Optimization method and system for training sample of wind power plant power prediction model

Abstract

The invention provides an optimization method and system for training samples of a wind power plant power prediction model, relates to the technical field of sample adjustment and optimization, and aims to improve the sample learning ability of the model in power prediction, comprising the steps of sampling training data to obtain a plurality of sampling samples for a tree model; training by adopting the samples based on wind speed meteorological characteristics through a tree model; obtaining a prediction value of each sample, and calculating a prediction error absolute value; carrying out equal interval division on the sampling sample, and dividing the sampling sample into a large-probability scene sample and a small-probability scene sample; different sample weights are given to the different samples respectively; performing first-stage training through a deep learning model based on the sample weight; gradually adjusting the sample weights, and carrying out second-stage training through a deep learning model until the sample weights of all samples are the same; the method has the advantage of reducing model identification errors caused by different sample distribution.

Description

Technical Field

[0001] The present invention relates to the technical field of sample adjustment optimization, and in particular to an optimization method and system for training samples of a wind power plant power prediction model. Background Art

[0002] Power forecasting is a common problem faced by wind power plants.

[0003] The distribution of wind power plant power generation data is a long-tail distribution. When sample balancing is not performed, the low-power samples account for too large a proportion, while the high-power samples and low-probability scenario samples account for a small proportion in the calculation of the model loss function. This causes the prediction model to over-learn the low-power range and common scenarios, which in turn leads to two problems: the model prediction value is therefore overall low, resulting in the predicted daily power generation being lower than the actual power generation average; the model's learning level of low-probability scenario data is weak, resulting in insufficient utilization of meteorological characteristics, which in turn leads to poor prediction results in low-probability scenarios.

[0004] Therefore, it is necessary to optimize the data samples to improve the model's learning ability for samples in power prediction. Summary of the invention

[0005] The object of the present invention is to provide a method and system for optimizing training samples of a wind power plant power prediction model, which can improve the model's learning ability for samples in power prediction.

[0006] The present invention is achieved through the following technical solutions:

[0007] The optimization method for training samples of a wind power plant power prediction model comprises the following steps:

[0008] Sampling the training data to obtain multiple sampling samples for the tree model;

[0009] Using the sampled samples to perform training based on wind speed meteorological characteristics through a tree model;

[0010] Obtaining the predicted value of each sample, and calculating the absolute value of the prediction error between the predicted value of each sample and the corresponding true value;

[0011] Divide the sampled samples into equal intervals according to the absolute value of the prediction error, into samples of a high-probability scenario and samples of a low-probability scenario, wherein the sampled samples in the low-probability scenario have a larger absolute value of the prediction error;

[0012] Assigning different sample weights to different sampling samples in the high-probability scenario samples and the low-probability scenario samples respectively;

[0013] Performing first-stage training through a deep learning model based on the sample weights;

[0014] The sample weights are gradually adjusted to perform second-stage training through a deep learning model until the sample weights of the respective sampled samples are the same.

[0015] Preferably, the sampling method is stratified sampling, systematic sampling or cluster sampling.

[0016] Preferably, the method of training the tree model using sampling samples based on wind speed meteorological characteristics is:

[0017] Collecting data sets through historical transaction records;

[0018] Dividing the data set into a data set and a test set by random sampling;

[0019] Using the training set to train a tree model, and using the test set to evaluate the performance of the trained tree model;

[0020] The prediction error of the trained tree model is evaluated using the data set.

[0021] Preferably, the ratio of the data set to the test set is 7:3.

[0022] Preferably, the method of dividing the sampled samples into equal intervals according to the absolute value of the prediction error is:

[0023] Preset the prediction error absolute value threshold;

[0024] If the absolute value of the prediction error between the predicted value of the sample and the corresponding true value is greater than the prediction error absolute value threshold, the sample is classified as the low-probability scenario sample, otherwise it is classified as the high-probability scenario sample.

[0025] Preferably, the method of assigning different sample weights to different sampled samples in the high-probability scene samples and the low-probability scene samples is:

[0026] Assign larger sample weights to small probability scene samples to increase the learning of small probability scene samples.

[0027] Preferably, the method of gradually adjusting the sample weights is:

[0028] The sample weights of the sampling samples with larger sample weights are gradually reduced until the sample weights of the sampling samples are the same.

[0029] The present invention also provides an optimization system for training samples of a wind power plant power prediction model, which is applied to any one of the above-mentioned optimization methods for training samples of a wind power plant power prediction model, and comprises:

[0030] A sampling module is used to sample the training data to obtain multiple sampling samples for the tree model;

[0031] A tree model training module, used for training the tree model based on wind speed meteorological characteristics using the sampled samples;

[0032] A prediction error absolute value acquisition module is used to obtain the prediction value of each sample, and calculate the prediction error absolute value between the prediction value of each sample and the corresponding true value;

[0033] A probability scene division module, used for dividing the sampled samples into equal intervals according to the absolute value of the prediction error, into high probability scene samples and low probability scene samples, wherein the sampled samples in the low probability scene samples have a larger absolute value of the prediction error;

[0034] A weight assigning module, used to assign different sample weights to different sampled samples in the high probability scenario samples and the low probability scenario samples;

[0035] A first-stage training module, used for performing first-stage training through a deep learning model based on the sample weights;

[0036] The second stage training module is used to gradually adjust the sample weights and perform second stage training through a deep learning model until the sample weights of the various sampled samples are the same.

