Distributed photovoltaic-oriented intelligent power grid load prediction method and system

Abstract

The invention relates to the technical field of smart power grids, in particular to a distributed photovoltaic-oriented smart power grid load prediction method and system, and the method comprises the steps: collecting the total load data, distributed photovoltaic power generation power data and meteorological data of a target region and an adjacent region, and carrying out the preprocessing; extracting spatial correlation characteristics and time sequence evolution characteristics of the photovoltaic power generation power data, and fusing to generate a low-dimension spatial-temporal characteristic matrix; splicing the spatial-temporal feature matrix with the total load data and the meteorological data to form a multi-modal fusion feature vector, and inputting the multi-modal fusion feature vector into a machine learning prediction model to generate an initial load prediction result; and introducing real-time updated ultra-short-term weather forecast data, carrying out dynamic correction on an initial prediction result through a linear regression error correction model, and outputting a final load prediction value. According to the method, the space-time precision and the real-time adaptive capacity of load prediction in a distributed photovoltaic scene can be remarkably improved, and the method has good practicability and engineering generalizability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart grid, in particular to a smart grid load prediction method and system for distributed photovoltaic. BACKGROUND

[0002] As an important part of clean energy, distributed photovoltaic power generation is connected to the distribution network system in a high-density and high-penetration form. Due to the significant intermittency and volatility of distributed photovoltaic output, the power grid load presents complex dynamic variation characteristics in time and space. In order to ensure the safe and stable operation of the local power grid, it is urgent to establish a power grid load prediction method with time and space resolution, sensitivity to photovoltaic disturbance, and fast response speed, to provide reliable data support for distribution network dispatching optimization, source-grid-load collaborative control and auxiliary service strategy.

[0003] However, the existing power grid load prediction technology generally has the following technical bottlenecks: first, the data acquisition range is limited, and most of the target area single-point or single-source data is used as the basis, without effectively fusing the adjacent regional collaborative information and multi-source weather data, resulting in insufficient input feature dimension of the model and low prediction accuracy; second, the feature modeling means is still mainly based on static statistical method or simple recursive neural network, which is difficult to capture the spatio-temporal coupling evolution law of photovoltaic power; third, most models cannot cope with sudden weather changes, lack of error correction mechanism based on real-time weather update, resulting in lag of prediction results and serious error accumulation, which is difficult to meet the real-time and sensitivity requirements under the background of high proportion of new energy grid connection. SUMMARY

[0004] The present application provides a smart grid load prediction method and system for distributed photovoltaic, which fuses multi-regional and multi-modal features and has real-time dynamic correction capability, to improve the prediction accuracy and application effectiveness in complex scenarios.

[0005] The smart grid load prediction method for distributed photovoltaic includes the following steps: S1: Collecting the total load data, distributed photovoltaic power generation data of the target area and at least two adjacent areas in the historical same period and the current period, and the corresponding spatio-temporal scale weather data; S2: Based on the distributed photovoltaic power generation data collected in S1, extracting the spatial correlation features between different regions and the time sequence evolution features at different time points, and generating a spatio-temporal feature matrix; S3: Feature fusion of the spatio-temporal feature matrix with the total load data in the historical same period and the current period, and the weather data, and input into a pre-trained machine learning prediction model to generate an initial load prediction result of the target area; S4: Obtain real-time updated ultra-short-term weather forecast data for the target area, dynamically correct the initial load forecast result based on the real-time updated ultra-short-term weather forecast data, and output the final load forecast value for the target area.

[0006] Optionally, S1 includes: S11: Determine the target area and at least two adjacent areas to form an area set, and obtain the geographic location identification information of the area set; S12: Based on the geographic location identification information of the region set, collect the total load data of the region set in the same historical period and the current period from the power grid data platform, collect the distributed photovoltaic power generation data of the region set in the same historical period and the current period from the distributed photovoltaic monitoring system, and collect meteorological data covering the region set at the corresponding spatiotemporal scale from the meteorological data service interface. S13: Perform data cleaning and outlier processing on the collected total load data, distributed photovoltaic power generation data and meteorological data respectively to obtain the cleaned total load data, distributed photovoltaic power generation data and meteorological data; S14: Align and match the cleaned total load data, distributed photovoltaic power generation data, and meteorological data according to timestamps and regional identifiers to generate historical total load data, distributed photovoltaic power generation data, and meteorological data for the same period and current time in a standardized format.

