CN115759465A - Wind power prediction method based on multi-target collaborative training and NWP implicit correction - Google Patents
Wind power prediction method based on multi-target collaborative training and NWP implicit correction Download PDFInfo
- Publication number
- CN115759465A CN115759465A CN202211543600.3A CN202211543600A CN115759465A CN 115759465 A CN115759465 A CN 115759465A CN 202211543600 A CN202211543600 A CN 202211543600A CN 115759465 A CN115759465 A CN 115759465A
- Authority
- CN
- China
- Prior art keywords
- data
- wind power
- nwp
- prediction
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a wind power prediction method based on multi-target collaborative training and NWP implicit correction. When the model network is optimized and trained, the wind power and the NWP input error corresponding to the measured weather are optimized, the NWP correction result output error corresponding to the hidden layer of the wind power is optimized, and the problem that a two-step prediction method is possibly invalid is solved. In addition, the method of the invention only adopts one network to simultaneously realize the hidden correction and the power prediction of the NWP, thereby avoiding secondary calculation, saving the calculation and storage cost and improving the prediction performance through the NWP correction.
Description
Technical Field
The invention belongs to the technical field of wind power prediction, and particularly relates to a wind power prediction method based on multi-target collaborative training and NWP implicit correction.
Background
With the rapid development of new energy, how to build a novel power system and realize high-proportion new energy grid connection becomes a research hotspot of people, and the method is also a must path for realizing the double-carbon strategic target; wind power currently accounts for the highest proportion in new energy and is also the main body of future new energy. According to the statistical information of the national energy bureau, the installed capacity of the wind power grid-connected in China breaks through the 3 hundred million kilowatt major at 11-29 months in 2021, is 1.4 times of the wind power general installation of the European Union at the end of 2020 and 2.6 times of the United states, and is the first to stably live in the world for 12 continuous years. However, with the continuous increase of installed capacity of wind power, the proportion of wind power in a power grid is increased day by day, and new wind power problems are continuously generated and need to be solved urgently, so that a wind power prediction technology with higher precision needs to be researched.
Wind power prediction generally comprises ultra-short-term prediction, short-term prediction and medium-long-term prediction, wherein the ultra-short-term prediction mainly helps to optimize the frequency modulation rotating storage capacity, the short-term prediction is mainly used for scheduling of a power grid, the medium-long-term prediction duration is different from several days to several weeks or even several months, the prediction result can provide a plan for maintaining and debugging equipment of a wind power plant, and evaluation data can also be provided for construction feasibility of the wind power plant. At present, related research on short-term and ultra-short-term wind power prediction is relatively mature, the prediction result can basically meet the requirements of the national energy agency, research difficulties gradually change from short-term and ultra-short-term prediction to medium-term and long-term prediction, and improvement and optimization are mainly carried out on the basis of an intelligent learning method. The intelligent learning method mainly takes NWP (Numerical Weather prediction) data and relevant historical data as input, and takes a machine learning model as a main network for modeling. With the rapid development of artificial intelligence algorithms, intelligent learning models have stronger nonlinear fitting capability, and more intelligent learning models such as long-time and short-time memory networks, random forests, gradient boosting decision trees and the like are used for wind power prediction modeling.
The Chinese patent with the publication number of CN113379142A provides a short-term wind power prediction method based on a wind speed correction and fusion model, and aims to solve the problems that the accuracy of NWP forecast data needs to be improved, and the prediction effect is poor due to the deviation of the NWP forecast wind speed and the actual wind speed of a wind power plant; however, this patent technology only corrects the wind speed error and performs power prediction only by using the corrected wind speed, and does not consider characteristic values such as wind direction, temperature, humidity, pressure and the like having a large correlation with the wind power. Document [ Cai Zhenqi ] BP neural network wind power short-term prediction research [ D ]. Zhejiang university, 2012] based on numerical weather forecast NWP correction firstly corrects NWP by using measured meteorological data, and then predicts by using NWP correction results, which is called a two-step prediction method; the prediction method improves the accuracy of short-term wind power prediction, but does not perform example analysis of medium-term and long-term wind power prediction; moreover, when the error introduced by the two-step prediction method is larger than the error improvement brought by the power prediction model, the two-step prediction method is invalid, and even the prediction precision is obviously reduced.
The wind power prediction precision is improved from the error correction, model optimization and data preprocessing layer surface in the prior art, but algorithms all relate to a plurality of models, are complex and are not strong in applicability.
