A distributed photovoltaic power missing data restoration method based on similarity matching and cross-domain association modeling

Abstract

The application discloses a kind of distributed photovoltaic power missing data repair method based on similarity matching and cross-domain association modeling, belong to new energy and power system technical field.The method first utilizes deep similarity matching strategy, through auto-encoder the deep features of historical irradiance sequence are extracted and feature reconstruction is carried out, the multidimensional similarity of the time period to be repaired and historical period is calculated, constructs similar power candidate set and selects the optimal matching sample, forms similar feature matrix.On this basis, introduce deep cross network model to the association modeling of cross time domain similar features, through feature cross extraction high-order nonlinear characteristics, improve the accuracy and robustness of power data repair.The method can effectively alleviate the problem of power data missing caused by abnormal collection, channel congestion or equipment failure, significantly improve the input data quality of power prediction model, support the safe and efficient operation of distributed photovoltaic system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of new energy and power systems, in particular to a distributed photovoltaic power missing data repair method based on similarity matching and cross-domain association modeling. BACKGROUND

[0002] With the wide application of distributed photovoltaic systems in the energy internet, they play an increasingly important role in grid dispatching, power generation prediction and operation optimization. However, due to the influence of weather conditions, irradiance, temperature, component aging and shading, the power output of photovoltaic systems has significant volatility and uncertainty. In actual application scenarios, power data acquisition links are also easily disturbed by communication failures, sensor failures, equipment maintenance or sudden extreme weather, resulting in missing or abnormal power monitoring data. If these missing data cannot be repaired in time and accurately, it will have a serious impact on key business links such as short-term power prediction, load dispatching optimization and abnormal condition diagnosis, and reduce the economy and safety of distributed photovoltaic power station operation.

[0003] Traditional power missing data repair methods are mostly based on statistical modeling, time series completion or interpolation algorithms, such as mean interpolation, linear interpolation, Kalman filtering or simple substitution based on similar days. Although these methods are easy to implement, they usually assume the stationarity or linear characteristics of the power sequence, lack the ability to characterize complex nonlinear coupling characteristics, and cannot fully utilize the deep association information between multi-source data, resulting in insufficient accuracy of the repair results. In the context of the continuous expansion of distributed photovoltaic clusters, the significant increase in data dimensions and feature complexity, traditional methods often fail or perform poorly in high-dimensional, multi-modal and multi-scale scenarios.

[0004] In recent years, with the development of deep learning, transfer learning and multi-modal data processing technology, power missing data repair methods based on deep neural networks have gradually attracted attention. Such methods can extract complex spatio-temporal correlation features through multi-layer nonlinear mapping and have strong fitting ability. However, existing deep learning repair methods still have shortcomings: on the one hand, some models fail to fully utilize the pattern information of different time scales and spatial locations, resulting in unstable repair results; on the other hand, the models lack sufficient context dependence on missing data, making it difficult to achieve high-quality power recovery in the case of large-scale missing data or noise interference. In addition, most methods lack multi-index joint optimization strategies for actual business scenarios, and cannot balance the repair accuracy, generalization ability and computational efficiency.

[0005] Therefore, there is an urgent need for a distributed photovoltaic power missing data repair method based on similarity matching and cross-domain association modeling to solve the above problems. SUMMARY

[0006] The present application aims to solve the technical problems raised in the background art, and provides a distributed photovoltaic power missing data repair method based on similarity matching and cross-domain association modeling, comprising the following steps:

[0007] S1: intra-day weather pattern division: input the historical irradiance sequence into self-encoders, and determine the intra-day weather pattern label to which each sample belongs based on the reconstruction error results of each sample under different encoders , wherein is a preset number of weather pattern categories;

[0008] S2: similar feature extraction: taking the target period irradiance and the candidate period irradiance as inputs, calculating the local dynamic time warping irradiance similarity index , and calculating the physical consistency index based on the power-irradiance conversion model , fusing the two types of indexes to obtain the comprehensive similarity, and screening out candidate periods with the smallest similarity;

[0009] S3: similar feature matrix construction: extracting power, irradiance and weight multi-modal features corresponding to the missing position from the selected candidate periods to form a similar feature matrix , wherein is the number of candidate periods, is the feature dimension;

[0010] S4: cross-time domain deep cross modeling: inputting the similar feature matrix into the cross-time domain deep cross network, fusing multi-period and multi-modal information, and obtaining the target period power prediction sequence

[0011] S5: model training and optimization: taking the residual error between the repaired power and the target power as the supervision signal, and iteratively updating the network parameters;

[0012] S6: missing data repair output: outputting the repaired target period power sequence to complete the missing power data.