[0007] Optionally, the data cleaning and outlier handling in S13 specifically involves: using the sliding time window standard deviation method to identify and remove outliers in the total load data, the distributed photovoltaic power generation data, and the meteorological data, and using linear interpolation to fill in missing values.

[0008] Optionally, S2 includes: S21: Based on the standardized format of the distributed photovoltaic power generation data collected in S1, calculate the Pearson correlation coefficient of the time series of distributed photovoltaic power generation data of the target area and each adjacent area within a preset time window, and generate spatial correlation features that characterize the spatial dependence between regions. S22: Based on the standardized format of distributed photovoltaic power generation data collected in S1, a time convolutional network is used to extract the long-term trend features and short-term fluctuation features of the time series of distributed photovoltaic power generation data in each region, and generate time series evolution features that characterize the time evolution law of each region. S23: The spatial correlation features generated in S21 and the temporal evolution features generated in S22 are spliced ​​and fused, and the high-dimensional features after fusion are reduced in dimensionality by principal component analysis algorithm to generate a low-dimensional spatiotemporal feature matrix.

[0009] Optionally, in step S23, the high-dimensional features after fusion are reduced using principal component analysis (PCA) algorithm. Specifically, the spatial correlation features and the temporal evolution features are concatenated along the feature dimension to form a high-dimensional fused feature vector. PCA algorithm is then used to reduce the dimension of the high-dimensional fused feature vector to a preset dimension while maintaining more than 95% of the original feature variance, thereby generating the low-dimensional spatiotemporal feature matrix.

[0010] Optionally, S3 includes: S31: Align the spatiotemporal feature matrix, the total load data of the historical period and the current period in the standardized format, and the meteorological data in the standardized format according to the time series, and perform splicing operation along the feature dimension to generate a multimodal fusion feature vector; S32: The multimodal fusion feature vector is input into a pre-trained machine learning prediction model. The pre-trained machine learning prediction model performs nonlinear transformation and feature extraction on the multimodal fusion feature vector through an internal multi-layer neural network structure to generate a load prediction feature representation for a specific future time period. S33: The pre-trained machine learning prediction model performs a linear transformation on the load prediction feature representation through an internal fully connected layer to map the feature dimension to the prediction output dimension, generating the initial load prediction result for the target region.

[0011] Optionally, the splicing operation along the feature dimension in S31 specifically involves connecting the feature vectors corresponding to the spatiotemporal feature matrix, the total load data of the historical period and the current period in the standardized format, and the meteorological data in the standardized format along the feature dimension to form a joint feature vector, namely the multimodal fusion feature vector.

[0012] Optionally, S4 includes: S41: Acquire real-time updated ultra-short-term weather forecast data of the target area in real time, and compare the real-time updated ultra-short-term weather forecast data with the meteorological data collected in S1 for generating the initial load prediction results, and calculate the meteorological data deviation value. S42: Input the meteorological data deviation value into the pre-trained error correction model, and the pre-trained error correction model outputs a prediction error compensation value corresponding to the initial load prediction result based on the meteorological data deviation value. S43: The initial load prediction result generated in S3 is algebraically added to the prediction error compensation value obtained in S42 to complete the dynamic correction of the initial load prediction result, and the final target area load prediction value is output.

[0013] Optionally, the algebraic addition in S43 specifically involves adding the initial load prediction result and the prediction error compensation value according to a one-to-one correspondence at each time point.

[0014] A smart grid load forecasting system for distributed photovoltaic (PV) power, used to implement the aforementioned smart grid load forecasting method for distributed PV power, includes the following modules: Data acquisition module: used to collect total load data, distributed photovoltaic power generation data and corresponding spatiotemporal meteorological data of the target area and at least two adjacent areas during the same period in history and the current period, and to clean, align and standardize the data to generate a dataset in a standardized format; Feature extraction module: used to extract spatial correlation features between regions and temporal evolution features of each region based on the distributed photovoltaic power generation data in the standardized format, and to fuse and reduce the dimensionality of the features to generate a spatiotemporal feature matrix; Predictive modeling module: This module is used to fuse the spatiotemporal feature matrix with the total load data of the same period in the historical time and the current time in the standardized format, as well as the meteorological data in the standardized format, and input the data into the pre-trained machine learning prediction model to generate the initial load prediction results for the target area in the future for a specific time period. Dynamic correction module: It is used to receive real-time updated ultra-short-term weather forecast data of the target area, calculate the meteorological data deviation value, generate the prediction error compensation value through the error correction model, perform algebraic correction on the initial load prediction result, and output the final target area load prediction value.