Disclosure of Invention
In view of the above, the invention provides a wind power prediction method based on multi-target collaborative training and NWP implicit correction, the method takes actual measurement wind power as final output, actual measurement meteorological data as hidden layer output, and a multi-target loss function collaborative training network is used for enabling a model to have wind power prediction and meteorological prediction capabilities at the same time.
A wind power prediction method based on multi-target collaborative training and NWP implicit correction comprises the following steps:
(1) Acquiring wind power plant data of a target area in a past period, wherein the wind power plant data comprises wind power actual measurement data, NWP data and meteorological actual measurement data;
(2) Preprocessing the three types of data to obtain a large number of data samples in a time series form, wherein each group of data samples comprises a wind power actual measurement data sequence, an NWP data sequence and a meteorological actual measurement data sequence which correspond to fixed length time;
(3) Dividing all data samples into a training set and a testing set;
(4) Building a prediction model based on a neural network, wherein the prediction model adopts a multi-layer forward neural network and comprises an input layer, a hidden layer and an output layer, the number of neurons of the input layer is the number of NWP meteorological features, the middle hidden layer carries out feature extraction on NWP data through a nonlinear activation function and nested operation, and the output layer is used for predicting and outputting wind power;
(5) Training a network model by using an NWP data sequence in a training set sample as model input and using a wind power actual measurement data sequence and a meteorological actual measurement data sequence as truth value labels;
(6) And inputting the NWP data sequence in the test set sample into the trained prediction model, and directly outputting the prediction result to obtain the wind power and weather prediction result.
Further, the wind power actual measurement Data And the meteorological actual measurement Data are acquired by an SCADA (Supervisory Control And Data Acquisition) system of the wind farm, and the NWP Data is acquired from a related third-party company; the NWP data are updated every day and provide prediction data of 7 days in the future, and the NWP data collected in the step (1) are data of 24 hours in the future of the daily data updating time point, so that the quality of the NWP data is improved.
Further, the specific implementation manner of the step (2) is as follows: firstly, dividing three types of data into a plurality of sections with 24 hours as a fixed length, and integrating data values in each section into a time sequence form with 15 minutes as an interval; then identifying repeated values, abnormal values and null values caused by cleaning the wind turbine generator in the data sequence, and further replacing or filling the values by adopting correlation analysis; and finally, normalizing all data sequences, namely uniformly mapping the data values in the sequences to a [0,1] interval.
Further, wind power plant data of the target area in the past year are collected, the data sample of the previous 9 months is used as a training set, and the data sample of the next 3 months is used as a testing set.
Furthermore, the forward neural network adopts 5 hidden layers H1-H5, the number of neurons of H1-H5 is 100, 50, 40, 20 and 10 respectively, the output of the last 4 neurons in H3 is extracted to be used as a meteorological prediction result, and the output characteristics of the wind speed, the air temperature, the air pressure and the relative humidity at the high position of the hub of the wind turbine generator are 4 items.
Furthermore, the weather prediction result extracted from the hidden layer is input into the subsequent hidden layer to participate in the subsequent feature extraction.
Further, the nonlinear activation function in the hidden layer is realized by matching a ReLU function and a Sigmoid function.
Further, the specific implementation manner of the step (5) is as follows:
5.1 initializing model parameters, including a bias vector and a weight matrix of each layer, a learning rate and an optimizer;
5.2, inputting the NWP data sequence in the training set sample into a model, outputting the prediction result corresponding to the wind power and the weather through forward propagation of the model, and calculating a loss function L between the prediction result and a truth value label;
and 5.3, continuously iterating and updating model parameters by using an optimizer through a Stochastic Gradient Descent (SGD) method according to the loss function L until the loss function L is converged, and finishing training.
Further, the optimizer adopts an Adam algorithm in the training process, and adopts a BP algorithm to carry out gradient solution.
Further, the expression of the loss function L is as follows:
wherein: p is pred，i And, P real，i Respectively representing the predicted value and the measured value of the wind power and the electric power at the ith moment, C i Indicates the boot capacity at time i, W pred，i And W real，i Respectively representing a weather predicted value and a weather measured value at the ith moment, wherein beta is a weight coefficient, and n is the length of a data sequence.