[0013] As a preferred technical solution of the present application: the training target of the first self-encoder in the intra-day weather pattern division is:

[0014]

[0015] wherein: is a training set is a set of sample indexes currently labeled as patterns in the middle is an irradiance vector of a sample is a reconstruction vector of a sample is a reconstruction vector of a sample is a reconstruction vector of a sample is a self-encoder is a self-encoder is a preset number of weather pattern categories

[0016] As a preferred technical solution of the present application: based on the reconstruction results of different self-encoders of S1, sample pattern recognition and label assignment are performed, which satisfies the following formula:

[0017]

[0018]

[0019]

[0020] Wherein: is an irradiance vector of a sample is a reconstruction result of a sample under a self-encoder is a reconstruction result of a sample under a self-encoder is an intraday weather pattern label of a sample is a reconstruction error of a sample under a model is a measured irradiance of a sample at a time is a reconstructed irradiance of a sample at a time is a number of sampling points in a day As a preferred technical solution of the present application: the local dynamic time warping irradiance similarity index is calculated as follows:

[0021]

[0022]

[0023] Wherein: is a local irradiance similarity index of a target period and a candidate period is a target period irradiance sequence is a candidate period irradiance sequence is a local alignment path matrix ​​​​A set of paths that allow arbitrary start and end points and satisfy monotonicity and continuity constraints.

[0024] As a preferred technical solution of the present invention: the physical consistency index Determined using the power-irradiance conversion model:

[0025]

[0026]

[0027]

[0028] in: Parameters of the power-irradiance conversion model; :Target time period at time The actual power; :Target time period at time irradiance; :Irradiance sequence for candidate time periods; The actual power sequence for the target time period; : The repair power sequence after model transformation for candidate time periods; Diagonal weight matrix; Diagonal weight matrix The The diagonal elements, at time... When data is missing ,otherwise .

[0029] As a preferred technical solution of the present invention: in S2, and The overall similarity is obtained by fusion and sorted, and the top ones are selected. Multiple candidate time periods are used to extract multimodal features and construct... .

[0030] As a preferred technical solution of the present invention: the cross-temporal deep cross network is calculated layer by layer as follows:

[0031]

[0032] in: :No. Layer network input features; :No. Layer network output features; :No. Layer weight matrix; :No. Layer bias vector; :Hadamard element-wise multiplication operator.

[0033] As a preferred technical solution of the present application: the S5 model training stage is repaired with power And the residual error of the target power The iterative update of the network parameters is supervised by the residual error, which improves the repair accuracy and obtains the repair value of the missing point.

[0034] As a preferred technical solution of the present application: the diagonal weight matrix Satisfies And the comprehensive similarity is screened in the form of Top- To ensure The representativeness and consistency with the target mode.

[0035] The present application also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the method steps of a distributed photovoltaic power missing data repair method based on similarity matching and cross-domain association modeling.

[0036] Beneficial effects: first, the present application is different from the traditional matching method based on statistical similarity. The method relies on the feature reconstruction capability of the deep autoencoder, can accurately depict the intraday irradiance fluctuation mode characteristics, and quantifies the reconstruction performance of the historical irradiance segment to divide different weather modes. On this basis, the present method further combines the physical law and time sequence characteristics to select the most representative similar power segment, thereby constructing a high-precision similar feature matrix to provide reliable support for subsequent power data repair.

[0037] Secondly, the method uses the similar feature matrix containing cross-time domain information to realize deep modeling of cross-time domain association through deep feature cross mechanism, extract more expressive multi-dimensional deep features, and effectively improve the nonlinear fitting ability of the model. At the same time, the feature cross process can fully utilize the complementary relationship between different features to make up for the data sparsity problem caused by missing, so that the model has good stability and robustness in multiple scenarios, thereby realizing high-quality repair of power missing data.

[0038] Finally, the present application has significant advantages in engineering application and scalability. Through the modular feature extraction and deep modeling framework, the present method is not only suitable for photovoltaic cluster data repair of different scales and different regions, but also supports rapid deployment and iterative optimization of cloud and edge computing environment, significantly shortens the repair delay time. In addition, the algorithm structure has high interpretability, which is convenient for engineers to optimize and maintain parameters in actual operation, and further improves the operability and commercialization value of the system. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1A flow chart of a distributed photovoltaic power missing data repair method based on similarity matching and cross-domain association modeling is proposed for the present application. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0041] Reference Figure 1 , the present application is used for repairing power missing points in distributed photovoltaic power station operation data. The method works in the flow of "day weather pattern division, similar feature extraction, similar feature matrix construction, cross-time domain deep cross modeling, training and optimization, and output of repair value". The core is to use historical irradiance / power information for similarity matching, and to complete cross-domain feature fusion and missing point prediction by a deep cross network, and to improve robustness and precision by combining physical consistency constraints.