[0015] The beneficial effects of this invention are: This invention, by introducing a cross-regional collaborative data acquisition mechanism, comprehensively acquires total load data, distributed photovoltaic power generation data, and meteorological data of the target area and its adjacent areas, and uniformly performs standardized processing, effectively solving the problems of heterogeneous data sources, asynchronous time, and spatial discontinuity. Furthermore, it integrates inter-regional Pearson spatial correlation features with temporal evolution features extracted by temporal convolutional networks within the region to construct a multi-dimensional low-dimensional spatiotemporal feature matrix. This enables the model to more accurately capture the regional propagation effect and dynamic evolution law of photovoltaic output, thereby significantly improving the load prediction accuracy and robustness in complex multi-source scenarios.

[0016] This invention designs a multimodal joint modeling framework based on the fusion of spatiotemporal feature matrices, historical load sequences, and refined meteorological data. By inputting the fused features into a gated recurrent unit network model based on an attention mechanism, it achieves dynamic weighting of the model's contribution to the input features while maintaining temporal dependency modeling capabilities. This structure enables the model to possess stronger nonlinear mapping capabilities, allowing it to adaptively learn the combined effects of different meteorological conditions, historical operating conditions, and photovoltaic power disturbances on load fluctuations. This effectively alleviates the problem of traditional models' inaccurate predictions of abnormal weather, sudden changes in illumination, or complex terrain areas.

[0017] This invention introduces an error compensation mechanism based on real-time updated ultra-short-term weather forecast data. By calculating the meteorological data deviation and inputting it into a lightweight linear regression error correction model, the corresponding prediction error compensation value is output, achieving point-by-point dynamic correction of the initial load forecast results. This correction mechanism can rapidly adjust the forecast output based on minute-level meteorological changes, possessing the advantages of low latency and high sensitivity. It significantly improves the real-time performance and accuracy of the forecast system under short-term disturbances such as rapid fluctuations in photovoltaic power and sudden changes in cloud cover, meeting the technical requirements for refined and flexible grid control in the context of high distributed power penetration. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the system flow according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0021] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0022] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0023] like Figure 1 As shown, the smart grid load forecasting method for distributed photovoltaic power includes the following steps: S1: Collect total load data, distributed photovoltaic power generation data, and corresponding spatiotemporal meteorological data for the target area and at least two adjacent areas during the same historical period and the current time period. Specifically: S11: Determine the target area and at least two adjacent areas to form a set of areas, and obtain the geographic location identification information of the set of areas.

[0024] First, based on grid dispatch requirements and distributed photovoltaic deployment density, target areas requiring load forecasting are identified. Then, based on the grid topology, at least two areas directly connected to or adjacent to these target areas are selected to construct a region set. Subsequently, for each region in the region set, its corresponding geographic location identifier is extracted. This identifier consists of the latitude and longitude coordinates or standard geographic codes of each region. This information is retrieved through a Geographic Information System (GIS) platform and stored in a structured manner within the system for subsequent data retrieval and region attribution determination.

[0025] S12: Based on the geographic location identification information of the regional set, collect the total load data of the regional set in the same historical period and the current period from the power grid data platform, collect the distributed photovoltaic power generation data of the regional set in the same historical period and the current period from the distributed photovoltaic monitoring system, and collect the meteorological data of the covered regional set at the corresponding spatiotemporal scale from the meteorological data service interface.

[0026] Based on the geographic location information obtained from S11, the interface of the distribution master station data platform or the electricity consumption information collection system platform is called to collect the total load data of the regional set within the same historical period (e.g., the same season and time period in the past 5 years) and the current time period (e.g., the most recent 24 hours or the most recent 7 days), ensuring that the data granularity is at least 15 minutes. Simultaneously, distributed photovoltaic power generation data of the regional set within the same time range is retrieved from the regional distributed photovoltaic monitoring system (e.g., photovoltaic power plant SCADA system or remote monitoring platform), recording the power generation per unit area and real-time output power. Furthermore, the application programming interface for meteorological data services provided by the meteorological center or commercial meteorological service provider is called, using the geographic location identifier as input parameter, to request and obtain meteorological data covering the geographical range of the regional set, with kilometer-level spatial resolution and minute-level temporal resolution. The meteorological data includes irradiance, ambient temperature, and cloud cover, and is uniformly stored in the meteorological data cache database.

