Multi-source data reconstruction method, device and equipment of photovoltaic power station and storage medium

Abstract

The present application relates to the technical field of data reconstruction, and particularly relates to a multi-source data reconstruction method, device and equipment for a photovoltaic power station and a storage medium, which breaks away from the limitations of traditional photovoltaic data management relying on manual inspection and subjective judgment, realizes accurate reconstruction of power station electrical quantities and meteorological data through a complete algorithm link of multi-source heterogeneous data time sequence alignment, quantization error identification, differentiable physical constraint repair and multi-task joint reconstruction, and ensures physical consistency and result reproducibility of multi-source data reconstruction with the aid of a confidence scoring mechanism and a standardized data set management mode. The method solves problems such as data preprocessing loss, uncontrollable reconstruction error and violation of physical constraints, reduces costs, improves real-time performance and accuracy of data reconstruction, effectively guarantees data authenticity and operation and maintenance decision reliability, provides systematic technical support for photovoltaic power stations, and is suitable for standardized management and control scenes of data reconstruction of various new energy power stations.

Description

Technical Field

[0001] This invention relates to the field of data reconstruction technology, and in particular to a method, apparatus, equipment and storage medium for multi-source data reconstruction of a photovoltaic power plant. Background Technology

[0002] Currently, photovoltaic power plants face numerous technical challenges in multi-source data reconstruction and intelligent operation and maintenance. Existing data governance models rely excessively on manual on-site inspections, and the evaluation of data quality and reconstruction effectiveness largely depends on subjective experience, lacking objective quantitative indicators. This results in insufficient credibility of reconstruction results, making it difficult to support efficient operation and maintenance and scientific decision-making for power plants. Furthermore, data management only covers the acquisition and output stages, lacking real-time monitoring and correction mechanisms for the entire process, including outlier removal, time sequence alignment, and error calibration. This leads to problems such as information attenuation, time sequence misalignment, and violations of physical relationships during data transmission and processing.

[0003] Furthermore, the multi-source data reconstruction system operates independently from core business systems such as power plant energy management and condition monitoring. Inconsistent data formats and frequent duplicate data entry severely impact data accuracy and timeliness. Existing data processing platforms are cumbersome to operate and poorly adapted for mobile devices, failing to meet the demands of rapid on-site operations and further reducing the efficiency of data standardization.

[0004] In addition, the lack of physical constraints and joint optimization mechanisms in the reconstruction of multi-source heterogeneous data results in poor consistency between the reconstructed electrical quantities and micro-meteorological field data, uncontrollable errors, and data privacy leakage risks and difficulties in cross-power station collaboration in some scenarios. These issues have long constrained the digital and intensive operation and maintenance level of photovoltaic power stations and affected the stable and efficient operation of the energy system.

[0005] It is evident that existing technologies still need improvement and enhancement. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, the present invention aims to provide a multi-source data reconstruction method for photovoltaic power plants. It constructs a closed-loop system covering the entire process from multi-source data preprocessing, error identification, physical constraint repair to joint reconstruction and confidence assessment, thereby achieving accurate, physically consistent, and objectively quantitative evaluation of the reconstruction of electrical and meteorological data of the power plant.

[0007] The first aspect of this invention provides a method for reconstructing multi-source data from a photovoltaic power plant, comprising: acquiring an original multi-source heterogeneous dataset of the photovoltaic power plant; performing standardized preprocessing on the original multi-source heterogeneous dataset to obtain multi-source time-series data, wherein the multi-source time-series data includes initial total active power, initial three-phase electrical quantities, and initial meteorological data; inputting the initial total active power into a pre-trained quantization error identification model to obtain an error estimate; fusing the initial total active power with the error estimate to obtain a reconstructed total active power; performing physical constraint repair on the initial three-phase electrical quantities through a pre-trained differentiable physical guidance network to obtain repaired three-phase electrical quantities; inputting the error estimate, the repaired three-phase electrical quantities, and the initial meteorological data into a pre-trained multi-task joint learning model for data reconstruction processing to obtain reconstructed data, wherein the reconstructed data includes reconstructed three-phase electrical quantities, reconstruction error estimates, and reconstructed meteorological data; evaluating the credibility of the reconstructed data based on the reconstruction error estimates, generating a corresponding confidence score; and integrating the reconstructed three-phase electrical quantities, reconstructed total active power, reconstructed meteorological data, and the corresponding confidence scores into a standard dataset.

[0008] Optionally, in a first implementation of the first aspect of the present invention, the standardization preprocessing of the original multi-source heterogeneous dataset to obtain multi-source time-series data includes: removing outliers from the original multi-source heterogeneous dataset that exceed a preset threshold range; finding the location of missing values ​​in the original multi-source heterogeneous dataset according to a preset missing value determination rule, and marking the found missing value location; aligning the multi-source heterogeneous data after removing outliers and marking the missing value location according to timestamps to obtain the multi-source time-series data.

[0009] Optionally, in a second implementation of the first aspect of the present invention, before inputting the initial total active power into a pre-trained quantization error identification model to obtain an error estimate, the method includes: extracting a training total active power set from a preset training dataset, wherein the training total active power set consists of training total active power with different precision information under the same electrical quantity, the precision information being used to characterize the data recording precision of the training total active power, the data recording precision being characterized by the number of decimal places recorded in the training total active power; and selecting high-precision positive samples and low-precision negative samples from the training total active power set according to the precision information to construct learning sample pairs, wherein the high-precision positive samples are data records... The training total active power is defined as the data recording accuracy being greater than or equal to a preset accuracy threshold, where the low-precision negative samples are the training total active power with data recording accuracy less than the preset accuracy threshold. A Siamese network is constructed, and the learning samples are used to extract features from the input Siamese network to obtain a first feature vector of high-precision positive samples and a second feature vector of low-precision negative samples. The first and second feature vectors are input into a contrastive loss function for loss calculation to obtain a loss scalar value. The network parameters of the Siamese network are updated according to the loss scalar value and the optimizer to obtain an optimized Siamese network. The optimized Siamese network is combined with a decoder to obtain the pre-trained quantization error identification model.

[0010] Optionally, in a third implementation of the first aspect of the present invention, the step of inputting the initial total active power into a pre-trained quantization error identification model to obtain an error estimate includes: inputting the initial total active power into the pre-trained quantization error identification model, extracting features through an optimized Siamese network within the model to obtain an initial feature vector; and inputting the initial feature vector into the decoder of the pre-trained quantization error identification model for decoding to obtain the corresponding error estimate.

[0011] Optionally, in a fourth implementation of the first aspect of the present invention, before performing physical constraint repair on the initial three-phase electrical quantities using a pre-trained differentiable physical guidance network to obtain the repaired three-phase electrical quantities, the method includes: constructing a differentiable physical guidance network comprising an uncertainty estimation network and a differentiable repair network; inputting the training error estimate and training statistical features from a preset training dataset into the uncertainty estimation network to obtain dynamic weight coefficients; inputting the training three-phase electrical quantities from the preset training dataset into the differentiable repair network to obtain training repaired three-phase electrical quantities; encoding the preset electrical physical constraints into a differentiable physical loss function, and constructing a data reconstruction loss function comprising a data fidelity term and a data smoothing term based on the training repaired three-phase electrical quantities; weighting and summing the differentiable physical loss function and the data reconstruction loss function according to the dynamic weight coefficients to obtain a total loss function; and updating the network parameters of the differentiable repair network and the uncertainty estimation network according to the total loss function to obtain the pre-trained differentiable physical guidance network.

[0012] Optionally, in a fifth implementation of the first aspect of the present invention, before inputting the error estimate, the repaired three-phase electrical quantities, and the initial meteorological data into a pre-trained multi-task joint learning model for data reconstruction processing, the method includes: constructing an initial multi-task joint learning model comprising a shared feature extraction network, an electrical quantity reconstruction branch, an error correction branch, and a micro-meteorological field estimation branch; inputting the training error estimate, training three-phase electrical quantities, and training meteorological data from a preset training dataset into the shared feature extraction network to extract temporal and spatial features and concatenate them into shared features; inputting the shared features into the electrical quantity reconstruction branch, the error correction branch, and the micro-meteorological field estimation branch respectively to obtain the training reconstructed three-phase electrical quantities, the training reconstructed error estimate, and the training reconstructed meteorological data; calculating the multi-task joint loss based on the training three-phase electrical quantities, the training error estimate, and the training reconstructed meteorological data; and updating the parameters of the initial multi-task joint learning model based on the multi-task joint loss to obtain the pre-trained multi-task joint learning model.

