A new energy power supply guarantee curve prediction model construction method

Abstract

The application discloses a new energy power supply guarantee curve prediction model construction method; abnormal data is identified and repaired based on DBSCAN-TimaGAN; an NWP correction model is constructed based on ERA5 data; a training set power data is time series decomposed based on STL, wherein a periodic component is reduced in dimension through KernelPCA, decomposed in a Gabor frequency domain, TimesNet modules are respectively constructed according to different frequencies, and each module output is integrated through a multilayer perceptron; a trend component is multi-step predicted based on Autoformer after being filtered through a Gaussian low-pass filter; a residual component is predicted through a KAN model after extracting meteorological-residual sensitive features; the three component prediction models are integrated through Voting, the model is trained based on a dynamic loss function and a Lion optimizer, and model hyperparameters are optimized based on a Remora optimization algorithm, and finally, a new energy power supply guarantee curve prediction model is obtained. Through the method, new energy power supply guarantee curve prediction can be realized, a reference can be provided for power dispatching departments to formulate schemes, and the method has certain popularization value.

Description

Technical Field

[0001] The present invention relates to a method for constructing a new energy power supply guarantee curve prediction model, and belongs to the field of new energy power prediction. Background Art

[0002] With the construction of new power systems driven by the increasing proportion of renewable energy under the "dual carbon" goals, the inherent randomness and uncertainty of renewable energy pose challenges to the grid's ability to ensure power supply and promote energy consumption. Current research on renewable energy power forecasting focuses primarily on improving forecast accuracy. However, as the proportion of renewable energy in the power system continues to rise, if the forecasted renewable energy power significantly exceeds the actual power generation, and the grid dispatching department has already made dispatch plans based on this forecast, this may lead to power shortages and the inability to adjust backup power in a timely manner, posing challenges to the real-time balance of the power system. Conversely, if the forecast overemphasizes power supply security and uses a large proportional reduction factor for the forecasted power, the forecasted renewable energy power will consistently fall short of the actual power generation, potentially leading to an increase in the proportion of renewable energy curtailment. Errors in renewable energy power forecasts are unavoidable, so dispatching departments are prioritizing improving both forecast accuracy and the reliability of the forecasted power to support the formulation of dispatch plans and maximize the absorption of renewable energy generation while ensuring power supply reliability.

[0003] Currently, there is no technology for predicting the curve of new energy power supply security, and the existing new energy power technology is difficult to meet the actual production needs of the power system dispatching department. Therefore, this project proposes a new energy power supply security curve prediction method to provide technical support for the formulation of dispatching plans. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for constructing a new energy power supply guarantee curve prediction model to solve the problems existing in the above background technology.

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

[0006] A method for constructing a new energy power supply guarantee curve prediction model is characterized by following the steps below:

[0007] S1: Collect NWP data and power data of each station in the target new energy cluster as a sample data set, and identify and repair outliers and missing value anomalies in the data based on DBSCAN (Density-Based Spatial Clustering of Applications with Noise)-TimeGAN (Time-series Generative Adversarial Networks); then build an NWP correction model for each station in the cluster based on ERA5 (European Meteorological Center atmospheric reanalysis data, ECMWF Reanalysis version 5) atmospheric reanalysis data, and accumulate the power data of each station to build a revised sample set consisting of the cluster's new energy historical power data and cluster new energy historical NWP data. Then, divide the revised sample set into the training set and test set of the new energy supply curve modeling dataset;

[0008] S2: Based on STL, the training set power data obtained in S1 is decomposed into time series. The decomposed periodic components are subjected to high-dimensional nonlinear dimensionality reduction using KernelPCA (principal component analysis based on kernel function), Gabor (Gabor transform) frequency domain decomposition, and two-dimensional transformation. TimesNet modules are then constructed according to different frequencies. The outputs of each module are then integrated based on a multi-layer perceptron to form a periodic component prediction model. The trend component is subjected to Gaussian low-pass filtering and a trend component prediction model is constructed based on an Autoformer. The residual component is subjected to a KAN model after extracting meteorological-residual sensitive features to construct a residual component prediction model.