[0027] S13: Perform data cleaning and outlier processing on the collected historical and current total load data, distributed photovoltaic power generation data, and meteorological data to obtain the cleaned historical and current total load data, distributed photovoltaic power generation data, and meteorological data.

[0028] A unified data preprocessing workflow is applied to all raw datasets acquired through S12. First, outliers are identified using a sliding time window standard deviation method. Specifically, a fixed-length (e.g., 2-hour) sliding time window is constructed centered on each data record. The standard deviation of all data points within the window is calculated. If the value of the current data point exceeds the mean ± three times the standard deviation within the window, it is marked as an outlier and removed. Next, missing data points are filled using linear interpolation. Specifically, based on two adjacent valid data points in the time series, the estimated value of the missing point is linearly extrapolated. This processing workflow is independently applied to the total load data, distributed photovoltaic power generation data, and meteorological data for the same historical period and the current time period, ensuring consistency, continuity, and statistical validity for all types of data. The cleaned total load data, distributed photovoltaic power generation data, and meteorological data for the same historical period and the current time period are output after processing.

[0029] S14: Align and match the cleaned historical total load data, distributed photovoltaic power generation data, and meteorological data for the same period and the current period according to the timestamp and regional identifier to generate standardized format historical total load data, distributed photovoltaic power generation data, and meteorological data for the same period and the current period.

[0030] The cleaned data sources from different platforms are integrated in a unified temporal and spatial manner. First, the timestamps of all data records are uniformly converted to Coordinated Universal Time (UTC) to avoid record misalignment caused by local time differences between different platforms. Then, based on the regional code or latitude and longitude information carried by each data point, and according to the preset geographical location identifiers in the regional set, all data are matched by a secondary index of timestamp and region to achieve cross-data source alignment. After alignment and matching, the data records under each timestamp and region combination are normalized. Specifically, the total load value, photovoltaic power generation value, and meteorological variables are respectively normalized to the minimum-maximum range according to their historical range, and standardized to the [0,1] interval. Finally, all normalized data are encapsulated into standardized data tables with unified time series indexes and regional labels, and output as standardized formats of historical total load data, distributed photovoltaic power generation data, and meteorological data for the current period, providing a consistent data input basis for subsequent model training and prediction.

[0031] S2: Based on the distributed photovoltaic power generation data collected in S1, its spatial correlation characteristics across different regions and temporal evolution characteristics at different time points are extracted to generate a spatiotemporal feature matrix, specifically: S21: Based on the standardized format of distributed photovoltaic power generation data collected by S1, calculate the Pearson correlation coefficient of the time series of distributed photovoltaic power generation data of the target area and each adjacent area within a preset time window, and generate spatial correlation features that characterize the spatial dependence between regions.

[0032] The system reads the standardized distributed photovoltaic (PV) power data output by S1. Based on the region set determined by S11, the target region is paired with each of its adjacent regions. A preset time window of 24 hours is set, and with a 1-hour time granularity, the distributed PV power time series of the target region and each adjacent region are paired over 24 consecutive time points. Subsequently, the Pearson correlation coefficient formula is used to perform correlation analysis on the time series of each pair of regions, obtaining a correlation coefficient value representing the synchronicity of power changes between the target region and its adjacent regions. The Pearson correlation coefficient values ​​of all region pairs within this time window constitute a set of spatial correlation feature values, representing the spatial dependency between the target region and its adjacent regions during this period. This set of spatial correlation features serves as one of the important input features for subsequent modeling.

[0033] S22: Simultaneously, based on the standardized format of distributed photovoltaic power generation data collected in S1, a temporal convolutional network is used to extract the long-term trend characteristics and short-term fluctuation characteristics of the time series of distributed photovoltaic power generation data in each region, generating temporal evolution characteristics that characterize the temporal evolution pattern of each region.