[0013] Optionally, in a sixth implementation of the first aspect of the present invention, the step of evaluating the credibility of the reconstructed data based on the reconstruction error estimate and generating a corresponding confidence score includes: comparing the reconstruction error estimate with a preset upper limit of error; if the reconstruction error estimate is greater than the preset upper limit of error, then marking the corresponding reconstructed data as unreliable; if the reconstruction error estimate is less than or equal to the upper limit of error, then inputting the reconstruction error estimate into a preset monotonically decreasing mapping function to generate the confidence score.

[0014] A second aspect of the present invention provides a multi-source data reconstruction device for a photovoltaic power station, comprising: an acquisition module for acquiring the original multi-source heterogeneous dataset of the photovoltaic power station, performing standardized preprocessing on the original multi-source heterogeneous dataset to obtain multi-source time-series data, wherein the multi-source time-series data includes initial total active power, initial three-phase electrical quantities, and initial meteorological data; a fusion module for inputting the initial total active power into a pre-trained quantization error identification model to obtain an error estimate, fusing the initial total active power with the error estimate to obtain a reconstructed total active power; and a repair module for using a pre-trained differentiable physics-guided network. The initial three-phase electrical quantities are physically constrained and repaired to obtain repaired three-phase electrical quantities. A reconstruction module is used to input the error estimate, the repaired three-phase electrical quantities, and the initial meteorological data into a pre-trained multi-task joint learning model for data reconstruction processing to obtain reconstructed data, which includes reconstructed three-phase electrical quantities, reconstruction error estimates, and reconstructed meteorological data. An integration module is used to evaluate the reliability of the reconstructed data based on the reconstruction error estimates, generate corresponding confidence scores, and integrate the reconstructed three-phase electrical quantities, reconstructed total active power, reconstructed meteorological data, and corresponding confidence scores into a standard dataset.

[0015] A third aspect of the present invention provides a multi-source data reconstruction device for a photovoltaic power plant, the multi-source data reconstruction device for a photovoltaic power plant comprising: a memory and at least one processor, the memory storing instructions; at least one processor calling the instructions in the memory to cause the multi-source data reconstruction device for the photovoltaic power plant to execute the various steps of the multi-source data reconstruction method for the photovoltaic power plant described in any of the preceding claims.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the multi-source data reconstruction method for a photovoltaic power station as described above.

[0017] This invention overcomes the limitations of traditional photovoltaic data governance, which relies on manual inspection and subjective judgment. Through a complete algorithmic chain encompassing multi-source heterogeneous data time-series alignment, quantification error identification, differentiable physical constraint repair, and multi-task joint reconstruction, it achieves accurate reconstruction and objective quantitative evaluation of power plant electrical quantities and micro-meteorological field data. Utilizing a confidence scoring mechanism and standardized dataset management model, it ensures the traceability, physical consistency, and reproducibility of results throughout the multi-source data reconstruction process. This method not only specifically addresses core technical issues in existing technologies, such as missing data preprocessing, uncontrollable reconstruction errors, violation of physical constraints, cross-system data silos, and insufficient privacy and security, but also significantly reduces the manpower and time costs of power plant data governance through algorithmic collaboration and closed-loop process optimization. It improves the real-time performance and accuracy of data reconstruction, effectively ensuring data authenticity and the reliability of operation and maintenance decisions. This provides systematic technical support for the intelligent, distributed, privacy-secure, and efficient operation and maintenance of photovoltaic power plant clusters, and is applicable to standardized management scenarios for multi-source heterogeneous data fusion and state reconstruction in various new energy power plants. Attached Figure Description

[0018] Figure 1 This is a first flowchart of a multi-source data reconstruction method for a photovoltaic power plant provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a multi-source data reconstruction device for a photovoltaic power station provided in an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a multi-source data reconstruction device for a photovoltaic power station provided in an embodiment of the present invention. Detailed Implementation

[0019] This invention provides a method, apparatus, device, and storage medium for multi-source data reconstruction in a photovoltaic power plant. In this invention, the terms "first," "second," "third," "fourth," etc. (if applicable) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0020] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the multi-source data reconstruction method for photovoltaic power plants in this invention includes: 101. Obtain the original multi-source heterogeneous dataset of the photovoltaic power station, perform standardization preprocessing on the original multi-source heterogeneous dataset to obtain multi-source time series data, the multi-source time series data including initial total active power, initial three-phase electrical quantities and initial meteorological data; In this embodiment, the original multi-source heterogeneous dataset is a collection of data from photovoltaic power plants that comes from different sources and has different structures, such as electrical operation data recorded by inverters and environmental data recorded by weather stations. The multi-source time-series data is a time-series dataset formed by aligning data from different sources according to a unified timestamp. The three-phase electrical quantities are the voltage, current, and active power generated by each of the three phases of AC power. If the power plant is a single-phase grid-connected power plant, the three-phase electrical quantities are the voltage, current, and power of the connected single phase, and the data of the other two phases are marked as missing.

[0021] Data from the photovoltaic power plant and its environment are collected to obtain a raw multi-source heterogeneous dataset. Specifically, meteorological data, such as irradiance, ambient temperature, humidity, and wind speed, are collected from weather stations near the power plant. At first preset time intervals (e.g., five minutes), the inverter's reported three-phase voltage, three-phase current, active power, and total active power at the inverter's output are collected. At second preset time intervals (e.g., fifteen minutes), the energy meter's recorded three-phase voltage, three-phase current, and total active power are collected. It should be noted that the inverter-reported data is used for real-time monitoring and local control; the sampling frequency is high, but the accuracy is limited. The data from the energy meter is used for electricity billing and grid connection assessment; the accuracy is high, but the sampling frequency is lower. For example, the inverter-reported values ​​are retained to one decimal place, while the energy meter's recorded values ​​are retained to three decimal places. The first and second preset time intervals can be set according to actual needs; they can be the same or different, and there are no restrictions on this.

[0022] Because the data sources and sampling frequencies differ, the data formats are not uniform. Therefore, the collected raw multi-source heterogeneous datasets undergo standardized preprocessing. Specifically, the data is cleaned, outlier removed, and missing value marked. The processed data is then linearly interpolated and aligned according to a unified timestamp to obtain multi-source time-series data with a unified sampling frequency (e.g., every 5 minutes). This multi-source time-series data includes initial total active power, initial three-phase electrical quantities, and initial meteorological data. By collecting and preprocessing multi-source data, converting it into standard data, standardized input is provided for subsequent error identification and data repair.

[0023] 102. Input the initial total active power into the pre-trained quantization error identification model to obtain the error estimate, and fuse the initial total active power with the error estimate to obtain the reconstructed total active power; In this embodiment, the quantization error identification model is a neural network model pre-trained through contrastive learning. The error estimate is a numerical value representing the difference between the reconstructed total active power and the initial data.

[0024] The initial total active power is input into a pre-trained quantization error identification model. The model extracts features through its internal Siamese network to obtain feature vectors, and then performs multi-layer fully connected mapping processing through a decoder to output an error estimate, for example, 0.05 kW. This error estimate indicates that the model's judgment of the initial total active power deviates from the true value by approximately 0.05 kW. The initial total active power and the error estimate are fused. In this embodiment, the fusion method is addition, i.e., the reconstructed total active power equals the initial total active power plus the error estimate, for example, 3350.1 kW plus 0.05 kW equals 3350.15 kW. This reconstructed value retains two decimal places, providing higher numerical accuracy compared to the initial value. It is understood that the fusion method can also use weighted averaging or other combinations of operations, and this is not limited. By extracting error features from the low-precision initial total active power using the pre-trained quantization error identification model to obtain the reconstructed total active power, the accuracy of the total active power is improved.

[0025] 103. The initial three-phase electrical quantities are physically constrained and repaired using a pre-trained differentiable physical guidance network to obtain the repaired three-phase electrical quantities; In this embodiment, the differentiable physical guidance network is a neural network model pre-trained through joint optimization. This model encodes electrical physical constraints into differentiable loss functions, enabling it to force the output results to conform to electrical laws during the data repair process. The repaired three-phase electrical quantities are three-phase voltage, three-phase current, and three-phase power data that conform to physical laws.

