Distributed photovoltaic power station photovoltaic output prediction method and system based on multi-source heterogeneous characteristic data and power dispatching system

Abstract

The present disclosure relates to the technical field of power system dispatching, in particular to a distributed photovoltaic power station photovoltaic output prediction method and system based on multi-source heterogeneous feature data and a power dispatching system. The method comprises: performing spatio-temporal alignment processing on a multi-source heterogeneous feature data set, wherein the historical time period is divided into a plurality of time periods to be interpolated according to the solar elevation angle change rate and the data type of the multi-dimensional feature sequence, so as to select a matched interpolation mode to resample the multi-dimensional feature sequence with a sampling frequency lower than a preset reference frequency; inputting the multi-dimensional feature sequence after spatio-temporal alignment into a corresponding encoder for feature extraction, obtaining a multi-modal feature set through dimension reduction, mapping and splicing, inputting a joint feature vector obtained through feature fusion using a self-attention mechanism into a machine learning model to train a prediction model. The present disclosure can accurately interpolate the nonlinear change characteristics of photovoltaic meteorological data, and improve the prediction accuracy and robustness.

Description

Technical Field

[0001] This disclosure relates to the field of power system dispatching technology, specifically to a method and system for predicting photovoltaic power output of distributed photovoltaic power plants based on multi-source heterogeneous characteristic data, as well as a power dispatching system. Background Technology

[0002] Currently, with the continuous advancement of the "dual carbon" target, distributed photovoltaic (PV) power generation is increasingly accounting for a significant portion of the new power system, becoming an important component of the power system. Photovoltaic output refers to the active power actually output by photovoltaic (PV) power generation equipment to the grid or load at a specific moment. Due to the strong randomness, volatility, and intermittency of PV output, accurate prediction is crucial for ensuring the safe and stable operation of the power grid and optimizing dispatch decisions.

[0003] The accuracy of photovoltaic (PV) power output forecasting largely depends on the quality and completeness of the input data; a single data source is often insufficient to support accurate PV power output forecasting. In recent years, multi-source data fusion technology has become a key approach to improving PV forecasting accuracy. Research shows that multi-source data fusion can effectively improve forecast accuracy, but the data fusion process faces certain difficulties, the core challenge being data heterogeneity. Different data sources exhibit significant differences in sampling frequency, spatial resolution, data format, units of measurement, and spatiotemporal coverage, making it difficult for traditional simple splicing or linear interpolation methods to achieve effective data fusion, and potentially even introducing additional noise and bias.

[0004] Therefore, how to achieve accurate spatiotemporal fusion of multi-source heterogeneous data, avoid information loss and noise introduction, and thus improve the accuracy of photovoltaic power output prediction is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] To address the problems in related technologies, this disclosure provides a method and system for predicting photovoltaic power output of a distributed photovoltaic power station based on multi-source heterogeneous feature data, as well as a power dispatching system.

[0006] In a first aspect, this disclosure provides a method for predicting the photovoltaic output of a distributed photovoltaic power station based on multi-source heterogeneous feature data, including:

[0007] Obtain a multi-source heterogeneous feature dataset of the target distributed photovoltaic power station collected over a historical period. The multi-source heterogeneous feature dataset includes multiple multi-dimensional feature sequences from independent data sources. Each multi-dimensional feature sequence includes one multi-dimensional feature data or multiple multi-dimensional feature data arranged in chronological order. The multi-source heterogeneous feature dataset is subjected to time alignment and spatial alignment processing to obtain a spatiotemporally aligned multi-source heterogeneous feature dataset. Specifically, when the sampling frequency of the multi-dimensional feature sequence is lower than a preset reference frequency, the historical time period is divided into multiple interpolation time periods based on the rate of change of the solar altitude angle and the data type of the multi-dimensional feature sequence. Based on the data type of the multi-dimensional feature data within each interpolation time period and its corresponding interpolation time period, a matching interpolation method is selected for resampling to align the corresponding multi-dimensional feature data to the reference time axis determined by the reference frequency. For each moment on the reference time axis, all multidimensional feature sequences in the spatially aligned multi-source heterogeneous feature dataset are input into the corresponding encoder to extract initial features; each initial feature is subjected to dimensionality reduction and feature mapping to obtain a feature vector with a specified dimension; all feature vectors with the specified dimension are concatenated to form a multimodal heterogeneous feature set; the multimodal heterogeneous feature set is processed using a self-attention mechanism to obtain attention output features; the multimodal heterogeneous feature set and the attention output features are subjected to feature fusion processing to obtain a joint feature vector with the specified dimension at the corresponding moment. The joint feature vector at each moment on the reference time axis is input into the machine learning model for training. The parameters of the machine learning model are adjusted according to the difference between the predicted power generation and the actual power generation at each moment until the training stopping condition is met, thus obtaining the photovoltaic power output prediction model. The photovoltaic power output prediction model is then used to predict the photovoltaic power output of the target distributed photovoltaic power station.

[0008] According to embodiments of this disclosure, processing the multimodal heterogeneous feature set using a self-attention mechanism includes: The correlation weight matrix between each feature vector in the multimodal heterogeneous feature set is calculated using a self-attention mechanism; The multimodal heterogeneous feature set is updated by weighting based on the association weight matrix to obtain the attention output feature.

[0009] According to embodiments of this disclosure, the feature fusion process of the multimodal heterogeneous feature set and the attention output feature includes: The multimodal heterogeneous feature set and the attention output feature are normalized after residual connection to obtain fused intermediate features; the fused intermediate features are input into a multilayer perceptron for mapping, and then the mean aggregation of each feature vector after mapping is performed to obtain the joint feature vector with the specified dimension at the corresponding time.

[0010] According to embodiments of this disclosure, the sampling frequency of the multidimensional feature sequence is divided into high frequency, medium frequency, or low frequency; Wherein, the high frequency, the medium frequency and the low frequency correspond to the first set sampling frequency range, the second set sampling frequency range and the third set sampling frequency range, respectively, and the lower limit of the first set sampling frequency range is greater than the upper limit of the second set sampling frequency range, and the lower limit of the second set sampling frequency range is greater than the upper limit of the third set sampling frequency range. The reference frequency is a preset frequency value within the first set sampling frequency range, used to establish the reference time axis.

[0011] According to embodiments of this disclosure: When the multidimensional feature sequence includes multiple multidimensional feature data, each multidimensional feature data is a multivariable temporal feature data of a single spatial location under the spatial resolution of the corresponding data source. When the multidimensional feature sequence includes a multidimensional feature data, the multidimensional feature data is multivariate static feature data of a single spatial location at the spatial resolution of the corresponding data source; The method further includes: before performing the time alignment processing, decoupling the variable dimensions of each multidimensional feature sequence to obtain multiple corresponding univariate feature sequences; wherein, when the multidimensional feature sequence includes multiple multidimensional feature data, the univariate feature sequence includes multiple univariate time-series feature data arranged in chronological order; when the multidimensional feature sequence includes one multidimensional feature data, the univariate feature sequence includes one univariate static feature data; the univariate feature sequence is provided with corresponding position coordinates and variable identifiers.

[0012] According to embodiments of this disclosure, time alignment processing is performed on the multi-source heterogeneous feature dataset, including: When the sampling frequency of the multidimensional feature sequence is lower than the reference frequency, interpolation processing is performed on the multiple univariate feature sequences corresponding to the multidimensional feature sequence to align the univariate feature sequence to the reference time axis. When the sampling frequency of the multidimensional feature sequence is higher than the reference frequency, the multiple univariate feature sequences corresponding to the multidimensional feature sequence are downsampled respectively to align the univariate feature sequences to the reference time axis.

[0013] According to embodiments of this disclosure, the interpolation processing of the multiple univariate feature sequences corresponding to the multidimensional feature sequence includes: When the univariate feature sequence includes multiple univariate time series feature data arranged in chronological order, a matching interpolation method is selected for resampling based on the data type of the univariate time series feature data in each interpolation time period and the interpolation time period to which it belongs. When the univariate feature sequence includes a single univariate static feature data, the single variable static feature data is copied along the reference time axis to generate a time-series feature sequence covering each moment of the reference time axis. According to embodiments of this disclosure, the univariate time-series feature data includes any one of the following data types: radiation quantity data, environmental state quantity data, cumulative data, and vector data; The step of selecting a matching interpolation method for resampling based on the data type and interpolation time period of the univariate time series feature data within each interpolation time period includes: When the time period to be interpolated is a nighttime period, for the univariate time series feature data within the nighttime period, cubic spline interpolation is performed when the data type of the univariate time series feature data is environmental state data, and forced zeroing is performed when the data type of the univariate time series feature data is radiation data. When the time period to be interpolated is a stable daytime period, cubic spline interpolation is performed on the univariate time series feature data within the stable daytime period when the data type of the univariate time series feature data is radiation data or environmental state data. When the data type of the univariate time series feature data is cumulative, monotonic interpolation processing is performed. When the data type of the univariate time series feature data is vector data, component-independent cubic spline interpolation is performed.

[0014] According to embodiments of this disclosure, the method further includes: When missing data is detected in the univariate feature sequence, for each missing feature data in the univariate feature sequence, the following is executed: Multiple candidate auxiliary data sources are obtained, and the comprehensive spatiotemporal correlation coefficient between each candidate auxiliary data source and the missing feature data is calculated; a threshold is set based on the comprehensive spatiotemporal correlation coefficient. The multiple candidate auxiliary data sources are filtered according to the threshold to obtain one or more reliable auxiliary sources; The observation data of each trusted auxiliary source at the corresponding target time are obtained, and the observation data are converted into estimated values ​​with the same dimensions as the missing feature data through the corresponding physical inversion model, wherein the target time is the sampling time of the missing feature data; The corresponding estimates are weighted and fused according to the normalized weights of each reliable auxiliary source to generate the imputation values ​​for the missing feature data.