[0034] Standardized distributed photovoltaic (PV) power data is read, and independent time-series input tensors are constructed for each region according to regional division. A temporal convolutional network (using causal convolution and dilated convolution mechanisms, a convolutional neural network architecture specifically designed for processing time-series data, effectively capturing long-term dependencies) is constructed. The network structure has four layers, with each layer employing a convolution operation with a kernel size of 3. The input time-series length is set to 24 (corresponding to 24 hours). ReLU activation and BatchNorm regularization are performed after each convolution layer to enhance feature representation and avoid gradient vanishing. The feature vectors output from the shallow layers (layers 1 and 2) are used to capture high-frequency changes, defined as short-term fluctuation features; the feature vectors output from the deep layers (layers 3 and 4) are used to capture low-frequency changes, defined as long-term trend features. The long-term trend features and short-term fluctuation features of each region are concatenated to form a complete temporal evolution feature set, reflecting the PV power variation patterns within each region.

[0035] S23: The spatial correlation features generated in S21 and the temporal evolution features generated in S22 are spliced ​​and fused. The high-dimensional features after fusion are then reduced in dimensionality using the principal component analysis algorithm to generate a unified low-dimensional spatiotemporal feature matrix.

[0036] The spatial correlation features output by S21 and the temporal evolution features output by S22 are concatenated according to feature dimensions to form a unified high-dimensional fused feature vector. This vector contains spatial dependency information between the target region and adjacent regions, as well as the temporal evolution patterns of each region, and has a strong spatiotemporal comprehensive expressive capability. To reduce the computational complexity of the model and minimize feature redundancy, principal component analysis (PCA) is used to reduce the dimensionality of the high-dimensional fused feature vector. Specifically, the feature covariance matrix of the fused feature vector is solved, the principal component directions are calculated, and the top principal components are selected according to their contribution rates, ensuring that the retained feature dimensions cumulatively explain more than 95% of the variance. The result after dimensionality reduction is the low-dimensional spatiotemporal feature matrix, which serves as one of the core inputs to the subsequent machine learning prediction model.

[0037] S3: The spatiotemporal feature matrix is ​​fused with historical and current total load data and meteorological data, and then input into a pre-trained machine learning prediction model to generate the initial load prediction results for the target area. Specifically: S31: Align the spatiotemporal feature matrix, the total load data of the same period in the standardized format and the current period, and the refined meteorological data in the standardized format according to the time series, and perform splicing operation along the feature dimension to generate a multimodal fusion feature vector.

[0038] The low-dimensional spatiotemporal feature matrix output by S2, along with the standardized historical and current total load data and meteorological data generated by S1, are obtained. During data alignment, a unified UTC timestamp is used as an index to strictly synchronize the three types of feature data in time series, ensuring that the feature values ​​corresponding to each time point exist in all three data sources. After time alignment, feature vectors for the three features at each time point are extracted: the spatiotemporal feature vector of the spatiotemporal feature matrix, the load feature vector of the historical and current total load data, and the meteorological feature vector of the meteorological data. These three feature vectors are concatenated according to their feature dimensions (i.e., the horizontal dimension of the vector) to form a joint feature vector containing all three types of information, defined as the multimodal fusion feature vector. This fusion feature vector effectively integrates the spatiotemporal evolution pattern, historical load trajectory, and current meteorological environment, providing full information input for subsequent prediction models.

[0039] S32: Input the multimodal fusion feature vector into the pre-trained machine learning prediction model. The pre-trained machine learning prediction model performs nonlinear transformation and feature extraction on the multimodal fusion feature vector through its internal multi-layer neural network structure to generate a load prediction feature representation for a specific future time period.

[0040] The multimodal fusion feature vector output from S31 is input into a pre-trained machine learning prediction model, which is an attention-based gated recurrent unit (GRU) network model. The main structure of this model consists of multiple stacked GRU modules, capable of capturing the sequential dependencies of the fusion feature vector over time. An attention mechanism is integrated into the middle layer of the network to dynamically calculate the relative importance of different modalities (i.e., spatiotemporal feature matrix, historical and current total load data, and meteorological data) in the multimodal fusion feature vector to the current output at each time step, and assigns weights accordingly to improve the responsiveness to key features. Through recursive processing of multiple GRU units in the network and weighted summation using the attention mechanism, a load prediction feature representation for a specific future time period is extracted. This specific future time period is 4 to 24 hours from now. This feature representation retains a complex cross-time and cross-modal relationship structure, forming the core foundation for subsequent prediction output generation.