[0009] S3: Based on the periodic component prediction model, trend component prediction model, and residual component prediction model obtained in S2, an initial new energy power supply security curve prediction model is constructed based on the Voting integrated model. The dynamic loss function is then used as the target loss function, and the initial new energy power supply security curve prediction model is trained using the Lion optimizer. The performance of the trained initial new energy power supply security curve prediction model is evaluated based on the dynamic loss function, and the hyperparameters of the trained new energy power supply security curve prediction model are optimized using the Remora optimization algorithm. The optimized new energy power supply security curve prediction model is retrained and performance evaluated again. This process is repeated until the maximum number of iterations of the optimization algorithm is reached, and the final new energy power supply security curve prediction model is saved.

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

[0011] S1.1: Collect NWP and power data from each station in the target new energy cluster, with a temporal resolution of 15 minutes. Identify outliers by inputting the NWP-power data into a DBSCAN clustering model. Train a TimeGAN model using normal NWP and power data. The trained TimeGAN model then inpaints NWP-power data with outliers and missing values.

[0012] S1.2: Collect ERA5 atmospheric reanalysis data for the target new energy cluster region. Compute the probability density distribution (PDF) of the corresponding parameters between the ERA5 atmospheric reanalysis data and the NWP data collected by each station within the target new energy cluster. Build a revised NWP model for each station within the cluster based on the statistical differences. Then, accumulate the power data for each station within the cluster to obtain the historical NWP data and historical power data for the cluster's new energy resources. These two data sets constitute the revised sample set for the new energy cluster.

[0013] S1.3: Divide the revised sample set of the new energy cluster according to the ratio of training set to test set of 7:3 to obtain the training set and test set of the new energy power supply guarantee curve modeling data set.

[0014] Furthermore, the specific steps of S2 include:

[0015] S2.1: Perform STL time series decomposition on the training set power data obtained in S1.3 to obtain periodic components, trend components, and residual components. The periodic components and the training set meteorological data are subjected to high-dimensional nonlinear dimensionality reduction using KernelPCA. Then, frequency domain decomposition is performed based on Gabor transformation. The decomposed data of different frequencies are converted into two-dimensional data according to period folding. The converted data are padded to obtain two-dimensional time series data of consistent length. TimesNet modules are constructed based on the two-dimensional time series data of different frequencies. Finally, each TimesNet module is connected through a multi-layer perceptron to obtain a periodic component prediction model.

[0016] S2.2: Perform a Gaussian low-pass filter on the trend component decomposed in S2.1 to obtain a smoothed trend component of the historical power data. This trend component is converted into a two-dimensional data structure using a sliding window, and a trend component prediction model is constructed using the Autoformer model.

[0017] S2.3 performs correlation analysis on the residual components decomposed in S2.1 and the various meteorological characteristic parameters of the training set meteorological data obtained in S1.3, extracts the residual sensitive features, and constructs a residual component prediction model based on KAN (Kolmogorov-Arnold Networks).

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

[0019] S3.1: The period component prediction model, trend component prediction model and residual component prediction model constructed in S2.1, S2.2 and S2.3 are used to construct the initial new energy power supply guarantee curve prediction model through the Voting integration model;

[0020] S3.2: Construct a dynamic loss function loss; use this dynamic loss function loss as the target loss function of the Lion optimizer, use the training set meteorological data obtained in S1.3 as features, and the training set power data as labels to train the initial new energy power supply guarantee curve prediction model;

[0021]

[0022] Where, α is the error direction coefficient, and its value range is (1, +∞); β is the output amplitude coefficient, and its value range is (1, +∞); y pred is the power prediction value, y actual is the actual value of power, y pred and y actual The value range is [0,1];

[0023] S3.3: Use the dynamic loss function loss as the evaluation indicator and use the test set data obtained in S1.3 to evaluate the performance of the trained new energy power supply security curve prediction model. Optimize the number of neurons, regularization coefficient, number of network layers, and learning rate hyperparameters of the trained new energy power supply security curve prediction model based on the Remora optimization algorithm. Retrain the optimized new energy power supply security curve prediction model and perform performance evaluation again. Repeat this process until the Remora optimization algorithm reaches the maximum number of iterations, and save the final iterative new energy power supply security curve prediction model.