[0026] The initial three-phase electrical quantities are input into a pre-trained differentiable physics-guided network. This network processes the data using temporal convolution and attention mechanisms to extract the temporal variation patterns of each electrical quantity and the correlation features between the three phases, outputting the repaired three-phase electrical quantities. The repaired three-phase electrical quantities satisfy the following physical laws: the sum of the three-phase currents approaches zero; the sum of the three-phase voltage phasors approaches zero; and the sum of the three-phase power approaches the reconstructed total active power. For example, if the initial three-phase currents are I... a =15.2 amperes, I b =15.3 Amperes, I c =15.1 Amperes, the sum of the initial three-phase currents is 45.6 Amperes, significantly deviating from zero; after repair by the differentiable physical guidance network, the output repaired three-phase current may be I. a =15.23 amperes, I b =15.24 amperes, Ic = -30.47 Amperes, making the sum of the three-phase currents close to zero (considering the neutral point current). It should be noted that the sum of the three-phase currents being zero does not mean that the three current values ​​add up to zero, but rather that the phasor sum after considering the phase relationship is zero.

[0027] By using a pre-trained differentiable physics-guided network to physically constrain and repair the initial three-phase electrical quantities, the repaired three-phase electrical quantities can satisfy physical laws while maintaining data fidelity, thus providing physically self-consistent three-phase electrical data for subsequent multi-task joint learning.

[0028] 104. Input the error estimate, the repaired three-phase electrical quantities, and the initial meteorological data into a pre-trained multi-task joint learning model for data reconstruction processing to obtain reconstructed data, which includes reconstructed three-phase electrical quantities, reconstructed error estimates, and reconstructed meteorological data. In this embodiment, the multi-task joint learning model is a model with a graph attention network as its core architecture, capable of simultaneously performing three related tasks: electrical quantity reconstruction, error correction, and micro-meteorological field estimation. Specifically, it is trained using a preset training dataset, which includes sample error estimates, sample three-phase electrical quantities, sample meteorological data, and corresponding real three-phase electrical quantities, real error estimates, and real meteorological data. During training, the model optimizes its parameters by minimizing the multi-task joint loss function. The reconstructed three-phase electrical quantities, reconstructed error estimates, and reconstructed meteorological data are refined and applicable data.

[0029] The error estimate, the corrected three-phase electrical quantities, and the initial meteorological data are input into a pre-trained multi-task joint learning model. The model first extracts temporal and spatial features through a shared feature extraction network and concatenates them into shared features. These shared features are then input into three task branches, simultaneously executing three related tasks. Specifically, the electrical quantity reconstruction branch outputs reconstructed three-phase electrical quantities, the error correction branch outputs reconstructed error estimates, and the micro-meteorological field estimation branch outputs reconstructed meteorological data. The three tasks collaborate through a cross-task attention mechanism. For example, meteorological information helps the electrical quantity reconstruction branch more accurately estimate the impact of weather changes on power output, and electrical quantity features help the micro-meteorological field estimation branch more accurately infer local meteorological conditions. By reconstructing the data using the pre-trained multi-task joint learning model, reconstructed data is obtained, further improving the data's accuracy.

[0030] 105. Evaluate the reliability of the reconstructed data based on the estimated reconstruction error value, generate the corresponding confidence score, and integrate the reconstructed three-phase electrical quantities, reconstructed total active power, reconstructed meteorological data, and the corresponding confidence scores into a standard dataset. In this embodiment, the confidence score is a metric that quantifies the credibility of the reconstructed data, ranging from 0 to 1, with higher values ​​indicating more credible data. The standard dataset refers to a collection of data formed by integrating the reconstructed data types and their corresponding confidence scores in a unified format.

[0031] The reliability of the reconstructed data is assessed based on the estimated reconstruction error value. Specifically, the estimated reconstruction error value is input into a preset monotonically decreasing mapping function to generate a confidence score between 0 and 1. For example, a linear mapping function and the confidence score corresponding to the reconstruction error value can be used. This score represents the reliability of the reconstructed data. It can be understood that the generation of the estimated reconstruction error value is related to the total reconstructed active power, which is generated from three-phase data. Therefore, the confidence score generated based on the estimated reconstruction error value can represent the overall reliability of the reconstructed data. A sigmoid function or an exponentially decaying function can also be used as the mapping function, as long as it satisfies the monotonically decreasing characteristic; there are no restrictions on its specific characteristics.

[0032] The confidence scores are used as data quality identifiers and appended to the corresponding reconstructed data points. This involves associating confidence scores at the same timestamp with the reconstructed three-phase electrical quantities, reconstructed total active power, and reconstructed meteorological data, thus integrating them into a standard dataset. The dataset can be output in CSV, Parquet, or directly written to a database. By generating confidence scores that quantify the reliability of the data based on the reconstruction error estimate and integrating them with other data into a standard dataset, the accuracy, physical consistency, and objective quantitative evaluation of the data reconstruction are achieved.

[0033] This invention discloses a multi-source data reconstruction method for photovoltaic power plants, which overcomes the limitations of traditional photovoltaic data governance relying on manual inspection and subjective judgment. Through a complete algorithmic chain involving multi-source heterogeneous data time-series alignment, quantification error identification, differentiable physical constraint repair, and multi-task joint reconstruction, it achieves accurate reconstruction and objective quantitative evaluation of power plant electrical quantities and micro-meteorological field data. Utilizing a confidence scoring mechanism and standardized dataset management model, it ensures the traceability, physical consistency, and reproducibility of the entire multi-source data reconstruction process. This method not only specifically addresses core technical problems in existing technologies such as missing data preprocessing, uncontrollable reconstruction errors, violation of physical constraints, cross-system data silos, and insufficient privacy and security, but also significantly reduces the manpower and time costs of power plant data governance through algorithmic collaboration and closed-loop process optimization. It improves the real-time performance and accuracy of data reconstruction, effectively ensuring data authenticity and the reliability of operation and maintenance decisions. This provides systematic technical support for the intelligent, distributed, privacy-secure, and efficient operation and maintenance of photovoltaic power plant clusters, and is applicable to the standardized management and control needs of multi-source heterogeneous data fusion and state reconstruction in various new energy power plants.

[0034] In this embodiment of the invention, the standardization preprocessing of the original multi-source heterogeneous dataset to obtain multi-source time-series data includes: 201. Remove outliers from the original multi-source heterogeneous dataset that exceed a preset threshold range; In this embodiment, the preset threshold range is a pre-defined numerical interval for the physically reasonable value range of each data type. Data exceeding this interval is considered an outlier. The preset threshold range is pre-set based on the physical attributes of each data type, the rated parameters of the equipment, and the statistical distribution of historical operating data. Specifically, for electrical quantity data, its reasonable value range is determined based on the rated parameters of the power station equipment and grid connection assessment standards; for meteorological data, its reasonable value range is determined based on the statistical distribution of historical observation data. For example, historical data can be used as the preset threshold range according to a preset quantile interval (e.g., 1st quantile to 99th quantile), or the upper and lower limits of the preset threshold range can be determined using the standard deviation of the mean plus or minus a preset multiple (e.g., three times the standard deviation). The preset threshold range is updated periodically based on the equipment's operating status. Outliers are values ​​that deviate from the normal measurement range due to sensor malfunctions, communication interference, or recording errors.

[0035] Outliers exceeding a preset threshold range in the original multi-source heterogeneous dataset are removed. Specifically, each data point in the original multi-source heterogeneous dataset (a data point is the smallest data unit corresponding to a single timestamp in multi-source time-series data, containing all field values ​​collected or recorded at that moment) is traversed. The value of the current data point is compared with the preset threshold range of its corresponding data type. If the value is lower than the lower limit or higher than the upper limit, the value is marked as an outlier and removed from the dataset. For example, for active power or total active power data, the preset threshold range is 0 kW to 12 kW, where the lower limit of 0 kW is determined based on the power plant's no-output operating condition. The upper limit of 12 kW is determined by multiplying the installed capacity of the power station (e.g., 10 kW) by a preset coefficient (e.g., 1.2). For three-phase voltage data, the preset threshold range is 198 V to 242 V, which is ±10% of the rated voltage of 220 V. For three-phase current data, the preset threshold range is 0 A to the preset upper limit current value (e.g., set to 40 A based on the rated current of the circuit breaker). For irradiance data, the preset threshold range is 0 W / m² to 1200 W / m². For ambient temperature data, the preset threshold range is -40°C to 60°C. For humidity data, the preset threshold range is 0% to 100%. If the total active power collected at a certain moment is -1 kW or 15 kW, it is judged as an outlier and removed.

[0036] By removing outliers from the original multi-source heterogeneous dataset that are outside the physically reasonable range, the negative impact of extreme values ​​caused by sensor failures or recording errors on data quality is eliminated, providing a reliable data foundation for subsequent data repair and model training.

[0037] 202. Locate the missing value positions in the original multi-source heterogeneous dataset according to the preset missing value determination rules, and mark the found missing value positions; In this embodiment, the preset missing value determination rule refers to a set of pre-defined conditions for determining whether data is missing. The missing value location refers to the timestamp and field location corresponding to the missing data in the original multi-source heterogeneous dataset.