[0037] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0038] The present invention corrects the prediction variable deviation under power prediction by processing sample data, thereby reducing the prediction deviation;

[0039] The present invention can capture the diversity of data by sampling the training data multiple times and using the tree model for training, reduce the overfitting of the model to individual samples, and improve the robustness and generalization ability of the model;

[0040] By giving higher weights to samples of low-probability scenarios, the model can focus more on learning such samples, avoid these difficult-to-predict samples from excessively affecting the overall performance of the model, and improve its adaptability to extreme or rare events.

[0041] After assigning different weights, the present invention gradually makes the weights of all samples the same through training in other stages, ultimately ensuring more comprehensive learning of the model;

[0042] The present invention is reasonably designed, simple to establish, and suitable for power prediction in various scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1A schematic flow chart of a method for optimizing training samples of a wind power plant power prediction model provided in Embodiment 1 of the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0045] Example 1

[0046] This embodiment provides an optimization method for training samples of wind power plant power prediction model, see Figure 1 , including the following steps:

[0047] Sampling the training data to obtain multiple sampling samples for the tree model;

[0048] Using the sampled samples to perform training based on wind speed meteorological characteristics through a tree model;

[0049] Obtaining the predicted value of each sample, and calculating the absolute value of the prediction error between the predicted value of each sample and the corresponding true value;

[0050] Divide the sampled samples into equal intervals according to the absolute value of the prediction error, into samples of a high-probability scenario and samples of a low-probability scenario, wherein the sampled samples in the low-probability scenario have a larger absolute value of the prediction error;

[0051] Assigning different sample weights to different sampling samples in the high-probability scenario samples and the low-probability scenario samples respectively;

[0052] Performing first-stage training through a deep learning model based on the sample weights;

[0053] The sample weights are gradually adjusted to perform second-stage training through a deep learning model until the sample weights of the respective sampled samples are the same.

[0054] In this embodiment, the sampling method is stratified sampling, systematic sampling or cluster sampling.

[0055] As a preferred solution of this embodiment, the method of training by using a tree model based on wind speed meteorological characteristics using sampling samples is:

[0056] Collecting data sets through historical transaction records;

[0057] Dividing the data set into a data set and a test set by random sampling;

[0058] Using the training set to train a tree model, and using the test set to evaluate the performance of the trained tree model;

[0059] The prediction error of the trained tree model is evaluated using the data set.

[0060] Furthermore, the ratio of the data set to the test set is 7:3.

[0061] Next, the method of dividing the sample into equal intervals according to the absolute value of the prediction error is preferably:

[0062] Preset the prediction error absolute value threshold;

[0063] If the absolute value of the prediction error between the predicted value of the sample and the corresponding true value is greater than the prediction error absolute value threshold, the sample is classified as the low-probability scenario sample, otherwise it is classified as the high-probability scenario sample.

[0064] Specifically, the method of assigning different sample weights to different sampled samples in the high-probability scene samples and the low-probability scene samples may be:

[0065] Assign larger sample weights to small probability scene samples to increase the learning of small probability scene samples.

[0066] On the other hand, the method of gradually adjusting the sample weights is:

[0067] The sample weights of the sampling samples with larger sample weights are gradually reduced until the sample weights of the sampling samples are the same.

[0068] This embodiment improves the robustness and generalization ability of the model through multiple sampling, prediction error division and dynamic adjustment of sample weights. At the same time, through differentiated processing of samples under different probability scenarios, it can optimize the prediction effect and adapt to diverse data scenarios. The reason for extracting wind speed meteorological features is that such features are generally the main factor affecting power consumption.

[0069] It should be noted in particular that, in actual implementation, when different sample weights are assigned to different sampling samples in the high-probability scenario samples and the low-probability scenario samples, a solution that can be adopted may be, for example, to assign a uniform sample weight a to different sampling samples in the low-probability scenario samples, and to assign a uniform sample weight b to different sampling samples in the high-probability scenario samples, where a>b.

[0070] It is also possible to assign a value to each sampling sample separately, for example, the average sample weight of different sampling samples in the low probability scenario sample is set as a, and the average sample weight of different sampling samples in the high probability scenario sample is set as b.

[0071] For the different sampled samples in the small probability scene samples, the sampled samples are sorted from large to small according to the absolute value of the prediction error. Assuming the sequence is 1-n, the sample weight ρ of the sampled sample in the small probability scene sample numbered i is u for:

[0072] ρ i =ρ1*t i-1 ,i∈[2,n];

[0073] Where t is the proportional coefficient that can be set by yourself, and ρ1 can be solved by the following formula:

[0074]

[0075] The sample weights in the high-probability scenario samples can be obtained in the same way as the above method.

[0076] On the other hand, when gradually reducing the sample weights at the end, the reduction method for each step can be:

[0077] Arrange all sample weights, select the minimum sample weight as the benchmark, and assume that the minimum value is the kth sample weight ω k ;

[0078] Set the total number of steps to reduce m;

[0079] ω s The value of the sth sample weight, in the qth step adjustment, the sth sample weight ω s,q The adjustment value is:

[0080]

[0081] Example 2

[0082] This embodiment provides an optimization system for training samples of a wind power plant power prediction model, which is applied to the optimization method for training samples of a wind power plant power prediction model in the above embodiment, including:

[0083] A sampling module is used to sample the training data to obtain multiple sampling samples for the tree model;

[0084] A tree model training module, used for training the tree model based on wind speed meteorological characteristics using the sampled samples;

[0085] A prediction error absolute value acquisition module is used to obtain the prediction value of each sample, and calculate the prediction error absolute value between the prediction value of each sample and the corresponding true value;

[0086] A probability scene division module, used for dividing the sampled samples into equal intervals according to the absolute value of the prediction error, into high probability scene samples and low probability scene samples, wherein the sampled samples in the low probability scene samples have a larger absolute value of the prediction error;

[0087] A weight assigning module, used to assign different sample weights to different sampled samples in the high probability scenario samples and the low probability scenario samples;

[0088] A first-stage training module, used for performing first-stage training through a deep learning model based on the sample weights;

[0089] The second stage training module is used to gradually adjust the sample weights and perform second stage training through a deep learning model until the sample weights of the various sampled samples are the same.

[0090] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for optimizing training samples of a wind power plant power prediction model, characterized in that: The following steps are involved: Sampling the training data to obtain multiple sampling samples for the tree model; Using the sampled samples to perform training based on wind speed meteorological characteristics through a tree model; Obtaining the predicted value of each sample, and calculating the absolute value of the prediction error between the predicted value of each sample and the corresponding true value; Divide the sampled samples into equal intervals according to the absolute value of the prediction error, into samples of a high-probability scenario and samples of a low-probability scenario, wherein the sampled samples in the low-probability scenario have a larger absolute value of the prediction error; Assigning different sample weights to different sampling samples in the high-probability scenario samples and the low-probability scenario samples respectively; Performing first-stage training through a deep learning model based on the sample weights; The sample weights are gradually adjusted to perform second-stage training through a deep learning model until the sample weights of the respective sampled samples are the same.

2. The optimization method for training samples of wind power plant power prediction model according to claim 1, characterized in that: The sampling method is stratified sampling, systematic sampling or cluster sampling.

3. The optimization method for training samples of wind power plant power prediction model according to claim 1, characterized in that: The method of training by using a tree model based on wind speed meteorological characteristics using sampling samples is: Collecting data sets through historical transaction records; Dividing the data set into a data set and a test set by random sampling; Using the training set to train a tree model, and using the test set to evaluate the performance of the trained tree model; The prediction error of the trained tree model is evaluated using the data set.

4. The optimization method for training samples of wind power plant power prediction model according to claim 3, characterized in that: The ratio of the data set to the test set is 7:

3.

5. The optimization method for training samples of wind power plant power prediction model according to claim 1, characterized in that: The method for dividing the sample into equal intervals according to the absolute value of the prediction error is: Preset the prediction error absolute value threshold; If the absolute value of the prediction error between the predicted value of the sample and the corresponding true value is greater than the prediction error absolute value threshold, the sample is classified as the low-probability scenario sample, otherwise it is classified as the high-probability scenario sample.

6. The optimization method for training samples of wind power plant power prediction model according to claim 1, characterized in that: The method of assigning different sample weights to different sampled samples in the high-probability scene samples and the low-probability scene samples is: Assign larger sample weights to small probability scene samples to increase the learning of small probability scene samples.

7. The optimization method for training samples of wind power plant power prediction model according to claim 1, characterized in that: The method of gradually adjusting the sample weights is: The sample weights of the sampling samples with larger sample weights are gradually reduced until the sample weights of the sampling samples are the same.

8. An optimization system for training samples of a wind power plant power prediction model, applied to the optimization method for training samples of a wind power plant power prediction model according to any one of claims 1 to 7, characterized in that: include: A sampling module is used to sample the training data to obtain multiple sampling samples for the tree model; A tree model training module, used for training the tree model based on wind speed meteorological characteristics using the sampled samples; A prediction error absolute value acquisition module is used to obtain the prediction value of each sample, and calculate the prediction error absolute value between the prediction value of each sample and the corresponding true value; A probability scene division module, used for dividing the sampled samples into equal intervals according to the absolute value of the prediction error, into high probability scene samples and low probability scene samples, wherein the sampled samples in the low probability scene samples have a larger absolute value of the prediction error; A weight assigning module, used to assign different sample weights to different sampled samples in the high probability scenario samples and the low probability scenario samples; A first-stage training module, used for performing first-stage training through a deep learning model based on the sample weights; The second stage training module is used to gradually adjust the sample weights and perform second stage training through a deep learning model until the sample weights of the various sampled samples are the same.

Images