Photovoltaic cluster prediction method based on hybrid graph filter and TCN-Transformer

Abstract

The invention provides a photovoltaic cluster prediction method based on a hybrid graph filter and a TCN-Transformer, and the method comprises the steps: collecting the historical power data of a distributed photovoltaic power station, constructing a dynamic power similarity matrix through a time sliding window and a maximum mutual information coefficient, and extracting spatial features; meanwhile, acquiring historical meteorological data, extracting a power similarity matrix and meteorological features through an adaptive mixed graph filter, and constructing a meteorological feature sequence with spatial information; then, the TCN model extracts a characteristic relation between photovoltaic power and meteorological data, and related characteristics are screened through a space attention mechanism; thirdly, the Transform model adopts cross attention to fuse spatial-temporal features, feature weights are given, and the model is trained; and finally, constructing a photovoltaic cluster power prediction model through full connection layer optimization features. According to the method, the photovoltaic station cluster characteristics are effectively mined, the power prediction precision is improved, the problem of dynamic change of photovoltaic cluster correlation is solved, and the method has high popularization value.

Description

Technical Field

[0001] The present invention relates to a photovoltaic cluster prediction method based on a hybrid graph filter and TCN-Transformer, and belongs to the field of new energy power prediction. Background Art

[0002] With the rapid development of renewable energy, photovoltaic power generation, as a key component, is becoming crucial for the transformation of the global energy mix. However, the problem of photovoltaic absorption is becoming increasingly prominent. The volatility and intermittency of photovoltaic power generation pose challenges to the stable operation of the power grid. In particular, in the development and deployment of photovoltaic clusters, how to efficiently predict power generation and optimize grid dispatch to improve the utilization efficiency of photovoltaic power has become a critical issue that needs to be addressed.

[0003] A photovoltaic cluster refers to multiple photovoltaic systems connected together within a specific area. Compared to centralized photovoltaic systems, the decentralized and non-uniform nature of a photovoltaic cluster offers greater flexibility in power generation and management. However, this also results in dynamic correlations, and its forecasting accuracy is far inferior to that of centralized photovoltaic systems. Traditional cluster segmentation and forecasting methods rely on static segmentation followed by forecasting. This approach is not suitable for distributed photovoltaic systems, where correlations can change dynamically. Summary of the Invention

[0004] The purpose of this invention is to provide a photovoltaic cluster prediction method based on a hybrid graph filter and TCN-Transformer to address the problems existing in the above-mentioned background technology. This method can realize the mining of photovoltaic station cluster characteristics and efficient prediction, solve the problem of dynamic changes in photovoltaic cluster correlation, and improve the accuracy of photovoltaic power prediction, which has the potential for promotion.

[0005] The purpose of the present invention is achieved by the following technical measures:

[0006] A photovoltaic cluster prediction method based on a hybrid graph filter and TCN-Transformer is characterized by the following steps:

[0007] S1: Collect historical power data of distributed photovoltaic power plants, calculate the correlation of historical power data of each photovoltaic power plant based on the time window and maximum mutual information coefficient, and then construct a dynamic matrix of historical power correlation by sliding the time window;

[0008] S2: Collect historical meteorological data of distributed photovoltaic power stations, extract the historical power correlation dynamic matrix and meteorological data features obtained based on S1 through an adaptive hybrid graph filter, and construct a new meteorological feature sequence with spatial correlation;

[0009] S3: The TCN prediction model is used to extract the characteristic relationship between the historical power data and historical meteorological data of distributed photovoltaic power stations and then output it; the spatial attention mechanism is used to filter the spatially correlated feature information in the new spatially correlated meteorological feature sequence obtained based on S2 and then output it; the Transformer model learns the input spatially correlated feature information and assigns weights to realize the training of the Transformer model; then the fully connected layer is applied to further extract and process the Transformer model training results, and finally the photovoltaic cluster power prediction model is obtained.