[0024] The present invention saves the final new energy power supply guarantee curve prediction model and outputs the evaluation results of the final new energy power supply guarantee curve prediction model. The target cluster NWP data is input into the final new energy power supply guarantee curve prediction model, and the corresponding new energy power supply guarantee curve for the time period is output.

[0025] The method of the present invention can realize the prediction of new energy power supply guarantee curve, provide a reference for the power dispatching department to formulate plans, and provide technical support for improving the operational stability of the power system, which has certain promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a structural block diagram of the overall process of the present invention. DETAILED DESCRIPTION

[0027] 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 method for constructing a new energy power supply guarantee curve prediction model.

[0028] like Figure 1 As shown, a method for constructing a new energy power supply guarantee curve prediction model is characterized by following the steps below:

[0029] S1: Collect NWP data and power data from each station in the target new energy cluster as a sample data set, and identify and repair outliers and missing values ​​in the data based on DBSCAN-TimeGAN; then build an NWP correction model for each station in the cluster based on ERA5 atmospheric reanalysis data, and accumulate the power data of each station to build a revised sample set consisting of the cluster's new energy historical power data and cluster new energy historical NWP data. The revised sample set is then divided into the training set and test set of the new energy supply curve modeling dataset; the specific steps include:

[0030] S1.1: Collect NWP and power data from each station in the target new energy cluster, with a temporal resolution of 15 minutes. Identify outliers by inputting the NWP-power data into a DBSCAN clustering model. Train a TimeGAN model using normal NWP and power data. The trained TimeGAN model then inpaints NWP-power data with outliers and missing values.

[0031] S1.2: Collect ERA5 atmospheric reanalysis data for the target new energy cluster region. Compute the probability density distribution (PDF) of the corresponding parameters between the ERA5 atmospheric reanalysis data and the NWP data collected by each station within the target new energy cluster. Build a revised NWP model for each station within the cluster based on the statistical differences. Then, accumulate the power data for each station within the cluster to obtain the historical NWP data and historical power data for the cluster's new energy resources. These two data sets constitute the revised sample set for the new energy cluster.

[0032] S1.3: Divide the revised sample set of the new energy cluster according to the ratio of training set to test set of 7:3 to obtain the training set and test set of the new energy power supply guarantee curve modeling dataset;

[0033] S2: Based on STL, the training set power data obtained in S1 is decomposed into time series. The decomposed periodic components are subjected to KernelPCA high-dimensional nonlinear dimensionality reduction, Gabor frequency domain decomposition, and two-dimensional transformation. TimesNet modules are constructed according to different frequencies. The outputs of each module are then integrated based on a multi-layer perceptron to form a periodic component prediction model. The trend component is subjected to Gaussian low-pass filtering and a trend component prediction model is constructed based on the Autoformer. The residual component is subjected to the KAN model after extracting meteorological-residual sensitive features to construct a residual component prediction model. The specific steps include:

[0034] S2.1: Perform STL time series decomposition on the training set power data obtained in S1.3 to obtain periodic components, trend components, and residual components. The periodic components and the training set meteorological data are subjected to high-dimensional nonlinear dimensionality reduction using KernelPCA. Then, frequency domain decomposition is performed based on Gabor transformation. The decomposed data of different frequencies are converted into two-dimensional data according to period folding. The converted data are padded to obtain two-dimensional time series data of consistent length. TimesNet modules are constructed based on the two-dimensional time series data of different frequencies. Finally, each TimesNet module is connected through a multi-layer perceptron to obtain a periodic component prediction model.

[0035] S2.2: Perform a Gaussian low-pass filter on the trend component decomposed in S2.1 to obtain a smoothed trend component of the historical power data. This trend component is converted into a two-dimensional data structure using a sliding window, and a trend component prediction model is constructed using the Autoformer model.