[0042] Firstly, S1 day weather pattern division is performed. The historical irradiance sequence is input into autoencoders, which respectively learn the reconstruction features of different weather patterns. The training target of the first autoencoder is:

[0043]

[0044] Among them: is the index set of samples in the training set annotated as mode ; is the irradiance vector of sample ; = is the reconstruction vector thereof; is the first autoencoder; is the preset number of weather pattern categories.

[0045] Based on the above trained multiple autoencoders, the samples are used for pattern recognition and label assignment. For any sample , the reconstruction and label calculation of the multiple encoders satisfy:

[0046]

[0047]

[0048]

[0049] In the formula: : sample irradiance vector of the sample; : sample reconstruction result under the first autoencoder; : sample intra-day weather pattern label; sample reconstruction error under the first model; : sample measured irradiance at time ; sample reconstructed irradiance at time ; : number of sampling points in a day. The "reconstruction and label" process can be iterated alternately until the error and label are stable.

[0050] Enter S2 similar feature extraction. With the target period irradiance and historical candidate period irradiance as input, first calculate the local dynamic time warping (LDTW) irradiance similarity index (formula (5)):

[0051]

[0052] In the formula: , are the irradiance vectors of the target and candidate periods, respectively; is the local alignment path matrix; is the path set that allows arbitrary start and end points and satisfies the monotonicity and continuity constraints; is the matrix inner product. The local constraint of LDTW allows the similarity measure to focus on the local fluctuation characteristics of the irradiance.

[0053] In combination with the physical consistency of the power-irradiance conversion model, first estimate the model parameters from the target period data, and then map the candidate period to equivalent power and measure the deviation. The specific calculation method is as follows:

[0054]

[0055]

[0056]

[0057] In the formula: is the actual power of the target period at time ; is the corresponding irradiance; is the candidate period irradiance sequence; The target power sequence; The repair power of the candidate sequence obtained by the model; A power-irradiance conversion model; It is a diagonal weight matrix. For its first One diagonal element, missing point Non-missing points .Will and The fusion yields a comprehensive similarity score, which is then used to rank and filter historical samples. Candidate time periods.

[0058] Perform S3 similarity feature matrix construction. From the selected... Multimodal features, including power, irradiance, and weights corresponding to missing locations, are extracted from each candidate time period and stacked into a similarity feature matrix. ,in Number of candidate time periods This is the feature dimension. This matrix explicitly carries similarity information in a "cross-sample, cross-time" structure and is the core input for subsequent modeling.

[0059] In S4 cross-temporal deep cross modeling, Input a deep cross-network to perform multi-layer feature interaction and residual fusion.

[0060] Layer-by-layer calculation satisfies the following formula:

[0061]

[0062] In the formula: and The first and Input and output characteristics of layered networks; This is the weight matrix for this layer; It is the bias vector; :Hadamard element-wise multiplication operator, residual term It helps maintain low-level patterns and stable training. Rate prediction sequence .

[0063] During the S5 training optimization phase, power was repaired. With target power The residuals are used as monitoring signals to iteratively update the network parameters; the weight matrix It is used to suppress or amplify the loss at missing locations. In the early stage of training, the consistency of observable segments is emphasized, and in the later stage, the fitting weight of missing segments is gradually increased, so as to obtain a robust repair output.

[0064] S6 outputs the repair result: at the missing timestamp... Backfill, form a complete target power sequence. To ensure the normative of sample selection and weight setting, then Apply the constraint, and always use the Top- rule to filter the candidate period to ensure the representativeness of the weather pattern consistent with the target period.

[0065] In implementation, data preparation usually includes: cutting the original monitoring data by day, limiting the amplitude of outliers and marking null values, and uniformly sampling and aligning the irradiance and power sequences; model hyperparameters such as the number of encoders , the number of candidates , the number of network layers and hidden dimensions, learning rate and batch size can be adjusted according to the size of the site and the amount of data; when the system is running online, S1-S6 can be repeatedly executed in a rolling window, or only S2-S6 can be updated to reduce latency.

[0066] In device implementation, the processor loads the program stored in the computer readable storage medium, and sequentially executes the above steps, so that the distributed photovoltaic power missing data repair method can be realized;

[0067] In engineering scenarios, for example, a 30-minute continuous missing occurs in a 5-minute sampled rooftop station, the mode division of S1 is first completed according to the method, then and Top- candidates are calculated and filtered by S2, S3 constructs and inputs to the corresponding network to obtain , S5 is iterated to convergence, and S6 is backfilled and output. Offline evaluation shows that, compared with linear interpolation and single model method, MAPE and RMSE are significantly reduced, and the turning point position and amplitude are highly consistent with the true power; when multiple stations are linked, similar segments of adjacent stations can be integrated into the candidate set to further improve the accuracy and stability.

[0068] The above only describes the preferred embodiments of the present application and should not be used to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.​

Claims

1. A method for repairing missing distributed photovoltaic power data based on similarity matching and cross-domain association modeling, characterized in that, Includes the following steps: S1: Intraday Weather Pattern Classification: Historical Irradiance Series Enter to In an autoencoder, using Each autoencoder is modeled accordingly. Based on the irradiance fluctuation characteristics under different encoders, the intraday weather model label of each sample was determined. ; S2: Similar Feature Extraction: Based on irradiance during the target time period With candidate time period irradiance Using this as input, calculate the local dynamic time-warped irradiance similarity index. And based on the power-irradiance conversion model Calculate physical consistency index By combining the two types of indicators to obtain a comprehensive similarity score, the pair with the lowest similarity score is selected. One candidate time period; S3: Similarity Feature Matrix Construction: From the selected Power, irradiance, and weighted multimodal features corresponding to missing locations are extracted from each candidate time period to form a similarity feature matrix. ,in For the number of candidate time periods, For feature dimensions; S4: Cross-temporal deep cross modeling: combining similar feature matrices By inputting a deep cross-temporal network and fusing information from multiple time periods and multiple modalities, a power prediction sequence for the target time period is obtained. ; Calculate layer by layer to satisfy the following formula: ； In the formula: and The first and Input and output characteristics of layered networks; This is the weight matrix for this layer; It is the bias vector; :Hadamard element-wise multiplication operator, residual term It helps maintain low-level patterns and stable training; Rate prediction sequence ; S5: Model Training and Optimization: To Repair Power With target power The residuals serve as monitoring signals, and the network parameters are iteratively updated. S6: Missing Data Repair Output: Outputs the repaired power sequence for the target time period, thus completing the missing power data.

2. The method for repairing missing distributed photovoltaic power data based on similarity matching and cross-domain association modeling as described in claim 1, characterized in that: The first in the intraday weather pattern division The training objective of each autoencoder is: ； in: For training set The current label is mode. The set of sample indexes; For the sample irradiance vector; For the sample The reconstructed vector, : is the k-th autoencoder that can characterize the irradiance characteristics of the corresponding category.

3. The method for repairing missing distributed photovoltaic power data based on similarity matching and cross-domain association modeling according to claim 1, characterized in that: Based on the reconstruction results of different autoencoders in S1, sample pattern recognition and label assignment are performed, and the calculation satisfies the following formula: ； ； ； in: :Daily weather pattern tags; :sample In the Reconstruction results under an autoencoder; sample In the Reconstruction error under the model; :sample At any moment Measured irradiance; sample At any moment Reconstructed irradiance; Number of sampling points within a day.

4. The method for repairing missing distributed photovoltaic power data based on similarity matching and cross-domain association modeling as described in claim 1, characterized in that: The local dynamic time-warped irradiance similarity index Calculate using the following formula: ； in: Local similarity index of irradiance between the target time period and the candidate time period; :Target period irradiance sequence; :Irradiance sequence for candidate time periods; Local alignment path matrix; A set of paths that allow arbitrary start and end points and satisfy monotonicity and continuity constraints.

5. The method for repairing missing distributed photovoltaic power data based on similarity matching and cross-domain association modeling according to claim 1, characterized in that: The physical consistency index Determined using the power-irradiance conversion model: ； ； ； in: Parameters of the power-irradiance conversion model; :Target time period at time The actual power; :Target time period at time irradiance; :Irradiance sequence for candidate time periods; The actual power sequence for the target time period; : The repair power sequence after model transformation for candidate time periods; Diagonal weight matrix; Diagonal weight matrix The The diagonal elements, at time... When data is missing ,otherwise .

6. The method for repairing missing distributed photovoltaic power data based on similarity matching and cross-domain association modeling according to claim 1, characterized in that: In S2, and The overall similarity is obtained by fusion and sorted, and the top ones are selected. Multiple candidate time periods are used to extract multimodal features and construct... .

7. The method for repairing missing distributed photovoltaic power data based on similarity matching and cross-domain association modeling according to claim 1, characterized in that: The S5 model training phase aims to repair power. With target power The residual is used as a monitoring signal to iteratively update the network parameters, improve the repair accuracy, and obtain the repair value of the missing points.

8. The method for repairing missing distributed photovoltaic power data based on similarity matching and cross-domain association modeling according to claim 5, characterized in that: The diagonal weight matrix satisfy And a top-ranked method based on comprehensive similarity is used. The method of selecting candidate time periods is used to ensure The representativeness is consistent with the target model.

9. A computer-readable storage medium, characterized in that: It stores a computer program thereon, which, when executed by a processor, implements the steps of the method described in any one of claims 1 to 8.

Images