Based on the technical scheme, the invention has the following beneficial technical effects:
1. when the model network is optimized and trained, the wind power and the NWP input error corresponding to the measured weather are optimized, the NWP correction result output error corresponding to the hidden layer of the wind power is optimized, and the problem that a two-step prediction method is possibly invalid is solved.
2. The method of the invention only adopts one network to simultaneously realize the implicit correction and the power prediction of the NWP, thereby avoiding secondary calculation and saving the calculation and storage cost.
3. With the increase of the prediction duration and the reduction of the prediction accuracy, the method provided by the invention completes the NWP implicit correction and the wind power prediction while optimizing the model through an end-to-end training mode, and improves the prediction accuracy.
4. According to the method, the accuracy and robustness of wind power prediction are improved based on the NWP implicit correction algorithm of multi-target collaborative training, and the prediction performance can be improved through NWP correction.
Drawings
FIG. 1 is a schematic diagram of a neural network-based prediction model structure according to the present invention.
FIG. 2 is a schematic diagram illustrating the concept of the NWP implicit correction algorithm based on multi-target collaborative training according to the present invention.
Detailed Description
In order to more specifically describe the present invention, the following detailed description is provided for the technical solution of the present invention with reference to the accompanying drawings and the specific embodiments.
The invention discloses a wind power prediction method based on multi-target collaborative training and NWP implicit correction, which comprises the following steps of:
(1) And collecting actual measurement data of wind power, NWP data and actual measurement meteorological data in a prediction region within a period of time.
Collecting data of a certain wind power plant, wherein the actually measured wind power data and the actually measured meteorological data are collected by an SCADA system of the wind power plant, and NWP data are obtained from related companies; the NWP data is updated every day, prediction data of 7 days in the future are provided, and data of 24 hours in the future at the time point of updating the data every day are taken, so that the quality of the NWP data is improved.
(2) Preprocessing the data to obtain time sequence data with a time interval of 15min, and dividing the data into a training set and a test set.
2.1 unifying the three types of data into time sequence data with a time interval of 15 min.
2.2 identifying and cleaning null values, repetition values and abnormal data of the wind turbine, replacing the part with more data loss with another wind power plant data with larger relevance to the wind power plant data, searching the wind power plant with larger relevance to pass through the Pearson coefficient, and calculating the formula as follows:
and processing the wind power data of the wind power plant and the wind power data of the adjacent wind power plant based on the Pearson correlation coefficient, selecting the wind power plant data with the r value closest to 1, and performing data replacement on the part with more data loss or abnormal parts.
And 2.3, normalizing the acquired wind power time sequence, uniformly mapping the wind power data set to a [0,1] interval, taking the data of the first 9 months as a training set and the data of the last 3 months as a test set, and obtaining the normalized wind power time sequence training set and the normalized wind power time sequence test set.
The normalized formula is:
wherein: x normal Is normalized data, X is original wind power data, X max 、X min The maximum and minimum values of the raw data, respectively.
(3) As shown in fig. 2, a prediction network model based on a neural network is established, NWP data is used as input of the model, the model simultaneously outputs a wind power prediction result and a meteorological prediction result, and multi-target collaborative training is performed on the model by combining actually measured wind power data and actually measured meteorological data.
The prediction network model based on the neural network is a multi-layer forward neural network, as shown in fig. 1, the number of neurons of an input layer is equal to the number of NWP meteorological features, a middle hidden layer performs feature extraction on NWP data through a nonlinear activation function and nested operation, and an output layer is used for prediction of measured power.
The prediction network model based on the neural network adopts a 5-layer hidden layer neural network, and the number of neurons in a hidden layer is respectively 100, 50, 40, 20 and 10. The input layer is NWP data and constructed time characteristics, and 38-dimensional characteristics are calculated; and (3) taking the last 4 neuron outputs of the 3 rd layer hidden layer as the prediction outputs of the measured weather, and fitting 4 output characteristics of wind speed, air temperature, air pressure and relative humidity at the high position of the hub.
The activation function adopts the collocation of a ReLU function and a Sigmoid function, an Adam algorithm is adopted for optimization in the training process, the initial learning rate is set to be 0.0001, and an early-stop method is adopted for training;
the ReLU function is an activation function commonly used in artificial neural networks, and in a general sense, it refers to a ramp function in mathematics, and its formula is:
f(x)＝max(0，x)
the Sigmoid function is used for hidden layer neuron output, the value range is (0,1), a real number can be mapped to a region (0,1), and the two classes can be used.