[0036] S2.3 performs correlation analysis on the residual components decomposed in S2.1 and the various meteorological characteristic parameters of the training set meteorological data obtained in S1.3, extracts the residual sensitive features, and constructs a residual component prediction model based on KAN (Kolmogorov–Arnold Networks);

[0037] S3: Based on the periodic component prediction model, trend component prediction model, and residual component prediction model obtained in S2, an initial new energy power supply security curve prediction model is constructed based on the Voting integrated model. The dynamic loss function is then used as the target loss function to train the initial new energy power supply security curve prediction model through the Lion optimizer. The performance of the trained new energy power supply security curve prediction model is evaluated based on the dynamic loss function, and the hyperparameters of the trained new energy power supply security curve prediction model are optimized through the Remora (voting) optimization algorithm. The optimized new energy power supply security curve prediction model is retrained and performance evaluated again. This process is repeated until the maximum number of iterations of the optimization algorithm is reached, and the final new energy power supply security curve prediction model of the iteration is saved. The specific steps include:

[0038] S3.1: The period component prediction model, trend component prediction model and residual component prediction model constructed in S2.1, S2.2 and S2.3 are used to construct the initial new energy power supply guarantee curve prediction model through the Voting integration model;

[0039] S3.2: Construct a dynamic loss function loss; use this dynamic loss function loss as the target loss function of the Lion optimizer, use the training set meteorological data obtained in S1.3 as features, and the training set power data as labels to train the initial new energy power supply guarantee curve prediction model;

[0040]

[0041] Where α is the error direction coefficient, and its value range is (1, +∞). The closer the value is to 1, the more sensitive the new energy power supply guarantee curve prediction model is to the error direction; β is the output amplitude coefficient, and its value range is (1, +∞). The larger the value is, the more sensitive the new energy power supply guarantee curve prediction model is to the output amplitude; y pred is the power prediction value, y actual is the actual power value. Since the power is normalized based on the installed capacity, y pred and y actual The value range is [0,1]; in order to minimize y as much as possible while ensuring the accuracy of prediction pred Greater than y actual Forward error, α and β are usually 2 and 1.5 respectively;

[0042] S3.3: Use the dynamic loss function loss as the evaluation indicator and use the test set data obtained in S1.3 to evaluate the performance of the trained new energy power supply security curve prediction model. Optimize the number of neurons, regularization coefficient, number of network layers, and learning rate hyperparameters of the trained new energy power supply security curve prediction model based on the Remora optimization algorithm. Retrain the optimized new energy power supply security curve prediction model and perform performance evaluation again. Repeat this process until the Remora optimization algorithm reaches the maximum number of iterations, and save the final iterative new energy power supply security curve prediction model.

[0043] The final new energy power supply security curve prediction model is saved, and the evaluation results of the final new energy power supply security curve prediction model are output; the target cluster NWP data is input into the final new energy power supply security curve prediction model, and the corresponding new energy power supply security curve for the time period is output.

Claims

1. A method for constructing a new energy power supply guarantee curve prediction model, characterized in that Follow these steps: S1: Collect NWP data and power data from each station in the target new energy cluster as a sample data set, and identify and repair outliers and missing values ​​in the data based on DBSCAN-TimeGAN; Then, based on the ERA5 atmospheric reanalysis data, a NWP correction model is constructed for each station in the cluster. The power data of each station is accumulated to construct a revised sample set consisting of the cluster's new energy historical power data and the cluster's new energy historical NWP data. The revised sample set is then divided into the training set and test set of the new energy supply curve modeling dataset. S2: Based on STL, the training set power data obtained in S1 is decomposed into time series. The decomposed periodic components are subjected to KernelPCA high-dimensional nonlinear dimensionality reduction, Gabor frequency domain decomposition, and two-dimensional transformation. TimesNet modules are constructed according to different frequencies. The outputs of each module are then integrated based on a multi-layer perceptron to form a periodic component prediction model. The trend component is subjected to Gaussian low-pass filtering and a trend component prediction model is constructed based on the Autoformer. The residual component is subjected to the KAN model after extracting meteorological-residual sensitive features to construct a residual component prediction model. S3: Based on the periodic component prediction model, trend component prediction model and residual component prediction model obtained in S2, an initial new energy power supply security curve prediction model is constructed based on the Voting integrated model. The dynamic loss function is then used as the target loss function, and the initial new energy power supply security curve prediction model is trained through the Lion optimizer; the performance of the trained new energy power supply security curve prediction model is evaluated based on the dynamic loss function, and the hyperparameters of the trained new energy power supply security curve prediction model are optimized through the Remora optimization algorithm. The optimized new energy power supply security curve prediction model is retrained and performance evaluated again. The process is repeated until the maximum number of iterations of the optimization algorithm is reached, and the final new energy power supply security curve prediction model of the iteration is saved.

2. According to the method for constructing a new energy power supply guarantee curve prediction model according to claim 1, it is characterized in that: The S1 step specifically includes: S1.1: Collect NWP and power data from each station in the target new energy cluster, with a temporal resolution of 15 minutes. Identify outliers by inputting the NWP-power data into a DBSCAN clustering model. Train a TimeGAN model using normal NWP and power data. The trained TimeGAN model then inpaints NWP-power data with outliers and missing values. S1.2: Collect ERA5 atmospheric reanalysis data for the target new energy cluster region. Compute the probability density distribution (PDF) of the corresponding parameters between the ERA5 atmospheric reanalysis data and the NWP data collected by each station within the target new energy cluster. Build a revised NWP model for each station within the cluster based on the statistical differences. Then, accumulate the power data for each station within the cluster to obtain the historical NWP data and historical power data for the cluster's new energy resources. These two data sets constitute the revised sample set for the new energy cluster. S1.3: Divide the revised sample set of the new energy cluster according to the ratio of training set to test set of 7:3 to obtain the training set and test set of the new energy power supply guarantee curve modeling data set.

3. According to the method for constructing a new energy power supply guarantee curve prediction model according to claim 2, it is characterized in that: The S2 step specifically includes: S2.1: Perform STL time series decomposition on the training set power data obtained in S1.3 to obtain periodic components, trend components, and residual components. The periodic components and the training set meteorological data are subjected to high-dimensional nonlinear dimensionality reduction using KernelPCA. Then, frequency domain decomposition is performed based on Gabor transformation. The decomposed data of different frequencies are converted into two-dimensional data according to period folding. The converted data are padded to obtain two-dimensional time series data of consistent length. TimesNet modules are constructed based on the two-dimensional time series data of different frequencies. Finally, each TimesNet module is connected through a multi-layer perceptron to obtain a periodic component prediction model. S2.2: Perform a Gaussian low-pass filter on the trend component decomposed in S2.1 to obtain a smoothed trend component of the historical power data. This trend component is converted into a two-dimensional data structure using a sliding window, and a trend component prediction model is constructed using the Autoformer model. S2.3 performs correlation analysis on the residual components decomposed in S2.1 and the various meteorological characteristic parameters of the training set meteorological data obtained in S1.3, extracts the residual sensitive features, and constructs a residual component prediction model based on KAN.

4. The method for constructing a new energy power supply guarantee curve prediction model according to claim 3, characterized in that: The S3 step is specifically as follows: S3.1: The period component prediction model, trend component prediction model and residual component prediction model constructed in S2.1, S2.2 and S2.3 are used to construct the initial new energy power supply guarantee curve prediction model through the Voting integration model; S3.2: Construct dynamic loss function lass; The dynamic loss function loss is used as the target loss function of the Lion optimizer. The training set meteorological data obtained in S1.3 is used as features, and the training set power data is used as labels to train the initial new energy power supply guarantee curve prediction model. Where, α is the error direction coefficient, and its value range is (1, +∞); β is the output amplitude coefficient, and its value range is (1, +∞); y pred is the power prediction value, y actual is the actual value of power, y pred and y actual The value range is [0,1]; S3.3: Use the dynamic loss function lass as the evaluation indicator and use the test set data obtained in S1.3 to evaluate the performance of the trained new energy power supply security curve prediction model. Optimize the number of neurons, regularization coefficient, number of network layers, and learning rate hyperparameters of the trained new energy power supply security curve prediction model based on the Remora optimization algorithm. Retrain the optimized new energy power supply security curve prediction model and perform performance evaluation again. Repeat this process until the Remora optimization algorithm reaches the maximum number of iterations, and save the final new energy power supply security curve prediction model of the iteration.

5. A method for constructing a new energy power supply guarantee curve prediction model according to claim 4, characterized in that: The target cluster NWP data is input into the final new energy power supply security curve prediction model, and the corresponding new energy power supply security curve for that time period is output.

Images