scenarios. It then uses the low-dimensional manifold structure and the bidirectional mapping module to obtain the net power prediction trajectories corresponding to each of the multiple future meteorological scenarios, generating a set of prediction trajectories. Continuous cohomology analysis is performed on the point cloud structure formed by the prediction trajectory set in high-dimensional phase space to extract topological evolution features over time, including connectivity, holes, and boundary curvature, resulting in a topological feature sequence. Based on the topological feature sequence, the system predicts the topological evolution of the net power feasible region in future time periods and calculates emerging risk indicators. When any emerging risk indicator exceeds a preset threshold, a risk warning signal for power grid dispatch is generated.

[0067] Furthermore, the result inversion component 15 is used to perform the following method: The emergence risk indicators include at least one of phase space trajectory divergence, topological mutation early warning index, and attractor steady-state switching probability.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A short-term power forecasting method for combined wind and solar power generation for hybrid power sources, characterized in that, The method includes: Obtain the historical operational dataset of the wind-solar hybrid power source within the target area and the corresponding high-dimensional meteorological and topographic dataset; The high-dimensional meteorological and topographic dataset is reconstructed into a phase space and fused to construct a high-dimensional phase space for describing the dynamics of regional net power. Using a neural manifold learning algorithm, combined with the historical running dataset, a low-dimensional manifold structure representing the net power dynamics of the region is extracted from the high-dimensional phase space, along with a bidirectional mapping module between the low-dimensional manifold structure and the high-dimensional physical space. Collect real-time datasets, analyze the coordinates of phase points on the low-dimensional manifold structure, predict the future motion trajectory of the phase points on the low-dimensional manifold structure, and obtain the manifold trajectory prediction results within a preset time period in the future. The manifold trajectory prediction results are inverted to a high-dimensional physical space through the bidirectional mapping module to obtain a deterministic prediction curve of the regional net power.

2. The short-term power forecasting method for combined wind and solar power for hybrid power sources as described in claim 1, characterized in that, Using a neural manifold learning algorithm, and combining the historical running dataset, a low-dimensional manifold structure representing the net power dynamics of the region is extracted from the high-dimensional phase space. A bidirectional mapping module between the low-dimensional manifold structure and the high-dimensional physical space is also included. Topological data analysis is performed on the historical time series data corresponding to the high-dimensional phase space to identify potential dimensions; Construct a manifold learning neural network that includes an encoder and a decoder; Based on the potential dimension, the manifold learning neural network is trained using the historical operating dataset and the high-dimensional meteorological and topographic dataset, so that the encoder learns to map high-dimensional phase space data to low-dimensional manifold coordinates, and the decoder learns to reconstruct the dynamic state of regional net power from the low-dimensional manifold coordinates. The bidirectional mapping module is obtained by encapsulating the manifold learning neural network trained to convergence.

3. The short-term power forecasting method for combined wind and solar power for hybrid power sources as described in claim 2, characterized in that, Topological data analysis is performed on the historical time-series data corresponding to the high-dimensional phase space to identify potential dimensions, including: Based on the multivariate time-series data collected in the high-dimensional phase space, an analog point cloud dataset in the high-dimensional phase space is constructed using a time-delay embedding method. The analog point cloud dataset is used to generate a persistent graph representing topological features through a continuous cohomology calculation method. The topological features include at least the birth time, death time, and persistence of topological structures in different dimensions. Analyze the topological features of the persistent graph, calculate the cumulative persistence, and take the dimension of the previous topological structure when the cumulative persistence first decreases by a preset value as the potential dimension.

4. The short-term power forecasting method for combined wind and solar power for hybrid power sources as described in claim 2, characterized in that, After identifying potential dimensions, the following is also included: Construct a range of candidate dimensions based on the potential dimensions; For each candidate dimension selected from the range of candidate dimensions, the following steps are performed: Step a: Construct a small sample set in the historical running dataset and the high-dimensional meteorological and topographic dataset according to a preset ratio, and use each candidate dimension as the intrinsic dimension to train the encoder and decoder on a small scale to obtain the corresponding encoder-decoder model. Step b: Perform forward prediction using the trained encoder-decoder model and calculate the root mean square error of the region net power prediction result; Step c: Select the candidate dimension that minimizes the root mean square error as the optimized potential dimension.