[0015] According to embodiments of this disclosure, the method further includes: After completing the time alignment process, multiple time-aligned single-variable time series feature sequences belonging to the same data source are recombined by spatial and variable mapping according to the corresponding position coordinates and variable identifiers to obtain time-aligned multidimensional feature sequences. The time-aligned multidimensional feature sequence is aligned with the reference time axis in the time dimension, and the spatial and variable dimensions are consistent with the multidimensional feature sequence before time processing.

[0016] According to embodiments of this disclosure, the method further includes: After completing the time alignment process, the time-aligned multidimensional feature sequence is subjected to spatial alignment processing, which includes: A unified fine grid is established for the target area, wherein the spatial resolution of the unified fine grid is higher than the spatial resolution of the data source corresponding to the multidimensional feature sequence, and the target area is the preset area range of the target distributed photovoltaic power station; For each moment on the reference time axis, the multidimensional feature sequence corresponding to each moment is mapped onto the unified fine grid to obtain the background value of each fine grid point; Obtain high-resolution digital elevation model grid data and land cover type grid data corresponding to each fine grid point in the unified fine grid; correct the background value of the corresponding grid point according to the digital elevation model grid data and the land cover type grid data to obtain a spatially aligned multidimensional feature sequence.

[0017] According to embodiments of this disclosure, the step of inputting all multidimensional feature sequences from the spatially aligned multi-source heterogeneous feature dataset into the corresponding encoder to extract initial features for each moment on the reference time axis includes: When the data structure of the multidimensional feature sequence is time-series data, a one-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is image data, a two-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is grid data, a three-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is scalar data, a fully connected encoder is used to extract the corresponding initial features.

[0018] Secondly, this disclosure provides a distributed photovoltaic power output prediction system based on multi-source heterogeneous feature data, comprising: The data input module is configured to acquire a multi-source heterogeneous feature dataset collected from the target distributed photovoltaic power station over a historical period. The multi-source heterogeneous feature dataset includes multiple multi-dimensional feature sequences from independent data sources. Each multi-dimensional feature sequence includes one multi-dimensional feature data or multiple multi-dimensional feature data arranged in chronological order. The spatiotemporal alignment module is configured to perform temporal and spatial alignment processing on the multi-source heterogeneous feature dataset to obtain a spatiotemporally aligned multi-source heterogeneous feature dataset. Specifically, when the sampling frequency of the multidimensional feature sequence is lower than a preset reference frequency, the historical time period is divided into multiple interpolation time periods based on the solar altitude angle change rate and the data type of the multidimensional feature sequence. Based on the data type of the multidimensional feature data within each interpolation time period and its corresponding interpolation time period, a matching interpolation method is selected for resampling to align the corresponding multidimensional feature data to a reference time axis determined by the reference frequency. The feature fusion module is configured to, for each moment on the reference time axis, input all multidimensional feature sequences from the spatially aligned multi-source heterogeneous feature dataset into the corresponding encoder to extract initial features; perform dimensionality reduction and feature mapping on each initial feature to obtain a feature vector with a specified dimension; concatenate all feature vectors with the specified dimension into a multimodal heterogeneous feature set; process the multimodal heterogeneous feature set using a self-attention mechanism to obtain attention output features; and perform feature fusion processing on the multimodal heterogeneous feature set and the attention output features to obtain a joint feature vector with the specified dimension at the corresponding moment. The photovoltaic power output prediction module is configured to input the joint feature vector of each moment on the reference time axis into the machine learning model for training, adjust the parameters of the machine learning model according to the difference between the predicted power generation and the actual power generation at each moment until the training stopping condition is met, and obtain the photovoltaic power output prediction model; and use the photovoltaic power output prediction model to predict the photovoltaic power output of the target distributed photovoltaic power station.

[0019] According to embodiments of this disclosure, when the multidimensional feature sequence includes multiple multidimensional feature data, each multidimensional feature data is a multivariate temporal feature data of a single spatial location at the spatial resolution of the corresponding data source; When the multidimensional feature sequence includes a multidimensional feature data, the multidimensional feature data is multivariate static feature data of a single spatial location at the spatial resolution of the corresponding data source; The system further includes a data preprocessing module configured to: decouple the variable dimensions of each multidimensional feature sequence before performing the time alignment processing, to obtain multiple corresponding univariate feature sequences; wherein, when the multidimensional feature sequence includes multiple multidimensional feature data, the univariate feature sequence includes multiple univariate time-series feature data arranged in chronological order; when the multidimensional feature sequence includes one multidimensional feature data, the univariate feature sequence includes one univariate static feature data; the univariate feature sequence is provided with corresponding position coordinates and variable identifiers.

[0020] According to embodiments of this disclosure, the spatiotemporal alignment module includes: a time dimension alignment submodule; the time dimension alignment submodule is configured to: When the sampling frequency of the multidimensional feature sequence is lower than the reference frequency, interpolation processing is performed on the multiple univariate feature sequences corresponding to the multidimensional feature sequence to align the univariate feature sequence to the reference time axis. When the sampling frequency of the multidimensional feature sequence is higher than the reference frequency, the multiple univariate feature sequences corresponding to the multidimensional feature sequence are downsampled respectively to align the univariate feature sequences to the reference time axis.

[0021] According to embodiments of this disclosure, the system further includes: a data reconstruction module; the data reconstruction module is configured to: After completing the time alignment process, multiple time-aligned single-variable time series feature sequences belonging to the same data source are recombined by spatial and variable mapping according to the corresponding position coordinates and variable identifiers to obtain time-aligned multidimensional feature sequences. The time-aligned multidimensional feature sequence is aligned with the reference time axis in the time dimension, and the spatial and variable dimensions are consistent with the multidimensional feature sequence before time processing.

[0022] According to embodiments of this disclosure, the spatiotemporal alignment module further includes a spatial dimension alignment submodule; the spatial dimension alignment submodule is configured to: A unified fine grid is established for the target area, wherein the spatial resolution of the unified fine grid is higher than the spatial resolution of the data source corresponding to the multidimensional feature sequence, and the target area is the preset area range of the target distributed photovoltaic power station; For each moment on the reference time axis, the multidimensional feature sequence corresponding to each moment is mapped onto the unified fine grid to obtain the background value of each fine grid point; Obtain high-resolution digital elevation model grid data and land cover type grid data corresponding to each fine grid point in the unified fine grid; correct the background value of the corresponding grid point according to the digital elevation model grid data and the land cover type grid data to obtain a spatially aligned multidimensional feature sequence.

[0023] According to embodiments of this disclosure, the feature fusion module includes a one-dimensional convolutional encoder, a two-dimensional convolutional encoder, a three-dimensional convolutional encoder, and a fully connected encoder; the feature fusion module is further configured to: When the data structure of the multidimensional feature sequence is time-series data, a one-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is image data, a two-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is grid data, a three-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is scalar data, a fully connected encoder is used to extract the corresponding initial features.

[0024] Thirdly, embodiments of this disclosure provide a power dispatching system, including a distributed photovoltaic output prediction system as described in any one of the second aspects; or including a memory and a processor, wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the prediction method as described in any one of the first aspects.

[0025] Fourthly, this disclosure provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the prediction method as described in any one of the first aspects.

[0026] According to the technical solution provided in this disclosure, a spatiotemporal alignment processing framework for multi-source heterogeneous feature data is constructed to achieve accurate fusion and feature extraction of multi-dimensional data. Specifically, this disclosure designs differentiated alignment strategies for the spatiotemporal characteristics of different data sources: In the time dimension, considering the nonlinear variation characteristics of photovoltaic meteorological data during sunrise and sunset, an adaptive interpolation method based on the rate of change of solar altitude angle is adopted, effectively overcoming the insufficient accuracy of traditional linear interpolation during periods of drastic changes in illumination; in the spatial dimension, refined spatial registration is performed by combining digital elevation models and land cover type data, eliminating spatial deviations caused by factors such as terrain shading and surface reflection. Simultaneously, through an encoder-self-attention mechanism multimodal feature fusion architecture, deep interaction and information complementarity of heterogeneous features are achieved.

[0027] This disclosure outperforms traditional splicing and interpolation methods, providing a high-quality dataset with a unified spatiotemporal benchmark for subsequent photovoltaic power output prediction. It effectively integrates multi-source data such as solar irradiance, meteorological data, and numerical weather prediction, enhancing not only the richness and discriminative power of feature representations but also strengthening the predictive model's adaptability to complex environmental changes. Significant improvements have been achieved in both model prediction accuracy and robustness, providing reliable technical support for refined power prediction of distributed photovoltaic power plants.

[0028] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0029] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments, taken in conjunction with the accompanying drawings. In the drawings: Figure 1 A flowchart is shown below illustrating a method for predicting photovoltaic power output of a distributed photovoltaic power station based on multi-source heterogeneous feature data according to an embodiment of the present disclosure; Figure 2 A comparison diagram showing the sampling time intervals of different multidimensional feature sequences in embodiments of this disclosure is provided. Figure 3 A comparison diagram showing the spatial resolution of different multidimensional feature sequences in embodiments of this disclosure is provided. Figure 4 This diagram illustrates the time alignment processing of meteorological temperature and NWP data in an embodiment of the present disclosure. Figure 5 A schematic diagram showing a performance comparison between an adaptive spatiotemporal fusion method according to an embodiment of the present disclosure and an existing fusion method is illustrated. Figure 6 A structural diagram of a distributed photovoltaic power plant photovoltaic output prediction system based on multi-source heterogeneous feature data, according to an embodiment of the present disclosure, is shown. Detailed Implementation

[0030] In the following, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings to enable those skilled in the art to readily implement them. Furthermore, for clarity, portions unrelated to the description of exemplary embodiments have been omitted from the drawings.

[0031] In this disclosure, it should be understood that terms such as “comprising” or “having” are intended to indicate the presence of features, figures, steps, behaviors, components, parts or combinations thereof disclosed in this specification, and are not intended to exclude the possibility of the presence or addition of one or more other features, figures, steps, behaviors, components, parts or combinations thereof.