[0038] The missing value locations in the original multi-source heterogeneous dataset are located according to a preset missing value determination rule. Specifically, each field in the dataset is traversed, and null values ​​or empty strings are checked. If the content of a field is empty, it is determined to be a missing value; or if certain fields contain preset placeholders (such as NULL, N / A, etc.), they are determined to be missing values. Each missing value location found in the dataset is written to a specific missing value tag (e.g., NaN) to replace the original null values ​​and placeholders. By finding the missing value locations in the original multi-source heterogeneous dataset according to the preset missing value determination rule and uniformly tagging them, missing value location information is provided for subsequent steps, avoiding the mistreatment of missing value tags as actual measured values.

[0039] 203. Align the multi-source heterogeneous data after removing outliers and marking the positions of missing values ​​according to timestamps to obtain the multi-source time-series data; In this embodiment, the multi-source time-series data refers to a data set that is continuous in time and has a uniform format, formed after outlier removal, missing value marking, and timestamp alignment. Specifically, based on the highest sampling frequency, a 5-minute sampling period for inverter data is selected as the basic interval to construct a target time axis. For example, the target time points of the target time axis are set to 10:00:00, 10:05:00, 10:10:00, 10:15:00, and so on. Then, the data collected from each data source that is closest to the target time point is searched. If the deviation between the timestamp of the collected data and the target time point is less than or equal to a preset time deviation threshold, the value of the collected data is assigned to the target time point, where the time deviation threshold is set to half of the sampling interval. If the deviation is greater than the time deviation threshold or the position has already been marked as a missing value, the field corresponding to the target time point is marked as a missing value. Finally, for positions that cannot be matched or have been marked as missing, the missing value marker is retained.

[0040] After timestamp alignment, the original multi-source heterogeneous data is transformed into a set of time-continuous, frequency-uniform data sequences. Each target time point corresponds to a complete record, containing the initial total active power, initial three-phase electrical quantities, and initial meteorological data at that moment. These data are arranged in chronological order to form the multi-source time-series data. By aligning the multi-source heterogeneous data after removing outliers and marking the locations of missing values ​​with uniform timestamps, standardized input is provided for subsequent data processing.

[0041] Before inputting the initial total active power into the pre-trained quantization error identification model to obtain the error estimate, the following steps are included in this embodiment of the invention: 301. Extract the total active power set from the preset training dataset, wherein the total active power set is composed of the total active power of training with different precision information under the same electrical quantity; In this embodiment, the preset training dataset is a training data set composed of historically collected data, the total active power set is a set composed of total active power sets with different precisions, and the precision information refers to the number of decimal places in the data record, which is used to distinguish between high-precision data and low-precision data.

[0042] The pre-set training dataset includes multi-source time-series data collected during the historical operation of the photovoltaic power plant. The total active power, an electrical quantity, has multiple recording sources with varying degrees of precision. For example, the total active power reported by the inverter is rounded to one decimal place (e.g., 3350.1 kW), the total active power recorded by the meter at the power outlet is rounded to three decimal places (e.g., 3350.123 kW), and the total active power recorded manually is an integer (e.g., 3350 kW). All total active power records are extracted from this historical data to form a training total active power set, with each record accompanied by its precision information. Extracting the training total active power set from the pre-set training dataset provides rich training data for subsequent model training.

[0043] 302. Based on the accuracy information, high-precision positive samples and low-precision negative samples are selected from the total active power set of the training to construct learning sample pairs; In this embodiment, the high-precision positive sample refers to a data record with higher precision, and the low-precision negative sample refers to a data record with lower precision. The learning sample pair is a training unit composed of a high-precision positive sample and a low-precision negative sample.

[0044] Based on the accuracy information, high-precision positive samples and low-precision negative samples are selected from the total active power set used in the training dataset. For example, data with three decimal places are selected as high-precision positive samples, and data with one decimal place or an integer are selected as low-precision negative samples. During the selection process, data records that are obviously abnormal or exceed a reasonable range are removed to ensure sample quality. High-precision positive samples and low-precision negative samples with the same timestamp are paired to form a learning sample pair. For example, if the total active power recorded by the meter at a certain moment is 3350.123 kW (high precision), and the total active power reported by the inverter at the same moment is 3350.1 kW (low precision), then these two values ​​are constructed into a learning sample pair, where 3350.123 kW is a positive sample and 3350.1 kW is a negative sample. If no data with the same timestamp exists, a time window matching method is used to pair the closest high-precision records with low-precision records within the time window (e.g., within 5 minutes before and after). It should be noted that the positive and negative samples in the learning sample pairs reflect the values ​​of the same electrical quantity under similar operating conditions.

[0045] By selecting high-precision positive samples and low-precision negative samples from the total active power set of training based on accuracy information, and constructing learning sample pairs, a clear comparative learning objective is provided for the Siamese network, enabling the network to learn discriminative features that distinguish accuracy levels.

[0046] 303. Construct a Siamese network, and extract features from the learning samples input into the Siamese network to obtain the first feature vector of high-precision positive samples and the second feature vector of low-precision negative samples; In this embodiment, the Siamese network refers to a model architecture consisting of two neural network branches with identical structures and shared parameters, used to extract feature vectors of input samples and perform comparative learning.

[0047] The Siamese network consists of two identical neural network branches with completely shared parameters. Each branch employs a hybrid architecture combining temporal convolutional networks and multi-head self-attention mechanisms, and both branches use identical weight parameters. The network's specific structure is as follows: the input layer receives a single total active power value or short sequence data; the intermediate layers consist of stacked multi-layer temporal convolutional modules and multi-head self-attention modules, followed by batch normalization and a non-linear activation function at each layer; and the output layer outputs a fixed-dimensional feature vector, which in this embodiment is set to 128 dimensions.

[0048] High-precision positive samples from the training sample pairs are input into the first branch of the Siamese network, while low-precision negative samples are input into the second branch. The two branches process their respective input samples in parallel and output corresponding feature vectors. Because the input data differs, the output feature vectors will be located in different positions in the vector space.

[0049] By constructing a parameter-shared Siamese network, the high-precision positive samples and low-precision negative samples in the learning sample pair are input into two branches for feature extraction, resulting in the corresponding first feature vector and second feature vector, which provides latent space feature representation for subsequent comparison loss calculation and model optimization.

[0050] 304. Input the first feature vector and the second feature vector into the contrastive loss function to calculate the loss and obtain the loss scalar value; In this embodiment, the contrastive loss function is a loss function used to bring positive sample pairs closer to each other and push negative sample pairs further apart from each other, such as the InfoNCE loss function.

[0051] The first and second feature vectors are input into the contrastive loss function. The contrastive loss function calculates the similarity between each high-precision positive sample feature vector and other high-precision positive sample feature vectors (positive sample pair similarity), and the similarity between each high-precision positive sample feature vector and all low-precision negative sample feature vectors (negative sample pair similarity). Then, for each high-precision positive sample, the ratio of the sum of its positive sample pair similarities to the total sum of positive and negative sample pair similarities is calculated, and the negative logarithm is taken to obtain the loss value for that sample. Finally, the loss values ​​of all samples within a batch are averaged, and the loss scalar value is output. The smaller the loss scalar value, the stronger the network's ability to distinguish between accuracy differences. By inputting feature vectors into the contrastive loss function and using the feature similarity of positive and negative sample pairs within a batch to calculate the loss scalar value, a quantitative basis for the parameter optimization of Siamese networks is provided.

[0052] 305. Update the network parameters of the Siamese network based on the loss scalar value and the optimizer to obtain the optimized Siamese network; In this embodiment, the optimizer is an algorithm that updates neural network parameters based on a scalar loss value, such as the Adam optimizer. The network parameters refer to the trainable weights and biases in the Siamese network. The optimizer updates the Siamese network parameters based on the scalar loss value. Specifically, the optimizer calculates the gradient of each trainable parameter in the Siamese network (e.g., the weights of a temporal convolutional layer, the weights of a multi-head self-attention layer, etc.) using a backpropagation algorithm based on the scalar loss value. The gradient represents the degree of influence of that parameter on the loss value. Subsequently, the Adam optimizer updates the Siamese network parameters in the direction of gradient descent based on the calculated gradients and a preset learning rate (e.g., 0.001), where the new parameters are equal to the old parameters minus the product of the learning rate and the gradient. The training is repeated iteratively until the decrease in the loss scalar value meets a preset convergence condition or reaches a preset maximum number of training epochs. Training then stops. The preset convergence condition is that the decrease in the loss scalar value is less than a preset convergence threshold for a consecutive preset number of epochs. The preset number of epochs is set to 10, and the convergence threshold is set to 0.001. That is, training stops when the decrease in the loss scalar value is less than 0.001 for 10 consecutive epochs, resulting in an optimized Siamese network. By updating the Siamese network parameters based on the loss scalar value and the optimizer, the network can effectively distinguish data at different accuracy levels.