[0041] S33: The pre-trained machine learning prediction model performs a linear transformation on the load prediction feature representation through an internal fully connected layer to map the feature dimension to the prediction output dimension, generating the initial load prediction result for the target region.

[0042] The load prediction feature representation generated by S32 is input to the output mapping module at the end of the machine learning prediction model. This module is a fully connected layer with a fixed structure. The fully connected layer performs a linear mapping operation, mapping the feature dimension of the load prediction feature representation to the prediction output dimension. Specifically, if the prediction time length is 4 hours, 8 hours, 12 hours, or 24 hours, the output dimension is set to 4, 8, 12, or 24, corresponding to one prediction point per hour. The weight parameters of the fully connected layer have been learned through historical data during the pre-training stage, ensuring that the potential high-order feature representation is transformed into load values ​​with actual physical meaning. Finally, the output prediction value sequence constitutes the initial load prediction result for the target area in a specific future time period. This result will be used by the subsequent real-time correction module for optimization, further improving prediction accuracy and practicality.

[0043] S4: Obtain real-time updated ultra-short-term weather forecast data for the target area, dynamically correct the initial load forecast results based on the real-time updated ultra-short-term weather forecast data, and output the final load forecast value for the target area. Specifically: S41: Acquire real-time updated ultra-short-term weather forecast data for the target area, and compare the real-time updated ultra-short-term weather forecast data with the refined meteorological data collected by S1 to generate the initial load forecast results, and calculate the meteorological data deviation value.

[0044] By connecting to a high-frequency weather forecast service interface, real-time updated ultra-short-term weather forecast data covering the target area is acquired. This data has a timescale of 15 to 60 minutes and a temporal resolution better than 5 minutes, mainly containing predicted values ​​for irradiance and cloud cover. This data is typically generated by fusing ground-based lidar observations, satellite cloud imagery tracking, and numerical forecast models, ensuring minute-level responsiveness and kilometer-level spatial accuracy. Subsequently, the aforementioned real-time updated ultra-short-term weather forecast data is matched with the standardized meteorological data collected in S1. The matching rule is as follows: using the same predicted future time point as a reference, irradiance and cloud cover values ​​from both meteorological data sources are extracted at each time point, and the numerical difference between the two is calculated point by point. This difference is the meteorological data deviation value, used to reflect the degree of deviation between the real-time weather conditions and the original predicted input data of the model. All deviation values ​​are combined in chronological order to form a complete meteorological data deviation value sequence, providing a direct basis for the error compensation mechanism.

[0045] S42: Input the meteorological data deviation value into the pre-trained error correction model. The pre-trained error correction model outputs a prediction error compensation value corresponding to the initial load prediction result based on the meteorological data deviation value.

[0046] The meteorological data deviation sequence obtained in S41 is used as the input vector and fed into a pre-trained error correction model. This model is a lightweight linear regression model specifically designed for rapid online correction. The pre-trained error correction model has been trained based on a large amount of historical sample data. The training samples use "meteorological data deviation values ​​(including historical irradiance and cloud cover deviations)" as input features and "historical prediction errors between the initial load forecast and the actual load value" as the supervision objective. The model parameters are learned by minimizing the mean squared error loss function. In the current operational phase, this error correction model maps the real-time input meteorological data deviation values ​​to a corresponding sequence of prediction error compensation values. The temporal distribution of the prediction error compensation values ​​corresponds one-to-one with the initial load forecast results, ensuring that subsequent correction operations can be performed directly.

[0047] S43: The initial load forecast result generated by S3 is algebraically added to the forecast error compensation value obtained by S42 to complete the dynamic correction of the initial load forecast result and output the final target area load forecast value.

[0048] The initial load forecast result generated in S3 is called and matched one-to-one with the forecast error compensation value sequence output by S42 in the time dimension. For each future time point, the initial forecast value and the corresponding error compensation value are algebraically added to obtain the final forecast value for that time point. This addition operation is expressed as: ; in, Indicates the first in the predicted sequence The addition operation is performed synchronously at all time points, thus completing a full correction of the initial prediction result.