[0032] It should also be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0033] In this disclosure, any operation involving the acquisition of user information or user data, or the display of user information or user data to others, is an operation authorized or confirmed by the user, or actively selected by the user.

[0034] As mentioned earlier, existing multi-source data fusion methods face many technical bottlenecks when dealing with photovoltaic power output prediction: First, traditional time alignment methods often employ fixed-interval linear interpolation or nearest-neighbor interpolation strategies, failing to fully consider the physical changes of different data sources over different time periods. For example, meteorological data is typically sampled at the minute or hour level, while photovoltaic power output data may be sampled at the second or minute level, resulting in a significant deviation between the interpolation results and the actual physical process, which in turn affects the input quality of subsequent prediction models.

[0035] Secondly, existing methods often employ simple grid resampling or nearest neighbor matching strategies. For example, the spatial resolution of satellite remote sensing data is typically at the kilometer level, while ground observation data reflects local conditions. Direct spatial matching can lead to scale effect errors and fail to accurately reflect the impact of local microclimates on photovoltaic power output.

[0036] To address the aforementioned technical problems, this disclosure focuses on the fusion and processing of multi-source heterogeneous data, and provides a method for predicting the photovoltaic output of distributed photovoltaic power plants based on multi-source heterogeneous feature data, including: Obtain a multi-source heterogeneous feature dataset of the target distributed photovoltaic power station collected over a historical period. The multi-source heterogeneous feature dataset includes multiple multi-dimensional feature sequences from independent data sources. Each multi-dimensional feature sequence includes one multi-dimensional feature data or multiple multi-dimensional feature data arranged in chronological order. The multi-source heterogeneous feature dataset is subjected to time alignment and spatial alignment processing to obtain a spatiotemporally aligned multi-source heterogeneous feature dataset. Specifically, when the sampling frequency of the multi-dimensional feature sequence is lower than a preset reference frequency, the historical time period is divided into multiple interpolation time periods based on the rate of change of the solar altitude angle and the data type of the multi-dimensional feature sequence. Based on the data type of the multi-dimensional feature data within each interpolation time period and its corresponding interpolation time period, a matching interpolation method is selected for resampling to align the corresponding multi-dimensional feature data to the reference time axis determined by the reference frequency. For each moment on the reference time axis, all multidimensional feature sequences in the spatially aligned multi-source heterogeneous feature dataset are input into the corresponding encoder to extract initial features; each initial feature is subjected to dimensionality reduction and feature mapping to obtain a feature vector with a specified dimension; all feature vectors with the specified dimension are concatenated to form a multimodal heterogeneous feature set; the multimodal heterogeneous feature set is processed using a self-attention mechanism to obtain attention output features; the multimodal heterogeneous feature set and the attention output features are subjected to feature fusion processing to obtain a joint feature vector with the specified dimension at the corresponding moment. The joint feature vector at each moment on the reference time axis is input into the machine learning model for training. The parameters of the machine learning model are adjusted according to the difference between the predicted power generation and the actual power generation at each moment until the training stopping condition is met, thus obtaining the photovoltaic power output prediction model. The photovoltaic power output prediction model is then used to predict the photovoltaic power output of the target distributed photovoltaic power station.

[0037] This disclosure constructs a high-quality feature set that is complementary in information and consistent in time and space through precise spatiotemporal fusion and deep feature interaction of multi-source heterogeneous data, thereby improving the accuracy and robustness of photovoltaic power output prediction. It not only effectively overcomes the information loss and noise introduction problems of traditional data fusion methods when processing heterogeneous data, but also provides a high-quality data foundation with a unified spatiotemporal benchmark for the field of photovoltaic prediction, which has significant practical implications for promoting the large-scale application of distributed photovoltaics and ensuring the safe and stable operation of new power systems.

[0038] Figure 1 A flowchart is shown below illustrating a method for predicting photovoltaic power output of a distributed photovoltaic power station based on multi-source heterogeneous feature data, according to an embodiment of the present disclosure.

[0039] In this disclosure, the distributed photovoltaic power station is a photovoltaic power generation system installed on the user side or the distribution network side, with a small single-point scale, and mainly for local consumption.

[0040] It is known that distributed photovoltaic (PV) power stations are characterized by their large number, small scale at individual points, and significant susceptibility to local microclimates. Their output prediction differs fundamentally from centralized PV in terms of data sources, feature dimensions, and spatiotemporal scale. Addressing the common technical bottlenecks in distributed PV prediction, such as data disorder, multi-source heterogeneity, and inconsistent spatiotemporal benchmarks, this disclosure effectively solves the problem of the difficulty in collaboratively utilizing multi-source data in distributed scenarios.

[0041] like Figure 1 As shown, the prediction method includes the following steps S101~S104: In step S101, a multi-source heterogeneous feature dataset collected from the target distributed photovoltaic power station over a historical period is obtained. The multi-source heterogeneous feature dataset includes multiple multi-dimensional feature sequences from independent data sources. Each multi-dimensional feature sequence includes one multi-dimensional feature data or multiple multi-dimensional feature data arranged in chronological order.

[0042] Among them, the target distributed photovoltaic power station refers to the distributed photovoltaic power station that requires photovoltaic output prediction.

[0043] The historical time period refers to the historical data collection cycle covering the continuous operation of the target distributed photovoltaic power station. The time span can be set according to actual needs, such as no less than one year, to ensure that the output characteristics under different seasons and weather conditions are covered, i.e., the collected multi-source heterogeneous feature dataset.

[0044] Specifically, the data source for the multidimensional feature sequence is any one of the following: Solar radiation data, including solar radiation parameters such as total irradiance, direct irradiance, and diffuse irradiance; Surface meteorological observation data, including surface meteorological elements such as temperature, humidity, wind speed, wind direction, and air pressure; Numerical weather forecast (NWP) data, including numerical forecast results for meteorological elements such as temperature, humidity, wind speed, wind direction, and air pressure; Satellite remote sensing data, including satellite cloud images and information on the distribution of cloud systems; Geospatial static data, including digital elevation models (DEM), slope, aspect, land cover type, and other topographic and surface features; Cloud cover observation data, including cloud cover information such as cloud coverage and cloud type classification; Aerosol data, including atmospheric aerosol parameters such as aerosol optical thickness.

[0045] Figure 2 A comparison diagram showing the sampling time intervals of different multidimensional feature sequences in embodiments of this disclosure is provided. Figure 3 A comparison diagram showing the spatial resolution of different multidimensional feature sequences in embodiments of this disclosure is presented.

[0046] Since the aforementioned multidimensional feature sequences originate from different data acquisition systems and observation platforms, they not only vary in data dimensions but also exhibit significant differences in spatiotemporal scales.

[0047] like Figure 2 and Figure 3 It is evident that, in the time dimension, the sampling frequencies of different data sources vary. For example, solar radiation data can be sampled at a frequency of 1 minute, ground meteorological observation data at a frequency of 15 minutes, while NWP data is typically sampled at a frequency of 3 hours. In the spatial dimension, the spatial resolutions of different data sources also differ. For instance, NWP data has a spatial resolution of 3 km, satellite remote sensing data has a spatial resolution of 2 km, while geospatial static data has a spatial resolution of 30 m.

[0048] Based on the differences in the sampling frequencies mentioned above, the sampling frequencies of the multidimensional feature sequences can be divided into high frequency, medium frequency, or low frequency.

[0049] Wherein, the high frequency, the medium frequency and the low frequency correspond to the first set sampling frequency range, the second set sampling frequency range and the third set sampling frequency range, respectively, and the lower limit of the first set sampling frequency range is greater than the upper limit of the second set sampling frequency range, and the lower limit of the second set sampling frequency range is greater than the upper limit of the third set sampling frequency range.

[0050] The reference frequency is a preset frequency value within the first set sampling frequency range, used to establish the reference time axis; the preset frequency value is any frequency value within the first set sampling frequency range, which can be set according to user needs.

[0051] Specifically, starting from the beginning of the historical time period, time points with equal intervals are generated according to the time intervals corresponding to the reference frequency, which serve as a unified time reference for the spatiotemporal alignment of the multi-source heterogeneous feature dataset, i.e., the reference time axis.

[0052] Wherein, the reference frequency is assumed to be The corresponding sampling time interval is Let's assume the starting time of the historical period is... The end time is Then the reference time axis This can be expressed by the following formula: ,in, , It is a non-negative integer.

[0053] According to embodiments of this disclosure, when the multidimensional feature sequence includes multiple multidimensional feature data, each multidimensional feature data is a multivariate temporal feature data of a single spatial location at the spatial resolution of the corresponding data source; When the multidimensional feature sequence includes a multidimensional feature data, the multidimensional feature data is multivariate static feature data of a single spatial location at the spatial resolution of the corresponding data source.

[0054] According to embodiments of this disclosure, before performing the time alignment process, each multidimensional feature sequence is decoupled by variable dimensions to obtain multiple corresponding univariate feature sequences; wherein, when the multidimensional feature sequence includes multiple multidimensional feature data, the univariate feature sequence includes multiple univariate time-series feature data arranged in chronological order; when the multidimensional feature sequence includes one multidimensional feature data, the univariate feature sequence includes one univariate static feature data; the univariate feature sequence is provided with corresponding position coordinates and variable identifiers.

[0055] In this disclosure, a multidimensional feature sequence is a collection of data elements, and the number of data elements is not limited. When a multidimensional feature sequence contains multiple multidimensional feature data, it is time-series data; when a multidimensional feature sequence contains only one multidimensional feature data, it is static data, which can be regarded as a special sequence of length 1. Based on this, the variable dimension decoupling also applies to these two cases.

[0056] The following example uses a multidimensional feature sequence from the NWP data source, which includes multiple multidimensional feature data, to illustrate the decoupling of variable dimensions in the multidimensional feature sequence.