[0010] Furthermore, the specific steps of S1 include:

[0011] S1.1: Collect historical power data of distributed photovoltaic power stations and normalize the collected historical power data to ensure the uniformity of subsequent calculation and processing standards;

[0012] S1.2: Use a time window to extract historical power data features of a certain length. After using Fourier transform to transform the time domain features of the historical power data into frequency domain features, calculate the power similarity between different PV plants using the maximum mutual information coefficient to obtain the historical power correlation matrix.

[0013] S1.3: Slide the time window and repeat the S1.2 process to obtain a series of historical power correlation matrices; arrange the series of matrices according to the time series to obtain a historical power correlation dynamic matrix.

[0014] Furthermore, in the step S2, the specific steps include:

[0015] S2.1: Collect historical meteorological data from distributed photovoltaic power stations, normalize the characteristic sequences of the collected historical meteorological data from distributed photovoltaic power stations, and use the same width time window as in S1.2 to extract the characteristic sequences of the historical meteorological data. Then, use the pseudo-labeling method to label the characteristic sequences of the historical meteorological data to obtain pseudo-labels of meteorological features.

[0016] S2.2: Based on the meteorological feature pseudo-labels constructed in S2.1 and the historical power correlation dynamic matrix constructed in S1.3, the two types of information are integrated using the adaptive mechanism correlation coefficient to construct two types of data correlation feature functions;

[0017] S2.3: Through graph joint aggregation, the meteorological feature pseudo-labels constructed in S2.1 and the historical power correlation dynamic matrix constructed in S1.3 are respectively jointly calculated; the calculation results and the two types of data correlation feature functions constructed in S2.2 are then input into the adaptive hybrid graph filter for meteorological feature screening; finally, the adaptive hybrid graph filter is used to filter out the spatial correlation meteorological feature sequence that is consistent with the spatial variation law of photovoltaic station power.

[0018] Furthermore, the specific steps of S3 include:

[0019] S3.1: Input the normalized historical power data and historical meteorological data obtained after processing in S1.1 and S2.1 into the TCN prediction model, and effectively extract the time series characteristics of each data through the TCN prediction model;

[0020] S3.2: The spatial correlation meteorological feature sequence extracted based on S2.3 is assigned a weight to each spatial correlation meteorological feature through the spatial attention mechanism to achieve the screening of spatial correlation meteorological features. Combined with the TCN feedforward unit, the spatial correlation meteorological feature sequence that matches the TCN prediction model is obtained and output;

[0021] S3.3: Based on the cross-attention mechanism, the time series features of each data obtained by the TCN prediction model in S3.1 are integrated with the spatial correlation meteorological feature sequence obtained in S3.2. The Transformer model is used for training to establish a global dependency relationship between the spatial correlation meteorological feature sequence, the historical meteorological feature sequence, and the historical power data. Finally, the training results of the Transformer model are further processed through the fully connected layer to obtain the photovoltaic cluster power prediction model.

[0022] This method leverages the ability of adaptive hybrid graph filter networks to model complex relationships between photovoltaic power generation facilities, as well as the advantages of TCN-Transformer in processing time series data, to effectively improve the accuracy of photovoltaic cluster power generation forecasts. This new combined model can realize the mining of photovoltaic site cluster characteristics and efficient forecasting, address the dynamic changes in photovoltaic cluster correlation, and improve the accuracy of photovoltaic power forecasts.

[0023] The method of the present invention can realize the mining and efficient prediction of photovoltaic station cluster characteristics, solve the problem of dynamic changes in photovoltaic cluster correlation, improve the accuracy of photovoltaic power prediction, ensure the reliable supply and consumption of new energy, and has promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a structural block diagram of the overall process of the present invention. DETAILED DESCRIPTION

[0025] The following is a detailed description of the embodiments of the present invention with reference to the accompanying drawings. The purpose of the present invention is to provide a photovoltaic cluster prediction method based on a hybrid graph filter and a TCN-Transformer.