Adam (Adaptive motion Estimation) is a first-order optimization algorithm that can replace the traditional stochastic gradient descent process, and can iteratively update neural network weights based on training data. Adam is essentially RMSprop with momentum term, which dynamically adjusts the learning rate of each parameter by using the first moment estimation and the second moment estimation of the gradient, and has the advantages that after offset correction, each iteration learning rate has a certain range, so that the parameters are relatively stable, and the main formula is as follows:
m t ＝μ*m t-1 +(1-μ)*gt
the two equations are the first and second moment estimates of the gradient, respectively, which can be viewed as estimates of the expected E | gt |, E | gt ^2 |.
The above formula is a correction to the first order second moment estimate, which can be approximated as an unbiased estimate of the expectation. It can be seen that the direct moment estimation of the gradient has no additional requirements on the memory and can be dynamically adjusted according to the gradient.
The front part of the formula is a dynamic constraint formed on the learning rate eta and has a definite range.
The invention adopts SGD optimization algorithm to obtain approximate optimal solution of network parameters, adopts BP algorithm to carry out gradient solution, and the loss function of the target is as follows:
wherein: n is the number of training samples, f loss Representing a loss function of a single piece or group of data, y pred，i 、y real，i Respectively representing the predicted samples and the actual samples.
The updating formula of the SGD algorithm is as follows:
the objective loss function determines the objective of network optimization, the gradient determines the direction of parameter optimization, and the regression task selects a Mean Square Error (MSE) function, whose expression is:
wherein: y is pred，i Indicates the predicted value at time i, y real，i Representing the true value at time i.
The invention designs a loss function through a multi-target loss function collaborative training network, wherein the total loss function consists of a wind power prediction error and an actual measurement meteorological prediction error, and the expression is as follows:
wherein: p pred，i ，P real，i Respectively representing predicted wind power and actually measured wind power at moment i, W pred，i ，W real，i Respectively representing predicted weather and measured weather at time i, C i Representing the starting capacity at the moment i for normalizing the data; beta is a weight coefficient for adjusting the weight of the two, and the weight is 0.5 in the embodiment.
(4) Inputting the NWP data of the test set into a prediction network model, outputting a hidden layer of a neural network as an implicit correction result of the NWP, and inputting the implicit correction result into a subsequent hidden layer to obtain a final output power prediction value.
4.1, inputting the test set into the optimal neural network model according to the optimal neural network model and the weight obtained after training is completed, and obtaining a normalized wind power optimal prediction result.
And 4.2, performing inverse normalization on the power prediction result obtained in the step 4.1 to obtain the optimal wind power prediction result.
X＝X normal (X max -X min )+X min
A certain domestic wind power plant is selected as an actual example to carry out simulation analysis, the installed capacity of the wind power plant is 99MW, and the matching data comprises: the data are sampled in 2020, wherein the data are measured output data, measured meteorological data (9 items in total) and NWP data (22 items in total) at 15min intervals, wherein the NWP data are updated every day, and prediction data of 7 days in the future are provided. In the simulation example, the data of 1-8 months in 2020 is used as a training set, and the data of 9-10 months is used as a test set for example analysis.
The prediction result is evaluated by using a common longitudinal index square error (RMSE) and a common transverse index Correlation Coefficient (CC), and the specific calculation formula is as follows:
in the formula: n represents the number of samples and,
p i respectively representing the predicted value and the actual value at the ith moment, cov (p ^, p) representing the covariance of the two, and delta (p ^) representing the standard deviation of the two.
And establishing and comparing different prediction models. Firstly, NWP data is used as input, actual measurement power is used as output, a common power prediction model is established respectively based on a Support Vector Regression (SVR) model, a Gradient Boosting Decision Tree (GBDT) and an NN model of the invention, the NWP correction problem is not considered in the model, end-to-end training is directly carried out based on the NWP data and the actual measurement output, and the model is respectively called as an SVR model, a GBDT model and an NN model; secondly, a two-step prediction method is used based on the GBDT and NN models, namely NWP explicit correction is firstly carried out, and then wind power prediction is carried out, and the model is respectively called as the GBDT-T model and the NN-T model; and finally, establishing an NWP implicit correction model based on multi-target collaborative training, wherein the NWP implicit correction model is called an NN-V model as a virtual correction model and is a total of 6 models.