[0053] 306. Combine the optimized Siamese network with the decoder to obtain the pre-trained quantization error identification model; In this embodiment, the decoder is a neural network module that maps feature vectors to error estimates, typically composed of multiple fully connected networks. Specifically, the low-precision total active power is input into the optimized Siamese network to obtain feature vectors; these feature vectors are then input into the decoder to obtain prediction error values; the mean square error between the prediction error value and the true error value is calculated as the loss, and the decoder parameters are updated based on this loss. The above steps are repeated until the decoder converges. Then, the optimized Siamese network and the trained decoder are concatenated to form a unified model. This combined model serves as a pre-trained quantization error identification model. By concatenating the optimized Siamese network and the decoder to form a complete quantization error identification model, the model can estimate the quantization error, providing an error estimation basis for subsequent total active power reconstruction.

[0054] In this embodiment of the invention, the step of inputting the initial total active power into a pre-trained quantization error identification model to obtain an error estimate includes: 401. Input the initial total active power into the pre-trained quantization error identification model, and extract features through the optimized Siamese network within the model to obtain the initial feature vector; In this embodiment, the initial feature vector is the latent space vector representation output after feature extraction of the initial total active power. The accuracy of the acquired initial total active power is inconsistent or does not reach the preset high accuracy; therefore, it needs to be optimized. The initial total active power is input into the pre-trained quantization error identification model. The optimized Siamese network within the model extracts features from the input initial total active power. Specifically, a temporal convolutional layer extracts the local variation pattern of the value in the time series, a multi-head self-attention layer captures the long-distance dependency between the value and the values ​​at previous and subsequent time points, and finally, a fixed-dimensional initial feature vector is output through a fully connected mapping layer. By inputting the initial total active power into the pre-trained quantization error identification model, the initial feature vector is obtained, providing a latent space feature representation for subsequent decoding and generation of error estimates.

[0055] 402. Input the initial feature vector into the decoder of the pre-trained quantization error identification model to decode it and obtain the corresponding error estimate; In this embodiment, after the initial feature vector is input into the decoder, the decoder maps the 128-dimensional feature vector to a 64-dimensional intermediate representation, then maps the 64-dimensional feature vector to a 32-dimensional representation, then maps the 32-dimensional feature vector to a 16-dimensional representation, and finally maps the 16-dimensional feature vector to a 1-dimensional scalar value through a linear output layer. This scalar value is the error estimate. For example, for an initial total active power of 3350.1 kW, the error estimate output by the decoder is 0.05 kW, indicating that the model judges the deviation between the measured value and the true value to be approximately 0.05 kW. It can be understood that the error estimate has the same dimension (kW) as the initial total active power and can be positive or negative. A positive value indicates that the initial value is lower than the true value and needs to be corrected upwards; a negative value indicates that the initial value is higher than the true value and needs to be corrected downwards. By inputting the initial feature vector into the decoder for decoding and mapping, the corresponding error estimate is obtained, completing the end-to-end estimation from low-precision input to quantization error, providing a basis for error correction for subsequent data fusion.

[0056] In this embodiment of the invention, before performing physical constraint repair on the initial three-phase electrical quantities through a pre-trained differentiable physical guidance network to obtain the repaired three-phase electrical quantities, the following steps are included: 501. Construct a differentiable physical guidance network that includes an uncertainty estimation network and a differentiable repair network; In this embodiment, the differentiable physical guidance network is a neural network model that encodes electrical and physical constraints into differentiable loss functions and embeds them into the training process. It consists of two sub-networks: an uncertainty estimation network and a differentiable repair network.

[0057] First, an uncertainty estimation network is constructed. This network is a lightweight, multi-layer, fully connected network. Its inputs are the training error estimate and local statistical features of the training three-phase electrical quantities (e.g., the mean, variance, and rate of change of five time points before and after the current moment). The output is a dynamic weight coefficient α, ranging from 0 to 1. The network structure is as follows: the input layer receives a feature vector of approximately 20 dimensions; two hidden layers, each containing 32 neurons, use the ReLU activation function; and the output layer uses the Sigmoid activation function to compress the output value to between 0 and 1. When the error estimate is large or the local statistical features indicate abnormal data fluctuations, the network outputs a larger α value to enhance the guiding role of physical constraints; when the data quality is good, the network outputs a smaller α value to reduce the intervention of physical constraints.

[0058] Secondly, a differentiable repair network is constructed. The differentiable repair network is a hybrid architecture network based on temporal convolutional networks and multi-head self-attention mechanisms. Its input consists of initial three-phase electrical quantities (including phase A voltage, phase B voltage, phase C voltage, phase A current, phase B current, phase C current, and power of each phase), and its output consists of the repaired three-phase electrical quantities. The structure of the differentiable repair network is as follows: an input layer, three temporal convolutional layers, a multi-head self-attention layer, and an output layer. The input layer receives multi-channel time-series data; the three temporal convolutional layers are connected sequentially, with each convolutional kernel having a size of 3; the multi-head self-attention layer is connected after the temporal convolutional layers; the output layer is a fully connected layer with the same output dimension as the input dimension.

[0059] Finally, the uncertainty estimation network and the differentiable repair network are combined in parallel to form a differentiable physical guidance network. The uncertainty estimation network receives the error estimate and local statistical characteristics, and outputs dynamic weight coefficients, while the differentiable repair network receives the initial three-phase electrical quantities and outputs the repaired three-phase electrical quantities. The two sub-networks are jointly optimized during the training phase, but only the differentiable repair network is used for forward inference during the application phase.

[0060] By constructing a differentiable physical guidance network, a model foundation is provided for subsequent physical constraint repair training. The uncertainty estimation network is used to adaptively adjust the physical constraint strength, and the differentiable repair network is used to perform specific electrical quantity repairs.

[0061] 502. Input the training error estimate and training statistical features from the preset training dataset into the uncertainty estimation network to obtain dynamic weight coefficients; In this embodiment, the training error estimate is the error estimate output by the quantized error identification model based on the training data. The training statistical features are local statistical features extracted from the training three-phase electrical quantities in the preset training dataset, used to describe the distribution characteristics and variation patterns of data points within a local time window. The dynamic weight coefficient is a weight value that adjusts the proportion of physical constraints in the total loss, with a value range of 0 to 1; the worse the data quality (i.e., the larger the error), the larger the weight coefficient.

[0062] The training error estimate and its corresponding local statistical characteristics of the three-phase electrical quantities are input into the uncertainty estimation network. Specifically, a local time window is formed by selecting five time points before and after the current data point. The mean, variance, difference between the maximum and minimum values, median of the absolute value of the first-order difference, difference between the current data and the window mean, and difference between the current data and the data at the adjacent time points are calculated for each of the three-phase electrical quantities (including A-phase voltage, B-phase voltage, C-phase voltage, A-phase current, B-phase current, C-phase current, and power of each phase). Six statistics are extracted for each electrical quantity, for a total of nine electrical quantities across the three phases, resulting in a total of 54 statistical characteristics.

[0063] The training error estimate is concatenated with the training statistical features and input into the uncertainty estimation network. This network is then mapped layer by layer through hidden layers, ultimately outputting dynamic weight coefficients with values ​​ranging from 0 to 1. These dynamic weight coefficients reflect the quality of the current data points; that is, when the error estimate is large or the training statistical data fluctuates abnormally, the weight coefficients are larger, indicating a need to increase the proportion of physical constraints in the total loss.

[0064] By inputting the training error estimate and training statistical features into the uncertainty estimation network, dynamic weight coefficients are generated, enabling adaptive adjustment of the physical constraint strength to the data quality.

[0065] 503. Input the training three-phase electrical quantities from the preset training dataset into the differentiable repair network to obtain the training repair three-phase electrical quantities; In this embodiment, the training and repair of three-phase electrical quantities refers to the three-phase electrical quantity data output by the differentiable repair network after repairing the input data. It should be noted that the data at this time is an intermediate value used for loss function calculation, and is not the final reconstructed data.