[0057] Assuming the multidimensional feature sequence is an NWP data sequence (time × variable), where the variables are temperature, humidity, wind speed, air pressure, precipitation, etc.; then the multidimensional feature sequence includes multiple NWP data (time × variable) at a single spatial location under the spatial resolution of the corresponding NWP data, and the NWP data is the multivariate time series feature data.

[0058] The NWP data sequence is decoupled according to the variable dimension to obtain multiple univariate feature sequences, namely: temperature feature sequence, humidity feature sequence, wind speed feature sequence, air pressure feature sequence, and precipitation feature sequence. Each univariate feature sequence contains the location coordinates (e.g., grid point index) and the corresponding variable identifier (e.g., temperature, humidity).

[0059] The following example, a multidimensional feature sequence from a data source (geographic data) including a multidimensional feature data, is used to illustrate the decoupling of the variable dimensions of the multidimensional feature sequence.

[0060] Suppose that the multidimensional feature sequence includes multivariate static feature data of a single spatial location at the spatial resolution of the corresponding geographic data. This multivariate static feature data includes digital elevation model (DEM), slope, aspect, land cover type, etc.

[0061] The multivariate static feature data is decoupled according to the variable dimensions to obtain multiple univariate feature sequences. Each univariate feature sequence includes one univariate static feature data, namely: DEM data, slope data, aspect data, and land cover type data. Each univariate static feature data contains the location coordinates of the spatial location and the corresponding variable identifier (e.g., DEM, slope).

[0062] In step S102, the multi-source heterogeneous feature dataset is subjected to time alignment and spatial alignment processing to obtain a spatiotemporally aligned multi-source heterogeneous feature dataset. When the sampling frequency of the multidimensional feature sequence is lower than a preset reference frequency, the historical time period is divided into multiple interpolation time periods based on the solar altitude angle change rate and the data type of the multidimensional feature sequence. Based on the data type of the multidimensional feature data within each interpolation time period and the corresponding interpolation time period, a matching interpolation method is selected for resampling to align the corresponding multidimensional feature data to the reference time axis determined by the reference frequency.

[0063] Due to the rate of change of solar altitude angle This reflects the degree of drastic change in solar irradiance over time, and the calculation formula is shown below: ; in, Geographical latitude, The solar declination angle, Solar hour angle, (Angle / second) represents the rate of change of the hour angle. This represents the current solar altitude angle.

[0064] Specifically, this disclosure will use the absolute value of the rate of change of the solar altitude angle. The core criterion for determining the segmentation of the interpolation strategy is: during sunrise and sunset periods. Angle / second. At this point, the solar altitude angle and solar irradiance change drastically, resulting in large linear interpolation errors. Therefore, nonlinear interpolation is required. Angle / second, with a small rate of change, the spline maintains the continuity of the curve, and cubic spline interpolation is used; specifically, different interpolation strategies are used in segments on the reference time axis according to physical laws, so that the data better matches the actual values ​​during sunrise and sunset.

[0065] According to an embodiment of this disclosure, when the sampling frequency of the multidimensional feature sequence is lower than the reference frequency, interpolation processing is performed on the multiple univariate feature sequences corresponding to the multidimensional feature sequence to align the univariate feature sequences to the reference time axis.

[0066] Furthermore, when the univariate feature sequence includes multiple univariate time-series feature data arranged in chronological order, a matching interpolation method is selected for resampling based on the data type of the univariate time-series feature data within each interpolation time period and the interpolation time period to which it belongs.

[0067] The univariate time-series feature data includes any of the following data types: radiation data, environmental state data, cumulative data, and vector data.

[0068] According to an embodiment of this disclosure, when the time period to be interpolated is a nighttime period, for the univariate time series feature data within the nighttime period, cubic spline interpolation is performed when the data type of the univariate time series feature data is environmental state data, and forced zeroing processing is performed when the data type of the univariate time series feature data is radiation data.

[0069] According to an embodiment of this disclosure, when the time period to be interpolated is a stable daytime period, cubic spline interpolation is performed on the univariate time series feature data within the stable daytime period when the data type of the univariate time series feature data is radiation data or environmental state data.

[0070] According to embodiments of this disclosure, monotonic interpolation is performed when the data type of the univariate time series feature data is cumulative data.

[0071] Figure 4 This diagram illustrates the time alignment processing of meteorological temperature and NWP data in an embodiment of the present disclosure.

[0072] like Figure 4 As shown, the reference time axis is determined by the sampling frequency of irradiance, and time alignment processing is performed on the meteorological temperature and NWP data. The meteorological temperature data (solid green line, 1-minute sampling) is already basically synchronized with the reference time axis. For the environmental state data sampled every 15 minutes (dashed green line), cubic spline interpolation is used to generate a continuous sequence synchronized with the reference time axis. For the NWP data updated every 3 hours (dashed red line), because it contains irradiance components, forced zeroing is performed during the nighttime period, and cubic spline interpolation is used during the daytime period. Then, time shifting is used to correct its systematic time delay, aligning it with the reference time axis. Finally, all sequences are mapped to a unified minute-level time grid, achieving time dimension alignment.

[0073] According to embodiments of this disclosure, when the data type of the univariate time-series feature data is vector data, component-independent cubic spline interpolation is performed. That is, cubic spline interpolation is performed on each component of the vector data separately. Taking wind speed as an example, since the wind vector field has directionality, cubic spline interpolation needs to be performed independently on the orthogonal components (U and V) of the horizontal wind speed vector in the Cartesian coordinate system. After interpolation, the wind speed is calculated according to the vector composition formula, and the wind direction is determined according to the sign of the U / V components. It should be noted that it is forbidden to interpolate the scalar wind speed first, otherwise the sign and phase information of the components will be destroyed, resulting in the error of wind direction distortion in the interpolation result.

[0074] Furthermore, when the univariate feature sequence includes a univariate static feature data, the univariate static feature data is copied along the reference time axis to generate a time-series feature sequence covering each moment of the reference time axis.

[0075] It should be noted that static feature data is typically not stored in time-series form in the original data source. In this disclosure, in order to construct a unified time-series feature sequence, the static data is treated as a special sequence with constant values ​​on the time axis. Therefore, in the time alignment step, this constant value characteristic allows the static feature data to be directly mapped or broadcast to each moment of the reference time axis without the need for complex interpolation calculations.

[0076] According to an embodiment of this disclosure, when the sampling frequency of the multidimensional feature sequence is higher than the reference frequency, the multiple univariate feature sequences corresponding to the multidimensional feature sequence are downsampled respectively to align the univariate feature sequences to the reference time axis.

[0077] The inventors noted that in actual data acquisition, in addition to needing to adopt differentiated interpolation strategies for different data types and time periods, they often face the challenge of missing data. When a data point in a data source is missing due to equipment failure, communication interruption, or environmental interference, directly using simple linear interpolation or deletion will lead to discontinuity in the data sequence and information loss, thus affecting the accuracy of subsequent photovoltaic power prediction. Therefore, intelligent filling of the missing data point is also necessary.

[0078] According to embodiments of this disclosure, when data loss is detected in the univariate feature sequence, for each missing feature data in the univariate feature sequence, the following is executed: Multiple candidate auxiliary data sources are acquired, and the comprehensive spatiotemporal correlation coefficient between each candidate auxiliary data source and the missing feature data is calculated. A threshold is set based on the comprehensive spatiotemporal correlation coefficient. The multiple candidate auxiliary data sources are filtered according to the threshold to obtain one or more reliable auxiliary sources. The observation data of each reliable auxiliary source at the corresponding target time is acquired, and the observation data is converted into an estimated value with the same dimensions as the missing feature data through the corresponding physical inversion model, wherein the target time is the sampling time of the missing feature data. The corresponding estimated values ​​are weighted and fused according to the normalized weights of each reliable auxiliary source to generate the imputation value of the missing feature data.

[0079] The following is a detailed explanation of how to select a trustworthy auxiliary source from candidate auxiliary sources: For each candidate auxiliary source, calculate its comprehensive spatiotemporal correlation coefficient with the target missing point (missing feature data): ; in, , which is the Pearson correlation coefficient, used to measure the synchronicity between the target missing point and the historical time series of candidate auxiliary sources; This is a spatial inverse distance weighting coefficient used to measure the spatial proximity between the target missing point and the candidate auxiliary source (the closer the distance, the larger the value). : Prioritizes weights based on the time dimension; thresholds include: , , .

[0080] The candidate auxiliary source is determined as a reliable auxiliary source only when all three of the threshold conditions are met simultaneously; otherwise, it is discarded to avoid the introduction of bias by low-relevance sources.

[0081] The inventors also noted that even if multiple data sources in a multi-source heterogeneous feature dataset record the same physical process, the asynchronous clocks of the acquisition systems—for example, one source clock being faster or slower—can cause deviations in the time dimension of the univariate feature sequences. To address this, this disclosure introduces a sliding window alignment mechanism: within the longest overlap period of two source data, cross-correlation is performed on the two source data to find the time offset Δt that makes them most similar (with the highest correlation). Finally, the data source with the slower or biased clock is shifted by Δt, thereby achieving unified correction in the time dimension. Specifically, the clock of the high-frequency data is used as the reference, and other low-frequency or asynchronous multidimensional feature sequences are aligned to this reference.

[0082] According to an embodiment of this disclosure, using the clock of a high-frequency sampled univariate feature sequence as a reference clock, when a clock asynchrony is detected between a univariate feature sequence that records the same physical process as the high-frequency sampled univariate feature sequence, cross-correlation is performed within the longest overlapping period of the two sets of univariate feature sequences to obtain a time offset Δt that maximizes the correlation between the two sets of univariate feature sequences; the univariate feature sequences with clock deviation are shifted as a whole by Δt to achieve unified correction of the time dimension; wherein, the univariate feature sequences with clock deviation are low-frequency or asynchronous univariate feature sequences.