For fair comparison, models based on the NN network all adopt 5-layer hidden layer neural networks, the number of neurons in the hidden layer is 100, 50, 40, 20 and 10 respectively, and the input layer is NWP data and constructed time characteristics, which totally calculate 38-dimensional characteristics. For the two-step prediction method, 4 output characteristics of wind speed, air temperature, air pressure and relative humidity at the high position of the hub are fitted. For the algorithm model, the output of the last 4 neurons of the 3 rd layer hidden layer is used as the prediction output of the measured weather, the collocation of a ReLU function and a Sigmoid function is adopted for an activation function, an Adam algorithm is adopted for optimization, the initial learning rate is set to be 0.0001, and an early-stopping method is adopted for training.
Because the NWP data provides 7 prediction days in the future and each prediction day is updated, a corresponding model is established for each prediction day in the home position, power prediction with different time scales is realized, namely the 6 types of models are respectively established for each prediction day, and the RMSE indexes of different models on different prediction days are shown in the table 1, and the CC indexes are shown in the table 2.
TABLE 1
TABLE 2
According to the table, firstly, comparing the NN-T model with the NN model to find that the RMSE index of the NN-T model is superior to the NN model on the 1 st, 4 th and 5 th prediction days, and the CC index is superior to the NN model on the 1 st, 2 th, 5 th and 6 th prediction days; comparing the GBDT-T model with the GBDT model, the RMSE index and the CC index of the GBDT-T model are superior to those of the GBDT model in the 2 nd to 7 th prediction days. This demonstrates that NWP correction, while improving prediction performance, is accompanied by instability.
Secondly, comparing the NN-T model with the GBDT-T model, the RMSE index of the GBDT-T model is superior to that of the NN-T model on all prediction days, and the CC index is inferior to that of the NN-T model only on the 1 st prediction day. The GBDT-T model optimizes the RMSE index by 0.71% and the CC index by 1.77% on average relative to the GBDT model. While the NN-T model was optimized on average only 0.94% for the CC index relative to the NN model. This shows that the GBDT-based correction model is more stable and accurate than the NN-based correction model.
Finally, comparing the NN-V model with the rest models, the NN-V model has the following RMSE indexes which are suboptimal only on the 2 nd prediction day, the rest prediction days are optimal, the CC indexes of the NN-V model are obviously superior to those of the rest models on all prediction days, the effect is most stable, and the precision is highest. The NN-V model optimizes the RMSE index by 1.02% and the CC index by 5.21% on average relative to the NN model, and the effectiveness and superiority of the algorithm of the invention in improving the wind power prediction precision are fully demonstrated.
Meanwhile, to further illustrate the superiority of the algorithm of the present invention, table 3 shows the number of networks used by each model:
TABLE 3
As can be seen from Table 3, although the GBDT-T model works well, at least 5 different GBDT networks are required to realize the prediction of different meteorological features and wind power, and the number of models increases linearly with the increase of the meteorological features. The NN-V model can simultaneously carry out NWP implicit correction and wind power prediction by only 1 neural network, improves the prediction performance, saves the calculation and storage cost, and is simpler, more convenient and quicker.
In order to help illustrate the effectiveness of the algorithm of the present invention, the NWP implicit correction result is compared with the explicit correction result, and the specific results are shown in table 4:
TABLE 4
As can be seen from table 4, the corrected NWP index is significantly improved over the original NWP data. Through comparison, the NWP implicit correction result is slightly worse than the explicit correction result, because in the optimization process of the network, the loss function corresponding to the implicit correction result only accounts for a part of the total loss function, and in addition, the loss function value of an output layer is also available; namely, the NWP implicit correction aims to enable hidden layer neurons to capture the features of actually measured weather during feature extraction, optimize and weaken the explicit correction error brought by the two-step prediction method, and does not aim to obtain an accurate NWP explicit correction result.
In combination with the above comparison, we can conclude the following:
1. as the prediction duration increases, the prediction accuracy decreases, and performing NWP correction may improve the prediction performance.
2. The GBDT-based two-step prediction model is more stable and has better RMSE index and CC index than the NN-based model, but the NWP explicit correction error introduced by the two-step prediction method can reduce the prediction precision in some cases.
3. The NWP implicit correction algorithm based on multi-target collaborative training only needs 1 network, the wind power prediction precision is improved, the effect is most stable, and the RMSE indexes and CC indexes of short-term and medium-term prediction results are optimal.