[0066] Training three-phase electrical quantities are obtained from a pre-set training dataset. These training three-phase electrical quantities are then input into a differentiable repair network. The differentiable repair network employs a hybrid architecture combining temporal convolutional networks and multi-head self-attention mechanisms to extract features and model the temporal sequence of the input data. The output is a repaired three-phase electrical quantity with the same shape as the input, i.e., the training repaired three-phase electrical quantity. The repaired data is smoother and more continuous in time, and abnormal fluctuations are suppressed.

[0067] By inputting the training three-phase electrical quantities into a differentiable repair network, the training repair three-phase electrical quantities are obtained, providing repaired data for subsequent loss calculation.

[0068] 504. Encode the preset electrical and physical constraints into a differentiable physical loss function, and construct a data reconstruction loss function containing data fidelity terms and data smoothing terms based on the trained and repaired three-phase electrical quantities; In this embodiment, the preset electrical and physical constraints refer to pre-defined physical laws that must be followed, including Kirchhoff's current law, Kirchhoff's voltage law, and three-phase power balance constraints. The differentiable physical loss function is a loss function that encodes the physical constraints into a differentiable mathematical expression; the smaller its value, the more the repair result conforms to physical laws. The data reconstruction loss function is a loss function that measures the difference between the repaired data and the original data, including a data fidelity term and a data smoothing term. The data fidelity term represents the deviation between the repaired result and the original observation, while the data smoothing term represents drastic changes between adjacent time points.

[0069] The preset electrical physical constraints are encoded as differentiable physical loss functions. Specifically, Kirchhoff's current law, Kirchhoff's voltage law, and three-phase power balance constraints are each encoded as differentiable mathematical formulas. Specifically, based on Kirchhoff's current law, the square of the sum of the three-phase currents is calculated; the smaller this value, the closer the three-phase currents are to equilibrium. Based on Kirchhoff's voltage law, the square of the sum of the loop voltage drops is calculated. Based on the three-phase power balance, the square of the difference between the sum of the three-phase power and the corrected total active power is calculated. These three physical losses are then added together to obtain the differentiable physical loss function.

[0070] Based on the training and repair of the three-phase electrical quantities, a data reconstruction loss function is constructed, comprising a data fidelity term and a data smoothing term. Specifically, the data fidelity term is a comparison term between the repaired data and the initial data, used to calculate the difference between them. The smaller the difference, the smaller the fidelity term, indicating that the repair result does not deviate too far from the true measurement value. The smoothing term checks the repair results at adjacent time points, calculating the magnitude of change between them. If the change is too drastic, the smoothing term will increase to adjust for this unreasonable jump. The data reconstruction loss function is obtained by weighted summing of the fidelity term and the smoothing term.

[0071] By encoding the electrical and physical constraints into a differentiable physical loss function and constructing a data reconstruction loss function that includes fidelity and smoothing terms, a two-dimensional loss objective is provided for subsequent adaptive weighted joint optimization.

[0072] 505. Based on the aforementioned dynamic weighting coefficients, the differentiable physical loss function and the data reconstruction loss function are weighted and added together to obtain the total loss function; In this embodiment, the differentiable physical loss function and the data reconstruction loss function are weighted and added together according to the dynamic weighting coefficients. Specifically, the total loss function is: L total =L data +α×L physics L total Let L be the total loss function. data The loss function is used for data reconstruction, where α is the dynamic weighting coefficient and L is the weighting coefficient. physics The physical loss function is differentiable. It can be understood that the differentiable physical loss function is multiplied by a dynamic weighting coefficient α and then added to the total loss. This ensures that when α is large, the physical loss accounts for a higher proportion of the total loss, and the model is more inclined to satisfy physical constraints during training; when α is small, the physical loss accounts for a lower proportion, and the model is more inclined to preserve the original features of the data during training. By weighting and adding the differentiable physical loss function and the data reconstruction loss function according to the dynamic weighting coefficient to generate the total loss function, the adaptive adjustment of the physical constraint strength to data quality is achieved.

[0073] 506. Update the network parameters of the differentiable repair network and the uncertainty estimation network according to the total loss function to obtain the pre-trained differentiable physical guidance network; In this embodiment, the network parameters refer to the trainable weights and biases in the differentiable repair network and the uncertainty estimation network. The preset stopping condition refers to a pre-set training termination criterion, such as reaching the maximum number of training epochs or the convergence of the loss function value.

[0074] The network parameters of the differentiable repair network and the uncertainty estimation network are updated according to the total loss function. Specifically, the partial derivatives of the total loss function with respect to each parameter in the differentiable repair network and the partial derivatives with respect to each parameter in the uncertainty estimation network are calculated using the backpropagation algorithm. Then, the Adam optimizer is used to simultaneously update the parameters of the differentiable repair network and the uncertainty estimation network in the direction of gradient descent. Iterative training is repeated until the training results meet a preset stopping condition, i.e., the number of training iterations reaches a preset maximum (e.g., 200 iterations), or the decrease in the total loss function value is less than a preset threshold (e.g., 0.001) for several consecutive iterations. After training stops, a pre-trained differentiable physical guidance network is obtained. It is important to note that in the application phase, only the differentiable repair network within this network is used for forward inference to repair the initial three-phase electrical quantities.

[0075] By simultaneously updating the parameters of the differentiable repair network and the uncertainty estimation network according to the total loss function, joint optimization of the differentiable repair network and the uncertainty estimation network is achieved. This enables the repair network to enhance physical constraint guidance when the data quality is poor, and to maintain the original characteristics of the data when the data quality is good.

[0076] In this embodiment of the invention, before the data reconstruction processing of the pre-trained multi-task joint learning model, which inputs the error estimate, the repaired three-phase electrical quantities, and the initial meteorological data, the following steps are included: 601. Construct an initial multi-task joint learning model that includes a shared feature extraction network, an electrical quantity reconstruction branch, an error correction branch, and a micro-meteorological field estimation branch; In this embodiment, the shared feature extraction network is a network module that extracts general features from the input data, and it is composed of a stacked bidirectional LSTM and a graph convolutional network connected in parallel. The initial multi-task joint learning model refers to a multi-task neural network model that has not yet been trained, which is composed of the aforementioned shared feature extraction network and three task branches.

[0077] An initial multi-task joint learning model is constructed. Specifically, a shared feature extraction network is formed by parallel stacked bidirectional LSTM and graph convolutional networks. The stacked bidirectional LSTM extracts the temporal features of the input data, while the graph convolutional network extracts the spatial features between multiple power stations. The outputs of the two sub-networks are concatenated along the feature dimension to form the shared features. The input to the shared feature extraction network includes training error estimates, training three-phase electrical quantities, and training meteorological data.

[0078] An electrical quantity reconstruction branch is constructed through a decoder network. The network structure adopts deconvolution or fully connected mapping. Its input is shared features, and its output is the reconstructed three-phase electrical quantities (including three-phase voltage, three-phase current and power of each phase). The shared features are mapped to the output space of the same dimension as the input electrical quantities through the network structure.

[0079] An error correction branch is constructed by using a multi-layer fully connected network. The network structure is a multi-layer fully connected mapping structure. The input is a concatenated vector of shared features and training error estimates. This branch gradually reduces the dimensionality of the input features and finally outputs a scalar value.

[0080] A micro-meteorological field estimation branch is constructed through a temporal convolutional network. The input is shared features, and the output is reconstructed meteorological data. The shared features are mapped to the same dimension as the preset meteorological data through temporal convolutional layers and then output. The shared feature extraction network is then combined in parallel with the three task branches mentioned above to form an initial multi-task joint learning model. The model's input includes training error estimates, training three-phase electrical quantities, and training meteorological data. The three branches output their respective prediction results. By constructing this initial multi-task joint learning model, a model foundation is provided for subsequent multi-task joint training, achieving feature sharing and collaborative learning among the three related tasks.

[0081] 602. Input the training error estimate, training three-phase electrical quantities and training meteorological data from the preset training dataset into the shared feature extraction network, extract the temporal features and spatial features and concatenate them into shared features; In this embodiment, the training error estimate, training three-phase electrical quantities, and training meteorological data from the preset training dataset are input into the shared feature extraction network. The stacked bidirectional LSTM network in the shared feature extraction network extracts features at different time scales and models these features along the time dimension, outputting a temporal feature vector. The graph convolutional network in the shared feature extraction network extracts spatial features between multiple power stations. For a single power station scenario, the graph convolutional network can be omitted or its input can be set to a constant. Finally, the temporal and spatial features are concatenated along the feature dimension to form shared features. By extracting temporal and spatial features through the shared feature extraction network and concatenating them into shared features, feature fusion and information sharing of multi-source heterogeneous data are achieved.