5. The short-term power forecasting method for combined wind and solar power for hybrid power sources as described in claim 2, characterized in that, When training the manifold learning neural network, the training objectives include minimizing the reconstruction error while constraining the trajectory smoothness and dynamic predictability on the low-dimensional manifold.

6. The short-term power forecasting method for combined wind and solar power for hybrid power sources as described in claim 1, characterized in that, Real-time datasets are collected, the coordinates of phase points on the low-dimensional manifold structure are analyzed, and the future trajectories of the phase points on the low-dimensional manifold structure are predicted to obtain the manifold trajectory prediction results within a preset future time period, including: A neural differential equation network is constructed based on the low-dimensional manifold structure, and the neural differential equation network is used to learn vector fields on the manifold; Collect historical manifold coordinate time series based on the low-dimensional manifold structure, and train the neural differential equation network with the goal of predicting the coordinates of the next time step until the manifold simulation converges; Using the current manifold coordinates obtained by mapping the real-time dataset through the encoder in the bidirectional mapping module as initial conditions, the neural differential equation network is used to deduce the manifold coordinates for a future preset time period and generate the manifold trajectory prediction result.

7. The short-term power forecasting method for combined wind and solar power for hybrid power sources as described in claim 6, characterized in that, The neural differential equation network takes manifold coordinates and time as input and outputs the derivative of the manifold coordinates with respect to time.

8. The short-term power forecasting method for combined wind and solar power for hybrid power sources as described in claim 1, characterized in that, The method also includes: Connect to the meteorological forecasting platform to collect continuous time-series numerical weather forecasts and generate multiple future meteorological scenarios; The net power prediction trajectories corresponding to the multiple future meteorological scenarios are obtained through the low-dimensional manifold structure and the bidirectional mapping module, and a set of prediction trajectories is generated. A continuous cohomology analysis is performed on the point cloud structure formed by the predicted trajectory set in the high-dimensional phase space to extract the topological evolution features that evolve over time, including connectivity, holes and boundary curvature, to obtain a topological feature sequence. Based on the topological feature sequence, predict the topological evolution of the net power feasible region in the future time period and calculate the emergence risk index; When any of the emerging risk indicators exceeds a preset threshold, a risk warning signal for power grid dispatch is generated.

9. The short-term power forecasting method for combined wind and solar power for hybrid power sources as described in claim 8, characterized in that, The emergence risk indicators include at least one of phase space trajectory divergence, topological mutation early warning index, and attractor steady-state switching probability.

10. A short-term power forecasting system for combined wind and solar power sources, characterized in that, For implementing the short-term wind-solar combined power forecasting method for hybrid power sources as described in any one of claims 1-9, the system comprises: Data acquisition component: Acquires historical operational datasets of wind-solar hybrid power sources and corresponding high-dimensional meteorological and topographic datasets within the target area; Phase space reconstruction component: The high-dimensional meteorological and topographic dataset is reconstructed into a phase space and fused to construct a high-dimensional phase space for describing the regional net power dynamics; Manifold structure extraction component: Utilizing a neural manifold learning algorithm, combined with the historical running dataset, extracts a low-dimensional manifold structure representing the net power dynamics of the region from the high-dimensional phase space, as well as a bidirectional mapping module between the low-dimensional manifold structure and the high-dimensional physical space; Trajectory prediction component: Collects real-time datasets, analyzes the coordinates of phase points on the low-dimensional manifold structure, predicts the future motion trajectory of the phase points on the low-dimensional manifold structure, and obtains the manifold trajectory prediction results within a preset time period in the future; Result Inversion Component: The manifold trajectory prediction results are inverted to a high-dimensional physical space through the bidirectional mapping module to obtain a deterministic prediction curve of the regional net power.

Images