[0026] S1: Collect historical power data of distributed photovoltaic power plants, calculate the correlation of historical power data of each photovoltaic power plant based on the time window and maximum mutual information coefficient, and then construct a dynamic matrix of historical power correlation using a sliding time window. The specific steps are as follows:

[0027] S1.1: Collect historical power data of distributed photovoltaic power stations and normalize the collected historical power data to ensure the uniformity of subsequent calculation and processing standards;

[0028] S1.2: Use a time window to extract historical power data features of a certain length. After using Fourier transform to transform the time domain features of the historical power data into frequency domain features, calculate the power similarity between different PV plants using the maximum mutual information coefficient to obtain the historical power correlation matrix.

[0029] S1.3: Slide the time window and repeat the S1.2 process to obtain a series of historical power correlation matrices; arrange the series of matrices according to the time series to obtain a historical power correlation dynamic matrix.

[0030] S2: Collect historical meteorological data from distributed photovoltaic power stations, extract the historical power correlation dynamic matrix and meteorological data features obtained in S1 through an adaptive hybrid graph filter, and construct a new meteorological feature sequence with spatial correlation. The specific steps are as follows:

[0031] S2.1: Collect historical meteorological data from distributed photovoltaic power stations, normalize the characteristic sequences of the collected historical meteorological data from distributed photovoltaic power stations, and use the same width time window as in S1.2 to extract the characteristic sequences of the historical meteorological data. Then, use the pseudo-labeling method to label the characteristic sequences of the historical meteorological data to obtain pseudo-labels of meteorological features.

[0032] S2.2: Based on the meteorological feature pseudo-labels constructed in S2.1 and the historical power correlation dynamic matrix constructed in S1.3, the two types of information are integrated using the adaptive mechanism correlation coefficient to construct two types of data correlation feature functions;

[0033] S2.3: Through graph joint aggregation, the meteorological feature pseudo-labels constructed in S2.1 and the historical power correlation dynamic matrix constructed in S1.3 are respectively jointly calculated; the calculation results and the two types of data correlation feature functions constructed in S2.2 are then input into the adaptive hybrid graph filter for meteorological feature screening; finally, the adaptive hybrid graph filter is used to filter out the spatial correlation meteorological feature sequence that is consistent with the spatial variation law of photovoltaic station power.

[0034] S3: The TCN prediction model is used to extract the feature relationship between the historical power data and historical meteorological data of the distributed photovoltaic power station and output it. The spatial attention mechanism is used to filter the spatially correlated feature information in the new spatially correlated meteorological feature sequence obtained based on S2 and output it. The Transformer model learns the input spatially correlated feature information and assigns weights to implement the Transformer model training. The fully connected layer is then used to further extract and process the Transformer model training results, and finally the photovoltaic cluster power prediction model is obtained. The specific steps are as follows:

[0035] S3.1: Input the normalized historical power data and historical meteorological data obtained after processing in S1.1 and S2.1 into the TCN prediction model, and effectively extract the time series characteristics of each data through the TCN prediction model;

[0036] S3.2: The spatial correlation meteorological feature sequence extracted based on S2.3 is assigned a weight to each spatial correlation meteorological feature through the spatial attention mechanism to achieve the screening of spatial correlation meteorological features. Combined with the TCN feedforward unit, the spatial correlation meteorological feature sequence that matches the TCN prediction model is obtained and output;

[0037] S3.3: Based on the cross-attention mechanism, the time series features of each data obtained by the TCN prediction model in S3.1 are integrated with the spatial correlation meteorological feature sequence obtained in S3.2. The Transformer model is used for training to establish a global dependency relationship between the spatial correlation meteorological feature sequence, the historical meteorological feature sequence, and the historical power data. Finally, the training results of the Transformer model are further processed through the fully connected layer to obtain the photovoltaic cluster power prediction model.