[0082] 603. Input the shared features into the electrical quantity reconstruction branch, the error correction branch, and the micro-meteorological field estimation branch respectively to obtain the training reconstructed three-phase electrical quantities, the training reconstructed error estimate, and the training reconstructed meteorological data; In this embodiment, the training and reconstruction data is the data after refining the error, i.e., the final reconstructed data. Shared features are input to the electrical quantity reconstruction branch. This branch uses a deconvolutional or fully connected mapping network to decode the shared features into an output with the same dimension as the input three-phase electrical quantities, i.e., training and reconstructing the three-phase electrical quantities, including the reconstructed A-phase voltage, B-phase voltage, C-phase voltage, A-phase current, B-phase current, C-phase current, and active power of each phase. The shared features and training error estimates are jointly input to the error correction branch. This branch first concatenates the shared features and training error estimates along the feature dimension, then maps them through a multi-layer fully connected network to output the training and reconstruction error estimate, i.e., the result after refining the input error estimate. Shared features are input to the micro-meteorological field estimation branch. This branch uses a temporal convolutional network to map the shared features into an output with the same dimension as the input meteorological data, i.e., training and reconstructing the meteorological data, including the reconstructed irradiance, temperature, humidity, and wind speed. By inputting shared features into three task branches and synchronously obtaining training reconstruction data, parallel processing and collaborative learning of three related tasks, as well as the acquisition of high-precision data, are achieved.

[0083] 604. Calculate the multi-task joint loss based on the training three-phase electrical quantities, the training error estimate, and the training reconstructed meteorological data; In this embodiment, the multi-task joint loss is a weighted sum of the outputs of the three task branches and their corresponding true values. Specifically, the true values ​​corresponding to the output data of each branch are obtained. For example, the true value of the electrical quantity reconstruction branch is the original training three-phase electrical quantity in the preset training dataset; the true value of the error correction branch is the difference between the training error estimate and the true error (i.e., the difference between the high-precision value and the low-precision value). The loss value of each branch is calculated based on the true values. Specifically, the electrical quantity reconstruction loss uses mean square error, calculating the mean of the squares of the differences between the training reconstructed three-phase electrical quantity and the original training three-phase electrical quantity; the error correction loss uses mean square error, calculating the mean of the squares of the differences between the training reconstructed error estimate and the true error value; and the micro-meteorological field estimation loss uses mean square error, calculating the mean of the squares of the differences between the training reconstructed meteorological data and the original training meteorological data.

[0084] The three loss values ​​are weighted and summed to obtain the multi-task joint loss. For example, L multi =λ1×L recon +λ2×L error +λ3×L meteoWhere λ1, λ2, and λ3 are preset weighting coefficients used to balance the importance of the three tasks. λ1, λ2, and λ3 are all real numbers greater than 0 and less than 1, and satisfy λ1 + λ2 + λ3 = 1. Their values ​​are preset according to the loss magnitude of each task branch. For example, setting λ1 = 0.5, λ2 = 0.3, and λ3 = 0.2, L multi For multi-task joint loss, L recon L error L meteo These are the loss values ​​for the electrical quantity reconstruction branch, the error correction branch, and the micro-meteorological field estimation branch, respectively. By calculating the joint loss of the three tasks based on the outputs of the three task branches and their corresponding true values, joint optimization of the three related tasks is achieved.

[0085] 605. Update the parameters of the initial multi-task joint learning model according to the multi-task joint loss to obtain the pre-trained multi-task joint learning model; In this embodiment, the gradient is calculated based on the multi-task joint loss. The partial derivatives of the multi-task joint loss with respect to the parameters of the shared feature extraction network, the electrical quantity reconstruction branch, the error correction branch, and the micro-meteorological field estimation branch are calculated using the backpropagation algorithm. Then, the optimizer updates the parameters of the shared feature extraction network and the three task branches simultaneously according to the direction of gradient descent. This iterative training is repeated until the multi-task joint loss converges below a preset threshold or reaches a preset maximum number of training rounds. Training then stops, resulting in a pre-trained multi-task joint learning model. At this point, the parameters within the model have been optimized, enabling the output of high-precision reconstructed data. By optimizing the parameters in the multi-task joint learning model, systematic quality repair of multi-source heterogeneous data is achieved, reconstructing low-precision data into a high-precision, physically consistent, standardized dataset that integrates meteorological information.

[0086] In this embodiment of the invention, the step of assessing the credibility of the reconstructed data based on the estimated reconstruction error and generating a corresponding confidence score includes: 701. Compare the estimated reconstruction error with a preset upper limit of error. If the estimated reconstruction error is greater than the preset upper limit of error, mark the corresponding reconstruction data as unreliable. In this embodiment, the preset upper limit of error refers to a pre-set error tolerance threshold used to determine whether the reconstructed data is reliable. The estimated reconstruction error represents the model's judgment on the deviation between the reconstructed total active power and the true value. The estimated reconstruction error is compared with the preset upper limit of error, which is set according to actual application requirements, such as 0.05 kW. If the estimated reconstruction error is greater than 0.05 kW, it indicates that the model's reconstruction error for the data is large. In this case, the reconstructed data corresponding to the estimated reconstruction error is marked as unreliable, and this data point will not participate in downstream tasks or will be assigned a low weight in subsequent standard datasets. By comparing the estimated reconstruction error with the preset upper limit of error, it is possible to effectively determine whether the error of the reconstructed data is within an acceptable range, avoiding irreversible effects caused by excessive data errors.

[0087] 702. If the estimated reconstruction error is less than or equal to the upper limit of the error, the estimated reconstruction error is input into a preset monotonically decreasing mapping function to generate the confidence score. In this embodiment, the confidence score is a metric for quantifying the reliability of the reconstructed data. If the estimated reconstruction error is less than or equal to a preset upper limit error value, a confidence score is generated based on the estimated reconstruction error. For example, a linear mapping function is used as a preset monotonically decreasing mapping function to generate the confidence score. Specifically, the confidence score equals 1 minus the ratio of the estimated reconstruction error to the upper limit error value. For example, if the estimated reconstruction error is 0.03 kW and the upper limit error is 0.5 kW, the confidence score is 1 minus 0.03 divided by 0.5, which is 0.94. This score indicates that the model's reliability for the reconstructed data is 94%. By generating the confidence score based on the error estimate, a quantitative assessment of the quality of the reconstructed data is provided, offering a high-quality data foundation for subsequent data use.

[0088] The multi-source data reconstruction method for a photovoltaic power station in the embodiments of the present invention has been described above. The multi-source data reconstruction device for a photovoltaic power station in the embodiments of the present invention will be described below. Please refer to [link / reference]. Figure 2 One embodiment of the multi-source data reconstruction device for a photovoltaic power station in this invention includes: The acquisition module 801 is used to acquire the original multi-source heterogeneous dataset of the photovoltaic power station, perform standardization preprocessing on the original multi-source heterogeneous dataset to obtain multi-source time series data, which includes initial total active power, initial three-phase electrical quantities and initial meteorological data. The fusion module 802 is used to input the initial total active power into a pre-trained quantization error identification model to obtain an error estimate, and to fuse the initial total active power with the error estimate to obtain a reconstructed total active power. Repair module 803 is used to perform physical constraint repair on the initial three-phase electrical quantities through a pre-trained differentiable physical guidance network to obtain repaired three-phase electrical quantities; The reconstruction module 804 is used to input the error estimate, the repaired three-phase electrical quantity and the initial meteorological data into a pre-trained multi-task joint learning model for data reconstruction processing to obtain reconstructed data. The reconstructed data includes reconstructed three-phase electrical quantity, reconstructed error estimate and reconstructed meteorological data. The integration module 805 is used to evaluate the credibility of the reconstructed data based on the reconstructed error estimate, generate a corresponding confidence score, and integrate the reconstructed three-phase electrical quantities, reconstructed total active power, reconstructed meteorological data, and corresponding confidence scores into a standard dataset.

[0089] Based on the same ideas as the methods in the above embodiments, the apparatus provided by the present invention can implement the methods in the above embodiments.

[0090] The above combination Figure 2 The multi-source data reconstruction device for a photovoltaic power station in this embodiment of the invention is described in detail from the perspective of modular functional entities. The multi-source data reconstruction device for a photovoltaic power station in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0091] Figure 3 This is a schematic diagram of the structure of a multi-source data reconstruction device for a photovoltaic power plant according to an embodiment of the present invention. The multi-source data reconstruction device 900 for the photovoltaic power plant can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 910 (e.g., one or more processors) and a memory 920, and one or more storage media 930 (e.g., one or more mass storage devices) storing application programs 933 or data 932. The memory 920 and storage media 930 can be temporary or persistent storage. The program stored in the storage media 930 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the multi-source data reconstruction device 900 for the photovoltaic power plant. Furthermore, the processor 910 may be configured to communicate with the storage media 930 and execute the series of instruction operations in the storage media 930 on the multi-source data reconstruction device 900 for the photovoltaic power plant to implement the steps of the multi-source data reconstruction method for the photovoltaic power plant provided in the above-described method embodiments.