Therefore, the algorithm is easier to realize, can be applied by slightly modifying the conventional NN model, and has stronger expansibility.
The foregoing description of the embodiments is provided to enable one of ordinary skill in the art to make and use the invention, and it is to be understood that other modifications of the embodiments, and the generic principles defined herein may be applied to other embodiments without the use of inventive faculty, as will be readily apparent to those skilled in the art. Therefore, the present invention is not limited to the above embodiments, and those skilled in the art should make improvements and modifications to the present invention based on the disclosure of the present invention within the protection scope of the present invention.
Claims (10)
1. A wind power prediction method based on multi-target collaborative training and NWP implicit correction comprises the following steps:
(1) Acquiring wind power plant data of a target area in a past period, wherein the wind power plant data comprises wind power actual measurement data, NWP data and meteorological actual measurement data;
(2) Preprocessing the three types of data to obtain a large number of data samples in a time series form, wherein each group of data samples comprises a wind power actual measurement data sequence, an NWP data sequence and a meteorological actual measurement data sequence which correspond to fixed length time;
(3) Dividing all data samples into a training set and a testing set;
(4) Building a prediction model based on a neural network, wherein the prediction model adopts a multi-layer forward neural network and comprises an input layer, a hidden layer and an output layer, the number of neurons of the input layer is the number of NWP meteorological features, the middle hidden layer carries out feature extraction on NWP data through a nonlinear activation function and nested operation, and the output layer is used for predicting and outputting wind power;
(5) Training a network model by using an NWP data sequence in a training set sample as model input and using a wind power actual measurement data sequence and a meteorological actual measurement data sequence as truth value labels;
(6) And inputting the NWP data sequence in the test set sample into the trained prediction model, and directly outputting the prediction result to obtain the wind power and weather prediction result.
2. The wind power prediction method of claim 1, characterized in that: the wind power actual measurement data and the meteorological actual measurement data are acquired by an SCADA system of the wind power plant, and the NWP data is acquired from a related third party company; the NWP data are updated every day and provide prediction data of 7 days in the future, and the NWP data collected in the step (1) are data of 24 hours in the future of the daily data updating time point, so that the quality of the NWP data is improved.
3. The wind power prediction method of claim 1, characterized in that: the specific implementation manner of the step (2) is as follows: firstly, dividing three types of data into a plurality of sections with 24 hours as a fixed length, and integrating data values in each section into a time sequence form with 15 minutes as an interval; then identifying repeated values, abnormal values and null values caused by cleaning the wind turbine generator in the data sequence, and further replacing or filling the values by adopting correlation analysis; and finally, normalizing all data sequences, namely mapping the data values in the sequences to a [0,1] interval uniformly.
4. The wind power prediction method of claim 1, characterized in that: and acquiring wind power plant data of the target area in the past year, wherein a data sample of the previous 9 months is used as a training set, and a data sample of the next 3 months is used as a testing set.
5. The wind power prediction method of claim 1, characterized in that: the forward neural network adopts 5 hidden layers H1-H5, the number of neurons of H1-H5 is respectively 100, 50, 40, 20 and 10, the output of the last 4 neurons in H3 is extracted to be used as a meteorological prediction result, and the output characteristics of the wind speed, the air temperature, the air pressure and the relative humidity at the high position of the hub of the wind turbine generator are 4 items.
6. The wind power prediction method of claim 5, characterized in that: the weather prediction result extracted from the hidden layer is input into the subsequent hidden layer to participate in the subsequent feature extraction.
7. The wind power prediction method of claim 1, characterized in that: the nonlinear activation function in the hidden layer is realized by matching a ReLU function and a Sigmoid function.
8. The wind power prediction method according to claim 1, characterized in that: the specific implementation manner of the step (5) is as follows:
5.1 initializing model parameters, including a bias vector and a weight matrix of each layer, a learning rate and an optimizer;
5.2, inputting the NWP data sequence in the training set sample into a model, carrying out forward propagation and output on the model to obtain a prediction result corresponding to the wind power and the weather, and calculating a loss function L between the prediction result and a truth value label;
and 5.3, continuously iterating and updating model parameters by using an optimizer through a Stochastic Gradient Descent (SGD) method according to the loss function L until the loss function L is converged, and finishing training.
9. The wind power prediction method of claim 8, characterized in that: in the training process, the optimizer adopts an Adam algorithm and a BP algorithm to carry out gradient solution.