Claims

1. A photovoltaic cluster prediction method based on hybrid graph filter and TCN-Transformer, characterized by Follow these steps: S1: Collect historical power data of distributed photovoltaic power stations, calculate the correlation of historical power data of each photovoltaic power station based on the time window and maximum mutual information coefficient, and then construct a dynamic matrix of historical power correlation by sliding the time window; S2: Collect historical meteorological data of distributed photovoltaic power stations, extract the historical power correlation dynamic matrix and meteorological data features obtained based on S1 through an adaptive hybrid graph filter, and construct a new meteorological feature sequence with spatial correlation; S3: The TCN prediction model is used to extract the characteristic relationship between the historical power data and historical meteorological data of distributed photovoltaic power stations and then output it; the spatial attention mechanism is used to filter the spatially correlated feature information in the new spatially correlated meteorological feature sequence obtained based on S2 and then output it; the Transformer model learns the input spatially correlated feature information and assigns weights to realize the training of the Transformer model; then the fully connected layer is applied to further extract and process the Transformer model training results, and finally the photovoltaic cluster power prediction model is obtained.

2. The photovoltaic cluster prediction method based on hybrid graph filter and TCN-Transformer according to claim 1 is characterized in that: In the step S1, the specific steps include: S1.1: Collect historical power data of distributed photovoltaic power stations and normalize the collected historical power data to ensure the uniformity of subsequent calculation and processing standards; S1.2: Use a time window to extract historical power data features of a certain length. After using Fourier transform to transform the time domain features of the historical power data into frequency domain features, calculate the power similarity between different PV plants using the maximum mutual information coefficient to obtain the historical power correlation matrix. S1.3: Slide the time window and repeat the S1.2 process to obtain a series of historical power correlation matrices; arrange the series of matrices according to the time series to obtain a historical power correlation dynamic matrix.

3. The photovoltaic cluster prediction method based on hybrid graph filter and TCN-Transformer according to claim 2 is characterized in that: In the step S2, the specific steps include: S2.1: Collect historical meteorological data from distributed photovoltaic power stations, normalize the characteristic sequences of the collected historical meteorological data from distributed photovoltaic power stations, and use the same width time window as in S1.2 to extract the characteristic sequences of the historical meteorological data. Then, use the pseudo-labeling method to label the characteristic sequences of the historical meteorological data to obtain pseudo-labels of meteorological features. S2.2: Based on the meteorological feature pseudo-labels constructed in S2.1 and the historical power correlation dynamic matrix constructed in S1.3, the two types of information are integrated using the adaptive mechanism correlation coefficient to construct two types of data correlation feature functions; S2.3: Through graph joint aggregation, the meteorological feature pseudo-labels constructed in S2.1 and the historical power correlation dynamic matrix constructed in S1.3 are respectively jointly calculated; the calculation results and the two types of data correlation feature functions constructed in S2.2 are then input into the adaptive hybrid graph filter for meteorological feature screening; finally, the adaptive hybrid graph filter is used to filter out the spatial correlation meteorological feature sequence that is consistent with the spatial variation law of photovoltaic station power.

4. The photovoltaic cluster prediction method based on hybrid graph filter and TCN-Transformer according to claim 3 is characterized in that: In the step S3, the specific steps include: S3.1: Input the normalized historical power data and historical meteorological data obtained after processing in S1.1 and S2.1 into the TCN prediction model, and effectively extract the time series characteristics of each data through the TCN prediction model; S3.2: The spatial correlation meteorological feature sequence extracted based on S2.3 is assigned a weight to each spatial correlation meteorological feature through the spatial attention mechanism to achieve the screening of spatial correlation meteorological features. Combined with the TCN feedforward unit, the spatial correlation meteorological feature sequence that matches the TCN prediction model is obtained and output; S3.3: Based on the cross-attention mechanism, the time series features of each data obtained by the TCN prediction model in S3.1 are integrated with the spatial correlation meteorological feature sequence obtained in S3.

2. The Transformer model is used for training to establish a global dependency relationship between the spatial correlation meteorological feature sequence, the historical meteorological feature sequence, and the historical power data. Finally, the training results of the Transformer model are further processed through the fully connected layer to obtain the photovoltaic cluster power prediction model.

Images