[0083] For example, suppose a high-frequency sampled univariate feature sequence A comes from solar radiation data, with a sampling frequency of 1 minute, and its clock is used as the reference clock; another univariate feature sequence B is detected from ground meteorological observation data, with a sampling frequency of 15 minutes, and there is an unknown clock deviation between B and A, with the preliminary estimate of the deviation range being within ±10 minutes.

[0084] At this point, the longest overlapping period between the two sequences is found to be 08:00-12:00 on a certain day, totaling 240 minutes. Therefore, this 240-minute period is selected as the calculation window. At this time, sequence A contains 240 data points, and sequence B contains 16 data points. Sequence B is interpolated and resampled to the same minute scale as sequence A, resulting in 240 data points to achieve time alignment. Then, the cross-correlation coefficient between sequence A and the interpolated sequence B within the range of [-10 minutes, +10 minutes] is calculated, and the time offset corresponding to the maximum value of the cross-correlation coefficient is taken as Δt. If the calculated Δt = -3 minutes, then all data of sequence B is shifted 3 minutes to the left on the reference time axis, thereby completing the time synchronization correction.

[0085] According to embodiments of this disclosure, after the time alignment process is completed, multiple time-aligned single-variable time-series feature sequences belonging to the same data source are recombined by spatial and variable mapping based on their corresponding position coordinates and variable identifiers to obtain time-aligned multidimensional feature sequences; wherein, the time-aligned multidimensional feature sequences are aligned with the reference time axis in the time dimension, and the spatial and variable dimensions are consistent with the multidimensional feature sequences before time processing.

[0086] In order to accurately calibrate time, this disclosure splits a multidimensional data into multiple individual one-dimensional time series for processing during time alignment. After the time alignment is completed, each one-dimensional time series needs to be reassembled back into the original multidimensional data structure according to its original position coordinates and variable identifiers.

[0087] This ensures that the reconstructed data retains the original spatial distribution and variable relationships while eliminating errors caused by clock asynchrony, thus providing a foundation for subsequent spatial alignment processing.

[0088] According to embodiments of this disclosure, after completing the time alignment process, the time-aligned multidimensional feature sequence is subjected to spatial alignment processing, the spatial alignment process including: A unified fine grid is established for the target area, wherein the spatial resolution of the unified fine grid is higher than the spatial resolution of the data source corresponding to the multidimensional feature sequence, and the target area is the preset area range of the target distributed photovoltaic power station; for each moment on the reference time axis, the multidimensional feature sequence corresponding to each moment is mapped onto the unified fine grid to obtain the background value of each fine grid point; high-resolution digital elevation model grid data and land cover type grid data corresponding to each fine grid point in the unified fine grid are obtained; the background value of the corresponding grid point is corrected according to the digital elevation model grid data and the land cover type grid data to obtain the spatially aligned multidimensional feature sequence.

[0089] It is known that in terms of spatial alignment, existing technologies often use simple grid interpolation or inverse distance weighting methods, ignoring the impact of spatial factors such as terrain elevation and land cover type on photovoltaic power output, which leads to spatial mismatch problems in the fusion process of data with different spatial resolutions.

[0090] This disclosure constructs a high-resolution unified fine grid as a spatial alignment benchmark and integrates digital elevation model and land cover type data to physically correct the mapped background values, thereby achieving fine alignment of heterogeneous feature data in the spatial dimension. This effectively eliminates the spatial mismatch problem caused by differences in terrain elevation and land cover heterogeneity, and significantly improves the accuracy of photovoltaic power output prediction.

[0091] The following uses NWP temperature data and surface solar radiation data as examples to illustrate spatial alignment processing: The resolution of the uniform fine grid in the target area is set to 100m×100m, that is, the target area is uniformly divided into a 100m×100m grid, and each fine grid point is assigned a corresponding spatial coordinate.

[0092] Temperature correction: The 3km grid data of NWP temperature data is spatially interpolated onto a 100m grid to obtain the background temperature value for each fine grid point. For example, if the temperature at a 3km grid point is 25°C, after interpolation, the initial temperature of all 100m grid points within its coverage area will also be 25°C. Then, the high-resolution digital elevation model is read to obtain the elevation of each 100m fine grid point. Elevation of corresponding NWP coarse grid points Calculate the elevation difference Then, a correction is made based on the vertical temperature lapse rate: ;in, This represents the vertical temperature lapse rate, typically taken as 6.5°C / km. This is the background temperature value. This is the corrected temperature value.

[0093] Radiation correction: For surface solar radiation, the 3km grid data of NWP surface solar radiation data is spatially interpolated onto a 100m grid to obtain the radiation background value for each fine grid point. Land cover type data, such as buildings, bare soil, farmland, water bodies, and vegetation, are read; each type corresponds to an empirical albedo value. Correcting surface solar radiation data using albedo: ;in, This is the surface downwave shortwave radiation background value after NWP interpolation. This is to correct the net absorption of shortwave radiation on the Earth's surface.

[0094] In step S103, for each moment on the reference time axis, all multidimensional feature sequences in the spatially aligned multi-source heterogeneous feature dataset are input into the corresponding encoder to extract initial features; each initial feature is subjected to dimensionality reduction and feature mapping to obtain a feature vector with a specified dimension; all feature vectors with the specified dimension are concatenated into a multimodal heterogeneous feature set; the multimodal heterogeneous feature set is processed using a self-attention mechanism to obtain attention output features; the multimodal heterogeneous feature set and the attention output features are subjected to feature fusion processing to obtain a joint feature vector with the specified dimension at the corresponding moment.

[0095] In this disclosure, after completing the spatiotemporal alignment process, it is also necessary to align the dimensions of different multidimensional feature sequences.

[0096] Specifically, when the data structure of the multidimensional feature sequence is time-series data, a one-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is image data, a two-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is grid data, a three-dimensional convolutional encoder is used to extract the corresponding initial features; and when the data structure of the multidimensional feature sequence is scalar data, a fully connected encoder is used to extract the corresponding initial features.

[0097] Wherein, the time-series data are time-series observations of a single point; the image data are two-dimensional remote sensing images; the grid data are three-dimensional meteorological field data containing vertical layers; and the static data are geographic environmental parameters that do not change over time.

[0098] According to embodiments of this disclosure, each initial feature is subjected to dimensionality reduction processing, including: When the data structure of the multidimensional feature sequence is time-series data, global average pooling is performed along the time axis; when the data structure of the multidimensional feature sequence is image data, global average pooling is performed along the spatial axis; when the data structure of the multidimensional feature sequence is gridded data, global average pooling is performed along the three-dimensional spatial axis; when the data structure of the multidimensional feature sequence is scalar data, no pooling is required, and it is directly output. See Table 1.

[0099] Table 1. Dimension Alignment Operations for Multidimensional Feature Sequences of Different Dimensions

[0100] According to embodiments of this disclosure, each initial feature undergoing dimensionality reduction is subjected to feature mapping processing, including: For the initial features of the time series data that have undergone dimensionality reduction processing, global average pooling is performed along the time axis, followed by a fully connected layer to output the feature vector of the specified dimension.

[0101] For the initial features of the image data that have undergone dimensionality reduction processing, global average pooling is performed along the height × width, followed by a fully connected layer to output a vector of the same dimension.

[0102] For the initial features of the dimensionality reduction processing of the grid data, global average pooling is performed along depth × height × width, followed by a fully connected layer to output a vector of the same dimension.

[0103] The initial features for dimensionality reduction processing of the static scalar data are: they are already vectors, and can be directly output as vectors of the same dimension through a fully connected layer (or identity mapping).

[0104] In this disclosure, all heterogeneous feature data, after the above-mentioned dimensionality reduction and feature mapping processes, are transformed into flat vectors of the same dimension (e.g., 1×256), which can be directly concatenated into a multimodal feature set for use by downstream models.

[0105] The inventors noted that existing technologies, when processing multi-source heterogeneous data, typically employ direct splicing or simple weighted averaging, failing to fully exploit the inherent correlations and complementarities between different data sources. This results in insufficient utilization of feature information and may even introduce additional noise due to differences in units and inconsistent data distribution.

[0106] In this disclosure, a self-attention mechanism can also be used to calculate the correlation weight matrix between each feature vector in the multimodal heterogeneous feature set; the multimodal heterogeneous feature set is then updated with weights based on the correlation weight matrix to obtain the attention output feature.

[0107] Furthermore, the feature fusion process of the multimodal heterogeneous feature set and the attention output feature includes: normalizing the multimodal heterogeneous feature set and the attention output feature after residual connection to obtain fused intermediate features; inputting the fused intermediate features into a multilayer perceptron for mapping processing; and then performing mean aggregation processing on each feature vector after mapping processing to obtain a joint feature vector with the specified dimension at the corresponding time.

[0108] The generation of joint feature vectors in this disclosure will be described below with a specific embodiment.

[0109] Suppose that dimensionality reduction and feature mapping are performed on multiple initial features to obtain time modality features. Image modal features Numerical weather forecast modal characteristics Spatial static modal characteristics A multimodal heterogeneous feature set can be obtained by directly concatenating the features. .

[0110] Then, the multimodal heterogeneous feature set is processed using a self-attention mechanism: Query Matrix Key matrix Value matrix ; Attention output features V ;in, ; , , , are all learnable linear projection weight matrices, used to generate the query matrix, key matrix, and value matrix, respectively.

[0111] This attention mechanism essentially models the cross-modal interaction relationships between four modal sources: each row corresponds to one modality, and multi-source information is dynamically fused by calculating its attention weights to all modalities. In the final attention output features, each row is a modal representation enhanced with context—for example, when it's sunny. and Attention is highly valued; on cloudy days Weighting increased.