[0049] The revised forecast results constitute the final load forecast value for the target area. This value improves the real-time performance and accuracy of the forecast by taking into account the latest meteorological changes, and can be directly used for smart grid dispatching, real-time load balance optimization, and auxiliary energy management decision-making.

[0050] like Figure 2 As shown, the smart grid load forecasting system for distributed photovoltaic (PV) power generation, used to implement the aforementioned smart grid load forecasting method for distributed PV power generation, includes the following modules: Data acquisition module: used to collect total load data, distributed photovoltaic power generation data and corresponding spatiotemporal meteorological data of the target area and at least two adjacent areas during the same period in history and the current period, and to clean, align and standardize the data to generate a dataset in a standardized format; Feature extraction module: used to extract spatial correlation features between regions and temporal evolution features of each region based on distributed photovoltaic power generation data in a standardized format, and to fuse and reduce the dimensionality of the features to generate a spatiotemporal feature matrix; Predictive modeling module: This module is used to fuse the spatiotemporal feature matrix with the total load data of the same period in the historical time and the current time in the standardized format, as well as the meteorological data in the standardized format, and input them into the pre-trained machine learning prediction model to generate the initial load prediction results for the target area in the future for a specific time period. Dynamic correction module: It is used to receive real-time updated ultra-short-term weather forecast data of the target area, calculate the meteorological data deviation value, generate the prediction error compensation value through the error correction model, perform algebraic correction on the initial load prediction results, and output the final target area load prediction value.

[0051] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0052] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A smart grid load forecasting method for distributed photovoltaic power, characterized in that, Includes the following steps: S1: Collect total load data, distributed photovoltaic power generation data, and meteorological data at the corresponding spatiotemporal scales for the target area and at least two adjacent areas during the same historical period and the current period. S2: Based on the distributed photovoltaic power generation data collected in S1, extract its spatial correlation characteristics in different regions and temporal evolution characteristics at different time points to generate a spatiotemporal feature matrix; S3: The spatiotemporal feature matrix is ​​fused with the total load data of the same historical period and the current period and the meteorological data, and then input into the pre-trained machine learning prediction model to generate the initial load prediction result for the target area; S4: Obtain real-time updated ultra-short-term weather forecast data for the target area, dynamically correct the initial load forecast result based on the real-time updated ultra-short-term weather forecast data, and output the final load forecast value for the target area.

2. The smart grid load forecasting method for distributed photovoltaic power according to claim 1, characterized in that, S1 includes: S11: Determine the target area and at least two adjacent areas to form an area set, and obtain the geographic location identification information of the area set; S12: Based on the geographic location identification information of the region set, collect the total load data of the region set in the same historical period and the current period from the power grid data platform, collect the distributed photovoltaic power generation data of the region set in the same historical period and the current period from the distributed photovoltaic monitoring system, and collect meteorological data covering the region set at the corresponding spatiotemporal scale from the meteorological data service interface. S13: Perform data cleaning and outlier processing on the collected historical total load data, distributed photovoltaic power generation data, and meteorological data for the same period and current time, respectively, to obtain the cleaned historical total load data, distributed photovoltaic power generation data, and meteorological data for the same period and current time. S14: Align and match the cleaned total load data, distributed photovoltaic power generation data, and meteorological data according to timestamps and regional identifiers to generate historical total load data, distributed photovoltaic power generation data, and meteorological data for the same period and current time in a standardized format.

3. The smart grid load forecasting method for distributed photovoltaic power according to claim 2, characterized in that, The data cleaning and outlier handling in S13 specifically involves: using the sliding time window standard deviation method to identify and remove outliers in the total load data, the distributed photovoltaic power generation data, and the meteorological data, and using linear interpolation to fill in missing values.

4. The smart grid load forecasting method for distributed photovoltaic power according to claim 3, characterized in that, S2 includes: S21: Based on the standardized format of the distributed photovoltaic power generation data collected in S1, calculate the Pearson correlation coefficient of the time series of distributed photovoltaic power generation data of the target area and each adjacent area within a preset time window, and generate spatial correlation features that characterize the spatial dependence between regions. S22: Based on the standardized format of distributed photovoltaic power generation data collected in S1, a time convolutional network is used to extract the long-term trend features and short-term fluctuation features of the time series of distributed photovoltaic power generation data in each region, and generate time series evolution features that characterize the time evolution law of each region. S23: The spatial correlation features generated in S21 and the temporal evolution features generated in S22 are spliced ​​and fused, and the high-dimensional features after fusion are reduced in dimensionality by principal component analysis algorithm to generate a low-dimensional spatiotemporal feature matrix.

5. The smart grid load forecasting method for distributed photovoltaic power according to claim 4, characterized in that, In step S23, the high-dimensional features after fusion are reduced in dimensionality using the principal component analysis algorithm. Specifically, the spatial correlation features and the temporal evolution features are concatenated along the feature dimension to form a high-dimensional fused feature vector. The principal component analysis algorithm is then used to reduce the dimension of the high-dimensional fused feature vector to a preset dimension while maintaining more than 95% of the original feature variance, thereby generating the low-dimensional spatiotemporal feature matrix.

6. The smart grid load forecasting method for distributed photovoltaic power according to claim 5, characterized in that, S3 includes: S31: Align the spatiotemporal feature matrix, the total load data of the historical period and the current period in the standardized format, and the meteorological data in the standardized format according to the time series, and perform splicing operation along the feature dimension to generate a multimodal fusion feature vector; S32: The multimodal fusion feature vector is input into a pre-trained machine learning prediction model. The pre-trained machine learning prediction model performs nonlinear transformation and feature extraction on the multimodal fusion feature vector through an internal multi-layer neural network structure to generate a load prediction feature representation for a specific future time period. S33: The pre-trained machine learning prediction model performs a linear transformation on the load prediction feature representation through an internal fully connected layer to map the feature dimension to the prediction output dimension, generating the initial load prediction result for the target region.

7. The smart grid load forecasting method for distributed photovoltaic power according to claim 6, characterized in that, The splicing operation along the feature dimension in S31 specifically involves connecting the feature vectors corresponding to the spatiotemporal feature matrix, the total load data of the historical period and the current period in the standardized format, and the meteorological data in the standardized format along the feature dimension to form a joint feature vector, namely the multimodal fusion feature vector.

8. The smart grid load forecasting method for distributed photovoltaic power according to claim 7, characterized in that, S4 includes: S41: Acquire real-time updated ultra-short-term weather forecast data of the target area in real time, and compare the real-time updated ultra-short-term weather forecast data with the meteorological data collected in S1 for generating the initial load prediction results, and calculate the meteorological data deviation value. S42: Input the meteorological data deviation value into the pre-trained error correction model, and the pre-trained error correction model outputs a prediction error compensation value corresponding to the initial load prediction result based on the meteorological data deviation value. S43: The initial load prediction result generated in S3 is algebraically added to the prediction error compensation value obtained in S42 to complete the dynamic correction of the initial load prediction result, and the final target area load prediction value is output.

9. The smart grid load forecasting method for distributed photovoltaic power according to claim 8, characterized in that, The algebraic addition in S43 specifically involves adding the initial load prediction result and the prediction error compensation value according to a one-to-one correspondence at each time point.

10. A smart grid load forecasting system for distributed photovoltaic power generation, used to implement the smart grid load forecasting method for distributed photovoltaic power generation as described in any one of claims 1-9, characterized in that, Includes the following modules: Data acquisition module: used to collect total load data, distributed photovoltaic power generation data and corresponding spatiotemporal meteorological data of the target area and at least two adjacent areas during the same period in history and the current period, and to clean, align and standardize the data to generate a dataset in a standardized format; Feature extraction module: used to extract spatial correlation features between regions and temporal evolution features of each region based on the distributed photovoltaic power generation data in the standardized format, and to fuse and reduce the dimensionality of the features to generate a spatiotemporal feature matrix; Predictive modeling module: This module is used to fuse the spatiotemporal feature matrix with the total load data of the same period in the historical time and the current time in the standardized format, as well as the meteorological data in the standardized format, and input the data into the pre-trained machine learning prediction model to generate the initial load prediction results for the target area in the future for a specific time period. Dynamic correction module: It is used to receive real-time updated ultra-short-term weather forecast data of the target area, calculate the meteorological data deviation value, generate the prediction error compensation value through the error correction model, perform algebraic correction on the initial load prediction result, and output the final target area load prediction value.

Images