[0092] The multi-source data reconstruction device 900 for a photovoltaic power station may also include one or more power supplies 940, one or more wired or wireless network interfaces 950, one or more input / output interfaces 960, and / or one or more operating systems 931, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The multi-source data reconstruction equipment structure of the photovoltaic power plant shown does not constitute a limitation on the multi-source data reconstruction equipment of the photovoltaic power plant. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0093] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of a multi-source data reconstruction method for a photovoltaic power station.

[0094] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0095] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0096] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for reconstructing multi-source data from a photovoltaic power station, characterized in that, include: The original multi-source heterogeneous dataset of the photovoltaic power station is obtained, and the original multi-source heterogeneous dataset is standardized and preprocessed to obtain multi-source time series data, which includes initial total active power, initial three-phase electrical quantities and initial meteorological data. The initial total active power is input into a pre-trained quantization error identification model to obtain an error estimate. The initial total active power and the error estimate are then fused to obtain a reconstructed total active power. The initial three-phase electrical quantities are physically constrained and repaired using a pre-trained differentiable physical guidance network to obtain the repaired three-phase electrical quantities. The error estimate, the repaired three-phase electrical quantities, and the initial meteorological data are input into a pre-trained multi-task joint learning model for data reconstruction processing to obtain reconstructed data, which includes reconstructed three-phase electrical quantities, reconstructed error estimates, and reconstructed meteorological data. The reliability of the reconstructed data is evaluated based on the estimated reconstruction error value, and a corresponding confidence score is generated. The reconstructed three-phase electrical quantities, reconstructed total active power, reconstructed meteorological data, and corresponding confidence scores are integrated into a standard dataset.

2. The multi-source data reconstruction method for photovoltaic power plants according to claim 1, characterized in that, The standardization preprocessing of the original multi-source heterogeneous dataset to obtain multi-source time-series data includes: Remove outliers from the original multi-source heterogeneous dataset that exceed a preset threshold range; The missing value locations in the original multi-source heterogeneous dataset are found according to the preset missing value determination rules, and the found missing value locations are marked. The multi-source heterogeneous data, after removing outliers and marking the positions of missing values, is aligned according to timestamps to obtain the multi-source time-series data.

3. The multi-source data reconstruction method for photovoltaic power plants according to claim 1, characterized in that, Before inputting the initial total active power into the pre-trained quantization error identification model to obtain the error estimate, the process includes: A training total active power set is extracted from a preset training dataset. The training total active power set consists of training total active power with different precision information under the same electrical quantity. The precision information is used to characterize the data recording precision of the training total active power. The data recording precision is characterized by the number of decimal places recorded in the training total active power. Based on the accuracy information, high-precision positive samples and low-precision negative samples are selected from the total training active power set to construct learning sample pairs. The high-precision positive samples are the total training active power with data recording accuracy greater than or equal to a preset accuracy threshold, and the low-precision negative samples are the total training active power with data recording accuracy less than the preset accuracy threshold. Construct a Siamese network, and extract features from the learning samples input into the Siamese network to obtain a first feature vector of high-precision positive samples and a second feature vector of low-precision negative samples; The first feature vector and the second feature vector are input into the contrastive loss function to calculate the loss and obtain a scalar value of loss. The network parameters of the Siamese network are updated based on the loss scalar value and the optimizer to obtain the optimized Siamese network; The optimized Siamese network is combined with the decoder to obtain the pre-trained quantization error identification model.

4. The multi-source data reconstruction method for photovoltaic power plants according to claim 3, characterized in that, The step of inputting the initial total active power into a pre-trained quantization error identification model to obtain an error estimate includes: The initial total active power is input into the pre-trained quantization error identification model, and the initial feature vector is obtained by extracting features through the optimized Siamese network within the model. The initial feature vector is input into the decoder of the pre-trained quantization error identification model for decoding to obtain the corresponding error estimate.

5. The multi-source data reconstruction method for photovoltaic power plants according to claim 1, characterized in that, Before obtaining the repaired three-phase electrical quantities by physically constraining the initial three-phase electrical quantities through a pre-trained differentiable physical guidance network, the process includes: Construct a differentiable physics-guided network that includes an uncertainty estimation network and a differentiable repair network; The training error estimate and training statistical features from the preset training dataset are input into the uncertainty estimation network to obtain dynamic weight coefficients; The training three-phase electrical quantities in the preset training dataset are input into the differentiable repair network to obtain the training repair three-phase electrical quantities; The preset electrical and physical constraints are encoded into a differentiable physical loss function, and a data reconstruction loss function containing data fidelity terms and data smoothing terms is constructed based on the trained and repaired three-phase electrical quantities. The differentiable physical loss function and the data reconstruction loss function are weighted and added together according to the dynamic weighting coefficients to obtain the total loss function; The network parameters of the differentiable repair network and the uncertainty estimation network are updated according to the total loss function to obtain the pre-trained differentiable physical guidance network.

6. The multi-source data reconstruction method for photovoltaic power plants according to claim 1, characterized in that, Before the step of inputting the error estimate, the repaired three-phase electrical quantities, and the initial meteorological data into the pre-trained multi-task joint learning model for data reconstruction processing includes: An initial multi-task joint learning model is constructed, which includes a shared feature extraction network, an electrical quantity reconstruction branch, an error correction branch, and a micro-meteorological field estimation branch. The training error estimate, training three-phase electrical quantity and training meteorological data in the preset training dataset are input into the shared feature extraction network to extract time-series features and spatial features and concatenate them into shared features. The shared features are respectively input into the electrical quantity reconstruction branch, the error correction branch, and the micro-meteorological field estimation branch to obtain the trained reconstructed three-phase electrical quantities, the trained reconstructed error estimate, and the trained reconstructed meteorological data. The multi-task joint loss is calculated based on the training three-phase electrical quantities, the training error estimate, and the training reconstructed meteorological data. The parameters of the initial multi-task joint learning model are updated based on the multi-task joint loss to obtain the pre-trained multi-task joint learning model.

7. The multi-source data reconstruction method for photovoltaic power plants according to claim 1, characterized in that, The step of assessing the credibility of the reconstructed data based on the estimated reconstruction error and generating a corresponding confidence score includes: The estimated reconstruction error is compared with a preset upper limit of error. If the estimated reconstruction error is greater than the preset upper limit of error, the corresponding reconstructed data is marked as unreliable. If the estimated reconstruction error is less than or equal to the upper limit of the error, the estimated reconstruction error is input into a preset monotonically decreasing mapping function to generate the confidence score.

8. A multi-source data reconstruction device for a photovoltaic power station, characterized in that, include: The acquisition module is used to acquire the original multi-source heterogeneous dataset of the photovoltaic power station, perform standardization preprocessing on the original multi-source heterogeneous dataset to obtain multi-source time series data, which includes initial total active power, initial three-phase electrical quantities and initial meteorological data. The fusion module is used to input the initial total active power into a pre-trained quantization error identification model to obtain an error estimate, and to fuse the initial total active power with the error estimate to obtain a reconstructed total active power. The repair module is used to perform physical constraint repair on the initial three-phase electrical quantities through a pre-trained differentiable physical guidance network to obtain the repaired three-phase electrical quantities. The reconstruction module is used to input the error estimate, the repaired three-phase electrical quantities, and the initial meteorological data into a pre-trained multi-task joint learning model for data reconstruction processing to obtain reconstructed data. The reconstructed data includes reconstructed three-phase electrical quantities, reconstructed error estimates, and reconstructed meteorological data. The integration module is used to assess the credibility of the reconstructed data based on the reconstructed error estimate, generate a corresponding confidence score, and integrate the reconstructed three-phase electrical quantities, reconstructed total active power, reconstructed meteorological data, and corresponding confidence scores into a standard dataset.

9. A multi-source data reconstruction device for a photovoltaic power station, characterized in that, The multi-source data reconstruction device of the photovoltaic power station includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the multi-source data reconstruction device of the photovoltaic power plant to perform the steps of the multi-source data reconstruction method of the photovoltaic power plant as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the multi-source data reconstruction method for photovoltaic power plants as described in any one of claims 1-7.

Images