[0112] This disclosure takes into account the varying reliability of data from different data sources under different weather conditions and employs an adaptive weight allocation mechanism based on different weather conditions. In sunny scenarios, since solar irradiance data directly reflects the availability of solar energy, therefore... It has the highest weighting; in multi-cloud scenarios, satellite cloud imagery data is favored because it can capture changes in cloud cover. Weighting percentage; In rainy scenarios, NWP forecast data has a stronger predictive ability for large-scale weather systems, therefore, it is increased. Weighting percentage.

[0113] Table 2 provides examples of weight allocation for each underlying data source under different weather scenarios. It should be noted that the above multi-source data are used to construct a multimodal heterogeneous feature set. Previously, preliminary aggregation was performed according to modal type: time series data such as solar radiation, surface meteorological observations, cloud cover, and aerosols were aggregated into time modal features. Satellite remote sensing data as image modal features Numerical weather forecast data serves as a modal feature of numerical weather forecasting. Geographic static data as spatial static modal features Therefore, the self-attention mechanism models cross-modal interactions at a 4×4 modal level, while the weight allocation in Table 2 is reflected at a more granular data source level, with each serving a different level of abstraction.

[0114] Table 2 Adaptive weight allocation table for each data source under different weather scenarios

[0115] Finally, the multimodal heterogeneous feature set and the attention output features are fused to output a 256-dimensional joint feature vector, which is achieved through the following formula: ; ; ; in, For layer normalization operation; To complete the enhancement feature matrix for residual connections and layer normalization; For a two-layer fully connected network with shared weights, Each line is independently applied to the same MLP, and the output is F; This represents the enhanced feature direction corresponding to the i-th mode; The arithmetic mean of the four modal enhancement features is used as the joint feature vector.

[0116] In step S104, the joint feature vector at each moment on the reference time axis is input into the machine learning model for training. The parameters of the machine learning model are adjusted according to the difference between the predicted power generation and the actual power generation at each moment until the training stopping condition is met, thus obtaining the photovoltaic power output prediction model. The photovoltaic power output prediction model is then used to predict the photovoltaic power output of the target distributed photovoltaic power station.

[0117] In this disclosure, multi-source data fusion is used to make up for the observation blind spots and error defects of single data sources, spatiotemporal alignment eliminates feature misalignment caused by asynchronous sampling, and deep feature interaction uncovers the implicit correlation between meteorological elements and photovoltaic output. The resulting fused features not only have higher information density and less noise, but also more comprehensively characterize the photovoltaic response patterns under complex weather conditions, providing a more reliable and discriminative input basis for photovoltaic output prediction models.

[0118] The inventors further recognized that the fusion of the heterogeneous feature data, in addition to alignment in time, space, and dimension, also needs to conform to physical laws. For example, the total solar irradiance should not exceed the solar constant, and the sum of the direct, scattered, and reflected components of solar radiation should equal the total incident irradiance to satisfy the law of energy conservation; when cloud cover increases, the solar irradiance reaching the ground should decrease accordingly to conform to the laws of atmospheric transport.

[0119] According to embodiments of this disclosure, the physical constraints on the heterogeneous feature data include: Radiation balance constraint: The total irradiance after fusion should be equal to the sum of direct irradiance, diffuse irradiance and reflected irradiance, and should not exceed the solar constant at the top of the atmosphere.

[0120] Energy conservation constraint: During the spatiotemporal fusion process, when large-scale NWP data is downscaled to a small-scale grid, the total energy of the grid cells should remain conserved.

[0121] Topographic consistency constraints: Meteorological elements after correction of high-resolution topographic data should meet the altitude gradient law, such as the lapse rate constraint that the temperature decreases with increasing altitude.

[0122] Cloud cover-irradiance correlation constraints: The fused cloud cover data and irradiance data should satisfy a statistically inverse correlation, and irradiance should decrease accordingly when cloud cover increases.

[0123] The total physical constraint loss is obtained by summing the loss values ​​of the four physical constraints mentioned above. This total loss reflects the extent to which the fusion result violates physical laws; the smaller the loss value, the better the physical consistency.

[0124] In this disclosure, physical constraints can be incorporated into the optimization objective of the prediction model in two ways: one is the Lagrange multiplier method, which introduces physical constraints as hard conditions into the optimization process and balances the task objective and physical constraints by adjusting the multiplier coefficients; the other is the penalty function method, which directly adds the physical constraint loss as a regularization term to the total loss function and controls the strength of the physical constraints by adjusting the penalty coefficient.

[0125] This disclosure achieves effective fusion of multi-source data, including solar irradiance, meteorological data, and numerical weather prediction, which not only enhances the richness and discriminative power of feature representations but also strengthens the predictive model's adaptability to complex environmental changes. Through multi-source data redundancy complementarity, anomaly suppression mechanisms, and targeted spatiotemporal alignment strategies, significant improvements have been made in prediction accuracy and robustness, providing reliable technical support for refined power prediction of distributed photovoltaic power plants.

[0126] To fully verify the independent contribution and synergistic effect of this disclosure on data fusion quality, an ablation experiment was designed. Based on the full model, key modules were gradually removed or replaced, and the following five comparison schemes were constructed: Model A (Complete Model): Includes adaptive temporal alignment, multi-resolution spatial registration, heterogeneous feature joint encoding, physical constraint consistency fusion, and adaptive weight optimization.

[0127] Model B (Remove Adaptive Time Alignment): Replaces adaptive time alignment with simple linear interpolation alignment, leaving the rest of the modules unchanged.

[0128] Model C (Removal of Multi-Resolution Spatial Registration): Replaces spatial downscaling based on terrain and land cover with bilinear interpolation resampling, while the rest of the modules remain unchanged.

[0129] Model D (Removal of Physical Constraints Consistency Fusion): Removes physical constraints such as radiation balance, energy conservation, terrain consistency, and cloud cover-irradiance correlation, retaining only mathematical alignment and feature encoding.

[0130] Model E (Removal of Heterogeneous Feature Joint Encoding): The outputs of each encoder are directly concatenated without performing unified feature space mapping and joint representation learning.

[0131] Model F (Removal of Adaptive Weight Optimization): Employs a fixed equal-weight fusion strategy, without dynamically adjusting the weights of each data source based on weather scenarios.

[0132] The experiment evaluated the results from five dimensions: temporal alignment accuracy, spatial registration accuracy, data integrity, feature consistency, and fusion reliability. The results are shown in Table 3, which compares the ablation experiment results. Table 3 Comparison of ablation test results

[0133] As shown in Table 3, the results show that each module contributes positively to the fusion quality. Among them, adaptive temporal alignment and multi-resolution spatial registration are the basic modules to ensure the consistency of spatiotemporal references, physical constraint consistency fusion and heterogeneous feature joint encoding are the core modules to improve the fusion quality, and adaptive weight optimization is the key module to enhance the robustness of fusion.

[0134] Figure 5 A schematic diagram showing a performance comparison between an adaptive spatiotemporal fusion method according to an embodiment of the present disclosure and an existing fusion method is illustrated.

[0135] like Figure 5 As shown, based on actual data from a distributed photovoltaic demonstration zone in a certain province, the performance of the method (adaptive spatiotemporal fusion) provided in this disclosure is compared with existing fusion methods—simple stitching, linear interpolation alignment, and nearest neighbor resampling—from five dimensions: temporal alignment accuracy, spatial registration accuracy, data integrity, feature consistency, and fusion reliability. Figure 5 As can be seen, the method provided in this disclosure is superior to existing fusion methods in all evaluation indicators. In particular, the fusion credibility index of the method provided in this disclosure reaches 0.85, which is far better than the other three comparison methods. This shows that the method provided in this disclosure can effectively control information loss and error accumulation in the fusion process.

[0136] The following verifies the effectiveness of using the fused joint feature vector for photovoltaic power output prediction in this disclosure: Specifically, using the same LightGBM model, the fused data obtained from the following four data fusion schemes were input respectively to perform ultra-short-term (15 min) and short-term (4 h) power prediction. The prediction accuracy results are shown in Table 4: Table 4. Impact of different fused data on photovoltaic forecast accuracy

[0137] Therefore, the joint eigenvector of this disclosure reduces the nRMSE (normalized root mean square error) of ultra-short-term prediction from 8.42% to 6.62%, a decrease of 21.4%, and the nRMSE of short-term prediction from 18.65% to 15.72%, a decrease of 15.7%, directly demonstrating the significant gain of the fused data of this disclosure for downstream prediction tasks.

[0138] This disclosure addresses the problem of multi-source heterogeneous data fusion in distributed photovoltaic power output prediction, proposing a complete spatiotemporal fusion technology system: First, it constructs a four-layer fusion framework, realizing full-process control from data access, preprocessing, fusion core to quality assessment, providing a systematic method for multi-source data fusion; Second, it proposes an adaptive time alignment strategy and multi-resolution spatial registration technology to effectively solve the problem of differences in time and spatial scales of multi-source data.

[0139] This disclosure proposes a complete spatiotemporal fusion technology system by focusing on the fusion processing of multi-source heterogeneous data: constructing a multi-source heterogeneous data fusion framework for distributed photovoltaic scenarios; proposing an optimization strategy that combines adaptive time alignment and multi-resolution spatial registration; and designing a fusion quality evaluation system based on information entropy and physical constraints.

[0140] Figure 6 A structural diagram of a distributed photovoltaic power plant photovoltaic output prediction system based on multi-source heterogeneous feature data, according to an embodiment of the present disclosure, is shown.

[0141] like Figure 6 As shown, the prediction system 600 includes a data input module 610, a spatiotemporal alignment module 620, a feature fusion module 630, and a photovoltaic output prediction module 640.

[0142] The data input module 610 is configured to acquire a multi-source heterogeneous feature dataset collected by the target distributed photovoltaic power station over a historical period. The multi-source heterogeneous feature dataset includes multiple multi-dimensional feature sequences from independent data sources. Each multi-dimensional feature sequence includes one multi-dimensional feature data or multiple multi-dimensional feature data arranged in chronological order. The spatiotemporal alignment module 620 is configured to perform temporal alignment and spatial alignment processing on the multi-source heterogeneous feature dataset to obtain a spatiotemporally aligned multi-source heterogeneous feature dataset. Specifically, when the sampling frequency of the multidimensional feature sequence is lower than a preset reference frequency, the historical time period is divided into multiple interpolation time periods based on the solar altitude angle change rate and the data type of the multidimensional feature sequence. Based on the data type of the multidimensional feature data within each interpolation time period and its corresponding interpolation time period, a matching interpolation method is selected for resampling to align the corresponding multidimensional feature data to the reference time axis determined by the reference frequency. The feature fusion module 630 is configured to, for each moment on the reference time axis, input all multidimensional feature sequences in the spatially aligned multi-source heterogeneous feature dataset into the corresponding encoder to extract initial features; perform dimensionality reduction and feature mapping on each initial feature to obtain a feature vector with a specified dimension; concatenate all feature vectors with the specified dimension into a multimodal heterogeneous feature set; process the multimodal heterogeneous feature set using a self-attention mechanism to obtain attention output features; and perform feature fusion processing on the multimodal heterogeneous feature set and the attention output features to obtain a joint feature vector with the specified dimension at the corresponding moment. The photovoltaic power output prediction module 640 is configured to input the joint feature vector of each moment on the reference time axis into the machine learning model for training, adjust the parameters of the machine learning model according to the difference between the predicted power generation and the actual power generation at each moment until the training stopping condition is met, and obtain the photovoltaic power output prediction model; and use the photovoltaic power output prediction model to predict the photovoltaic power output of the target distributed photovoltaic power station.

[0143] According to embodiments of this disclosure, when the multidimensional feature sequence includes multiple multidimensional feature data, each multidimensional feature data is a multivariate temporal feature data of a single spatial location at the spatial resolution of the corresponding data source; When the multidimensional feature sequence includes a multidimensional feature data, the multidimensional feature data is multivariate static feature data of a single spatial location at the spatial resolution of the corresponding data source; The system further includes a data preprocessing module configured to: decouple the variable dimensions of each multidimensional feature sequence before performing the time alignment processing, to obtain multiple corresponding univariate feature sequences; wherein, when the multidimensional feature sequence includes multiple multidimensional feature data, the univariate feature sequence includes multiple univariate time-series feature data arranged in chronological order; when the multidimensional feature sequence includes one multidimensional feature data, the univariate feature sequence includes one univariate static feature data; the univariate feature sequence is provided with corresponding position coordinates and variable identifiers.

[0144] According to embodiments of this disclosure, the spatiotemporal alignment module includes: a time dimension alignment submodule; the time dimension alignment submodule is configured to: When the sampling frequency of the multidimensional feature sequence is lower than the reference frequency, interpolation processing is performed on the multiple univariate feature sequences corresponding to the multidimensional feature sequence to align the univariate feature sequence to the reference time axis. When the sampling frequency of the multidimensional feature sequence is higher than the reference frequency, the multiple univariate feature sequences corresponding to the multidimensional feature sequence are downsampled respectively to align the univariate feature sequences to the reference time axis.

[0145] According to embodiments of this disclosure, the system further includes: a data reconstruction module; the data reconstruction module is configured to: After completing the time alignment process, multiple time-aligned single-variable time series feature sequences belonging to the same data source are recombined by spatial and variable mapping according to the corresponding position coordinates and variable identifiers to obtain time-aligned multidimensional feature sequences. The time-aligned multidimensional feature sequence is aligned with the reference time axis in the time dimension, and the spatial and variable dimensions are consistent with the multidimensional feature sequence before time processing.

[0146] According to embodiments of this disclosure: the spatiotemporal alignment module 620 further includes: a spatial dimension alignment submodule; the spatial dimension alignment submodule is configured as follows: A unified fine grid is established for the target area, wherein the spatial resolution of the unified fine grid is higher than the spatial resolution of the data source corresponding to the multidimensional feature sequence, and the target area is the preset area range of the target distributed photovoltaic power station; For each moment on the reference time axis, the multidimensional feature sequence corresponding to each moment is mapped onto the unified fine grid to obtain the background value of each fine grid point; Obtain high-resolution digital elevation model grid data and land cover type grid data corresponding to each fine grid point in the unified fine grid; correct the background value of the corresponding grid point according to the digital elevation model grid data and the land cover type grid data to obtain a spatially aligned multidimensional feature sequence.

[0147] According to embodiments of this disclosure, the feature fusion module includes a one-dimensional convolutional encoder, a two-dimensional convolutional encoder, a three-dimensional convolutional encoder, and a fully connected encoder; the feature fusion module is further configured to: When the data structure of the multidimensional feature sequence is time-series data, a one-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is image data, a two-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is grid data, a three-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is scalar data, a fully connected encoder is used to extract the corresponding initial features.

[0148] In particular, according to embodiments of this disclosure, the methods described above can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program tangibly embodied on a machine-readable medium, the computer program containing program code for performing the methods described above. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium.

[0149] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0150] The units or modules described in the embodiments of this disclosure can be implemented in software or programmable hardware. The described units or modules can also be located in a processor, and the names of these units or modules do not necessarily constitute a limitation on the unit or module itself.

[0151] In another aspect, this disclosure also provides a computer-readable storage medium, which may be a computer-readable storage medium included in the electronic device or computer system described above; or it may be a standalone computer-readable storage medium not assembled into a device. The computer-readable storage medium stores one or more programs, which are used by one or more processors to perform the methods described in this disclosure.

[0152] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

Claims

1. A method for predicting photovoltaic power output of a distributed photovoltaic power station based on multi-source heterogeneous feature data, characterized in that, include: Obtain a multi-source heterogeneous feature dataset of the target distributed photovoltaic power station collected over a historical period. The multi-source heterogeneous feature dataset includes multiple multi-dimensional feature sequences from independent data sources. Each multi-dimensional feature sequence includes one multi-dimensional feature data or multiple multi-dimensional feature data arranged in chronological order. The multi-source heterogeneous feature dataset is subjected to time alignment and spatial alignment processing to obtain a spatiotemporally aligned multi-source heterogeneous feature dataset. Specifically, when the sampling frequency of the multi-dimensional feature sequence is lower than a preset reference frequency, the historical time period is divided into multiple interpolation time periods based on the rate of change of the solar altitude angle and the data type of the multi-dimensional feature sequence. Based on the data type of the multi-dimensional feature data within each interpolation time period and its corresponding interpolation time period, a matching interpolation method is selected for resampling to align the corresponding multi-dimensional feature data to the reference time axis determined by the reference frequency. For each moment on the reference time axis, all multidimensional feature sequences in the spatially aligned multi-source heterogeneous feature dataset are input into the corresponding encoder to extract initial features; each initial feature is subjected to dimensionality reduction and feature mapping to obtain a feature vector with a specified dimension; all feature vectors with the specified dimension are concatenated to form a multimodal heterogeneous feature set; the multimodal heterogeneous feature set is processed using a self-attention mechanism to obtain attention output features; the multimodal heterogeneous feature set and the attention output features are subjected to feature fusion processing to obtain a joint feature vector with the specified dimension at the corresponding moment. The joint feature vector at each moment on the reference time axis is input into the machine learning model for training. The parameters of the machine learning model are adjusted according to the difference between the predicted power generation and the actual power generation at each moment until the training stopping condition is met, thus obtaining the photovoltaic power output prediction model. The photovoltaic power output prediction model is then used to predict the photovoltaic power output of the target distributed photovoltaic power station.

2. The prediction method according to claim 1, characterized in that, The process of using a self-attention mechanism to process the multimodal heterogeneous feature set includes: The correlation weight matrix between each feature vector in the multimodal heterogeneous feature set is calculated using a self-attention mechanism; The multimodal heterogeneous feature set is updated by weighting based on the association weight matrix to obtain the attention output feature.

3. The prediction method according to claim 1, characterized in that, The feature fusion process of the multimodal heterogeneous feature set and the attention output feature includes: The multimodal heterogeneous feature set and the attention output feature are normalized after residual connection to obtain fused intermediate features; the fused intermediate features are input into a multilayer perceptron for mapping, and then the mean aggregation of each feature vector after mapping is performed to obtain the joint feature vector with the specified dimension at the corresponding time.

4. The method according to claim 1, characterized in that, The sampling frequency of the multidimensional feature sequence is divided into high frequency, medium frequency, or low frequency; Wherein, the high frequency, the medium frequency and the low frequency correspond to the first set sampling frequency range, the second set sampling frequency range and the third set sampling frequency range, respectively, and the lower limit of the first set sampling frequency range is greater than the upper limit of the second set sampling frequency range, and the lower limit of the second set sampling frequency range is greater than the upper limit of the third set sampling frequency range. The reference frequency is a preset frequency value within the first set sampling frequency range, used to establish the reference time axis.

5. The prediction method according to claim 4, characterized in that: When the multidimensional feature sequence includes multiple multidimensional feature data, each multidimensional feature data is a multivariable temporal feature data of a single spatial location under the spatial resolution of the corresponding data source. When the multidimensional feature sequence includes a multidimensional feature data, the multidimensional feature data is multivariate static feature data of a single spatial location at the spatial resolution of the corresponding data source; The method further includes: before performing the time alignment processing, decoupling the variable dimensions of each multidimensional feature sequence to obtain multiple corresponding univariate feature sequences; wherein, when the multidimensional feature sequence includes multiple multidimensional feature data, the univariate feature sequence includes multiple univariate time-series feature data arranged in chronological order; when the multidimensional feature sequence includes one multidimensional feature data, the univariate feature sequence includes one univariate static feature data; the univariate feature sequence is provided with corresponding position coordinates and variable identifiers.

6. The prediction method according to claim 5, characterized in that, Time alignment processing of the multi-source heterogeneous feature dataset includes: When the sampling frequency of the multidimensional feature sequence is lower than the reference frequency, interpolation processing is performed on the multiple univariate feature sequences corresponding to the multidimensional feature sequence to align the univariate feature sequence to the reference time axis. When the sampling frequency of the multidimensional feature sequence is higher than the reference frequency, the multiple univariate feature sequences corresponding to the multidimensional feature sequence are downsampled respectively to align the univariate feature sequences to the reference time axis.

7. The prediction method according to claim 6, characterized in that, The interpolation process performed on the multiple univariate feature sequences corresponding to the multidimensional feature sequence includes: When the univariate feature sequence includes multiple univariate time series feature data arranged in chronological order, a matching interpolation method is selected for resampling based on the data type of the univariate time series feature data in each interpolation time period and the interpolation time period to which it belongs. When the univariate feature sequence includes a univariate static feature data, the univariate static feature data is copied along the reference time axis to generate a time-series feature sequence covering each moment of the reference time axis.

8. The prediction method according to claim 7, characterized in that, The univariate time-series feature data includes any of the following data types: radiation quantity data, environmental state quantity data, cumulative data, and vector data; The step of selecting a matching interpolation method for resampling based on the data type and interpolation time period of the univariate time series feature data within each interpolation time period includes: When the time period to be interpolated is a nighttime period, for the univariate time series feature data within the nighttime period, cubic spline interpolation is performed when the data type of the univariate time series feature data is environmental state data, and forced zeroing is performed when the data type of the univariate time series feature data is radiation data. When the time period to be interpolated is a stable daytime period, cubic spline interpolation is performed on the univariate time series feature data within the stable daytime period when the data type of the univariate time series feature data is radiation data or environmental state data. When the data type of the univariate time series feature data is cumulative, monotonic interpolation processing is performed. When the data type of the univariate time series feature data is vector data, component-independent cubic spline interpolation is performed.

9. The prediction method according to claim 5, characterized in that, The method further includes: When missing data is detected in the univariate feature sequence, for each missing feature data in the univariate feature sequence, the following is executed: Multiple candidate auxiliary data sources are obtained, and the comprehensive spatiotemporal correlation coefficient between each candidate auxiliary data source and the missing feature data is calculated; a threshold is set based on the comprehensive spatiotemporal correlation coefficient. The multiple candidate auxiliary data sources are filtered according to the threshold to obtain one or more reliable auxiliary sources; The observation data of each trusted auxiliary source at the corresponding target time are obtained, and the observation data are converted into estimated values ​​with the same dimensions as the missing feature data through the corresponding physical inversion model, wherein the target time is the sampling time of the missing feature data; The corresponding estimates are weighted and fused according to the normalized weights of each reliable auxiliary source to generate the imputation values ​​for the missing feature data.

10. The prediction method according to claim 5, characterized in that, The method further includes: After completing the time alignment process, multiple time-aligned single-variable time series feature sequences belonging to the same data source are recombined by spatial and variable mapping according to the corresponding position coordinates and variable identifiers to obtain time-aligned multidimensional feature sequences. The time-aligned multidimensional feature sequence is aligned with the reference time axis in the time dimension, and the spatial and variable dimensions are consistent with the multidimensional feature sequence before time processing.

11. The prediction method according to claim 10, characterized in that, The method further includes: After completing the time alignment process, the time-aligned multidimensional feature sequence is subjected to spatial alignment processing, which includes: A unified fine grid is established for the target area, wherein the spatial resolution of the unified fine grid is higher than the spatial resolution of the data source corresponding to the multidimensional feature sequence, and the target area is the preset area range of the target distributed photovoltaic power station; For each moment on the reference time axis, the multidimensional feature sequence corresponding to each moment is mapped onto the unified fine grid to obtain the background value of each fine grid point; Obtain high-resolution digital elevation model grid data and land cover type grid data corresponding to each fine grid point in the unified fine grid; correct the background value of the corresponding grid point according to the digital elevation model grid data and the land cover type grid data to obtain a spatially aligned multidimensional feature sequence.

12. The prediction method according to claim 1, characterized in that, For each moment on the reference time axis, the entire multidimensional feature sequence from the spatially aligned multi-source heterogeneous feature dataset is input into the corresponding encoder to extract initial features, including: When the data structure of the multidimensional feature sequence is time-series data, a one-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is image data, a two-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is grid data, a three-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is scalar data, a fully connected encoder is used to extract the corresponding initial features.

13. A distributed photovoltaic power output prediction system based on multi-source heterogeneous characteristic data, characterized in that, include: The data input module is configured to acquire a multi-source heterogeneous feature dataset collected from the target distributed photovoltaic power station over a historical period. The multi-source heterogeneous feature dataset includes multiple multi-dimensional feature sequences from independent data sources. Each multi-dimensional feature sequence includes one multi-dimensional feature data or multiple multi-dimensional feature data arranged in chronological order. The spatiotemporal alignment module is configured to perform temporal and spatial alignment processing on the multi-source heterogeneous feature dataset to obtain a spatiotemporally aligned multi-source heterogeneous feature dataset. Specifically, when the sampling frequency of the multidimensional feature sequence is lower than a preset reference frequency, the historical time period is divided into multiple interpolation time periods based on the solar altitude angle change rate and the data type of the multidimensional feature sequence. Based on the data type of the multidimensional feature data within each interpolation time period and its corresponding interpolation time period, a matching interpolation method is selected for resampling to align the corresponding multidimensional feature data to a reference time axis determined by the reference frequency. The feature fusion module is configured to, for each moment on the reference time axis, input all multidimensional feature sequences from the spatially aligned multi-source heterogeneous feature dataset into the corresponding encoder to extract initial features; perform dimensionality reduction and feature mapping on each initial feature to obtain a feature vector with a specified dimension; concatenate all feature vectors with the specified dimension into a multimodal heterogeneous feature set; process the multimodal heterogeneous feature set using a self-attention mechanism to obtain attention output features; and perform feature fusion processing on the multimodal heterogeneous feature set and the attention output features to obtain a joint feature vector with the specified dimension at the corresponding moment. The photovoltaic power output prediction module is configured to input the joint feature vector of each moment on the reference time axis into the machine learning model for training, adjust the parameters of the machine learning model according to the difference between the predicted power generation and the actual power generation at each moment until the training stopping condition is met, and obtain the photovoltaic power output prediction model; and use the photovoltaic power output prediction model to predict the photovoltaic power output of the target distributed photovoltaic power station.

14. The prediction system according to claim 13, characterized in that: When the multidimensional feature sequence includes multiple multidimensional feature data, each multidimensional feature data is a multivariable temporal feature data of a single spatial location under the spatial resolution of the corresponding data source. When the multidimensional feature sequence includes a multidimensional feature data, the multidimensional feature data is multivariate static feature data of a single spatial location at the spatial resolution of the corresponding data source; The system further includes a data preprocessing module configured to: decouple the variable dimensions of each multidimensional feature sequence before performing the time alignment processing, to obtain multiple corresponding univariate feature sequences; wherein, when the multidimensional feature sequence includes multiple multidimensional feature data, the univariate feature sequence includes multiple univariate time-series feature data arranged in chronological order; when the multidimensional feature sequence includes one multidimensional feature data, the univariate feature sequence includes one univariate static feature data; the univariate feature sequence is provided with corresponding position coordinates and variable identifiers.

15. The prediction system according to claim 14, characterized in that, The spatiotemporal alignment module includes a time dimension alignment submodule; the time dimension alignment submodule is configured as follows: When the sampling frequency of the multidimensional feature sequence is lower than the reference frequency, interpolation processing is performed on the multiple univariate feature sequences corresponding to the multidimensional feature sequence to align the univariate feature sequence to the reference time axis. When the sampling frequency of the multidimensional feature sequence is higher than the reference frequency, the multiple univariate feature sequences corresponding to the multidimensional feature sequence are downsampled respectively to align the univariate feature sequences to the reference time axis.

16. The prediction system according to claim 14, characterized in that, The system further includes: a data reconstruction module; the data reconstruction module is configured to: After completing the time alignment process, multiple time-aligned single-variable time series feature sequences belonging to the same data source are recombined by spatial and variable mapping according to the corresponding position coordinates and variable identifiers to obtain time-aligned multidimensional feature sequences. The time-aligned multidimensional feature sequence is aligned with the reference time axis in the time dimension, and the spatial and variable dimensions are consistent with the multidimensional feature sequence before time processing.

17. The prediction system according to claim 14, characterized in that, include: The spatiotemporal alignment module further includes a spatial dimension alignment submodule; the spatial dimension alignment submodule is configured as follows: A unified fine grid is established for the target area, wherein the spatial resolution of the unified fine grid is higher than the spatial resolution of the data source corresponding to the multidimensional feature sequence, and the target area is the preset area range of the target distributed photovoltaic power station; For each moment on the reference time axis, the multidimensional feature sequence corresponding to each moment is mapped onto the unified fine grid to obtain the background value of each fine grid point; Obtain high-resolution digital elevation model grid data and land cover type grid data corresponding to each fine grid point in the unified fine grid; correct the background value of the corresponding grid point according to the digital elevation model grid data and the land cover type grid data to obtain a spatially aligned multidimensional feature sequence.

18. The prediction system according to claim 14, characterized in that, The feature fusion module includes a one-dimensional convolutional encoder, a two-dimensional convolutional encoder, a three-dimensional convolutional encoder, and a fully connected encoder; the feature fusion module is also configured to: When the data structure of the multidimensional feature sequence is time-series data, a one-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is image data, a two-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is grid data, a three-dimensional convolutional encoder is used to extract the corresponding initial features; when the data structure of the multidimensional feature sequence is scalar data, a fully connected encoder is used to extract the corresponding initial features.

19. A power dispatching system, characterized in that, The system includes a distributed photovoltaic power output prediction system as described in any one of claims 13 to 18; or it includes a memory and a processor; wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the prediction method as described in any one of claims 1 to 12.

20. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the computer instructions are executed by the processor, they implement the prediction method as described in any one of claims 1 to 12.

Images