10. The wind power prediction method of claim 8, characterized in that: the expression of the loss function L is as follows:
wherein: p pred,i And, P real,i Respectively representing the predicted value and the measured value of the wind power and the electric power at the ith moment, C i Indicates the boot capacity at time i, W pred,i And W real,i Respectively representing a weather predicted value and a weather measured value at the ith moment, beta is a weight coefficient,n is the length of the data sequence.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211543600.3A CN115759465A (en) | 2022-12-01 | 2022-12-01 | Wind power prediction method based on multi-target collaborative training and NWP implicit correction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211543600.3A CN115759465A (en) | 2022-12-01 | 2022-12-01 | Wind power prediction method based on multi-target collaborative training and NWP implicit correction |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115759465A true CN115759465A (en) | 2023-03-07 |
Family
ID=85343379
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211543600.3A Pending CN115759465A (en) | 2022-12-01 | 2022-12-01 | Wind power prediction method based on multi-target collaborative training and NWP implicit correction |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115759465A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116979533A (en) * | 2023-09-25 | 2023-10-31 | 西南石油大学 | Self-attention wind farm power prediction method integrating adaptive wavelet |
| CN117394306A (en) * | 2023-09-19 | 2024-01-12 | 华中科技大学 | Wind power prediction model establishment method based on new energy grid connection and application thereof |
-
2022
- 2022-12-01 CN CN202211543600.3A patent/CN115759465A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117394306A (en) * | 2023-09-19 | 2024-01-12 | 华中科技大学 | Wind power prediction model establishment method based on new energy grid connection and application thereof |
| CN116979533A (en) * | 2023-09-25 | 2023-10-31 | 西南石油大学 | Self-attention wind farm power prediction method integrating adaptive wavelet |
| CN116979533B (en) * | 2023-09-25 | 2023-12-08 | 西南石油大学 | Self-attention wind farm power prediction method integrating adaptive wavelet |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110474339B (en) | Power grid reactive power control method based on deep power generation load prediction | |
| CN110543929B (en) | Wind speed interval prediction method and system based on Lorenz system | |
| US20190265768A1 (en) | Method, system and storage medium for predicting power load probability density based on deep learning | |
| CN110969290B (en) | Runoff probability prediction method and system based on deep learning | |
| CN113128793A (en) | Photovoltaic power combination prediction method and system based on multi-source data fusion | |
| CN115759465A (en) | Wind power prediction method based on multi-target collaborative training and NWP implicit correction | |
| CN108921339B (en) | Quantile regression-based photovoltaic power interval prediction method for genetic support vector machine | |
| CN107766986A (en) | Leak integral form echo state network on-line study photovoltaic power Forecasting Methodology | |
| CN113107626B (en) | Load prediction method of combined cycle generator set based on multivariable LSTM | |
| CN111144663B (en) | Ultra-short-term wind power prediction method for offshore wind farm considering output fluctuation process | |
| CN112085285B (en) | Bus load prediction method, device, computer equipment and storage medium | |
| CN109615124B (en) | SCADA master station load prediction method based on deep learning | |
| CN111079989A (en) | Water supply company water supply amount prediction device based on DWT-PCA-LSTM | |
| CN114970952B (en) | Photovoltaic output short-term prediction method and system considering environmental factors | |
| CN113554466A (en) | Short-term power consumption prediction model construction method, prediction method and device | |
| CN115860177A (en) | Photovoltaic power generation power prediction method based on combined machine learning model and application thereof | |
| CN111783947A (en) | Energy consumption prediction method based on LSTM neural network | |
| CN114757427A (en) | Autoregressive corrected LSTM intelligent wind power plant ultra-short term power prediction method | |
| CN111178612B (en) | LSTM load prediction method of grid user based on big data ODPS engine | |
| CN116578551A (en) | GRU-GAN-based power grid data restoration method | |
| CN114595542A (en) | Line loss prediction method | |
| CN117787915A (en) | Digital twin intelligent brain construction method for power distribution network | |
| CN113095547A (en) | Short-term wind power prediction method based on GRA-LSTM-ICE model | |
| CN116187540B (en) | Wind power station ultra-short-term power prediction method based on space-time deviation correction | |
| CN115239029B (en) | Wind power prediction method and system considering power time sequence and meteorological dependent characteristics |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |