CN114066095A

CN114066095A - Prediction of wind power generation methods using artificial neural network and support vector regression model

Info

Publication number: CN114066095A
Application number: CN202111420806.2A
Authority: CN
Inventors: 李昱坤; 米磊; 董显奕; 郝俊伟; 刘阳; 王旭鹏; 张哲�; 李喜平
Original assignee: Huaneng New Energy Co Ltd Shanxi Branch
Current assignee: Huaneng New Energy Co Ltd Shanxi Branch
Priority date: 2021-11-26
Filing date: 2021-11-26
Publication date: 2022-02-18

Abstract

The invention relates to a method for predicting wind power generation by using artificial neural network and support vector regression model. The randomness of wind power generation limits the increase in the share of wind power in the power system and challenges the market consolidation of wind power. A method for forecasting wind power generation using artificial neural networks and support vector regression models, processing and using actual data from a running wind farm in order to test the forecasting model at the wind farm to predict the level of wind power generation; using forecasting software tools, training ANN and The SVR model, in the context of non-exhaustive training, considers the forecast horizon 24 hours ago every hour; evaluates cases using typical evaluation metrics to evaluate the individual performance of neural networks and support vector regression models; Combined with support vector machines, the optimal wind power forecast results are obtained under the assumption of non-exhaustive model training. The invention is applied to the field of wind power generation.

Description

Method for predicting wind power generation by using artificial neural network and support vector regression model

Technical Field

The invention relates to a method for predicting wind power generation by using an artificial neural network and a support vector regression model.

Background

The randomness of wind power generation limits the increase of the share of wind power generation in the power system and poses a challenge to market integration of wind power generation, mainly because the wind farms built today need to cope with more dynamic pricing mechanisms. In this new environment, advanced bidding strategies require the introduction of new elements from wind power generation participants to support wind power generation and address the inherent changing effects to be considered a prerequisite.

The energy storage is a novel element, which can support higher wind energy permeability and ensure the dispatching of the wind power plant under optimized operating conditions and benefits. On the other hand, energy storage technologies, mainly battery technologies, have not been promising until now, and may be cost-reducing, thereby potentially making this solution cost a broader spectrum of wind energy applications than just some niche market aspects (e.g., off-grid and island power systems dominate petroleum power recommendations to increase operating costs, which may justify the coupling of wind power generation and energy storage) And a prediction horizon that can yield short-term and long-term predictions that can be effectively extended even to the future day. Internal software tools developed by a soft energy application laboratory team are used, various integrated artificial intelligence prediction technologies and algorithms are developed, the prediction application is customized, and different aspects of wind power generation prediction are researched in the current work. More specifically, by using actual operational data of existing wind plants, we first apply Artificial Neural Networks (ANNs) to make individual predictions of wind power generation within the next 24 hours. After evaluating the performance of two separate techniques, we continue to combine them to create an integrated model with the goal of producing optimal predicted results by clustering, without the need for constant testing.

Disclosure of Invention

The invention aims to provide a method for predicting wind power generation by using an artificial neural network and a support vector regression model.

The above purpose is realized by the following technical scheme:

a method for predicting wind power generation by using an artificial neural network and a support vector regression model comprises the following steps:

(1) processing and using actual data of an operating wind farm to test the level of wind power generation predicted by the predictive model at the wind farm;

(2) training the ANN and SVR models using a predictive software tool, taking into account a prediction horizon 24 hours ago per hour, in the context of non-exhaustive training;

(3) evaluating the cases by using typical evaluation indexes, and evaluating the individual performances of the neural network and the support vector regression model;

(4) and combining the neural network with the support vector machine through clustering to obtain an optimal wind power prediction result under the assumption of non-exhaustive model training.

The method for predicting wind power generation by using the artificial neural network and the support vector regression model comprises the following steps:

the prediction tool is an automatic AI platform based on an Encog ML framework, can solve regression problems, such as prediction of wind power generation, and the available range of the ML method is limited to two ML methods;

the ML method employed has a support vector machine and a neural network, the latter in the form of a feed-forward or recursive network. The main idea of SVR is to find a function f (x) with the largest deviation from the actual training target epsilon in all training patterns. The range of the tolerance ε is set to 0.1;

in order to process the capability and precision epsilon of the non-existent function, all approximate training pair SVR models are adapted to the soft edge loss function concept and are optimized by using the constraint of a relaxation variable;

in addition, a constant C is introduced in the objective function to penalize the use of those slack variables. The constant C is the superparameter influence function f (x), the value of which is limited by the fact that the experiment uses logarithmic steps and the appropriate values are training sets that are highly dependent on the training pattern, in case of separating methods that are non-linear in the N-dimensional space, rather than finding a higher degree curve, a possible increase in N, since this fact makes only the dot product of the test data needed, N, and even a transfer function, redundant. To this end, the dot product can be directly replaced by a kernel function;

in the developed tool, radial basis function kernels, or gaussian kernels, are used:

where x (test vector) and x' (support vector) are vectors to be projected into the new vector space, but are hyper-parameters, representing the inverse of the radius of influence of the selected support vector;

the main difference between feed forward and recursive networks for ANNs is that the former allows information to be propagated from input to output, while the latter allows information to be propagated in both directions. The latter is achieved by an additional layer of neurons, called the context layer, which stores the values of the hidden cells. The content of the context layer is fed back to the hidden layer in the input of the next stage, providing "memory" for the network.

The invention has the beneficial effects that:

1. the ANNs and SVR models of the invention are tested on the basis of different prediction ranges, and the actual wind speed and wind power generation measurement data of an operating wind power plant are used as case research. The model was tested using an internal prediction tool and the results obtained reflect an overall better fit of the SVR method, especially for predictions spanning more than 6 hours. Meanwhile, the combined prediction method for predicting wind power generation is optimized in an effort, the prediction method is improved through the field of clustering prediction, and the SVR method is enough to be used even for 24 hours.

2. The application of support vector regression methods has proven to be more efficient since the applied models perform better in predicting the peak and minimum values of power generation. However, even though the performance of SVR is generally better than that of ANNs, the combined results (ensemble model) of ANNs and SVR appear to be the best approach in a given power generation region. One example is to introduce a combination of two approaches by clustering, i.e., average predicted usage between 1500 kilowatts and 3800 kilowatts, as shown in the following figure. Finally, as the predicted horizon increases, the artificial neural network provides a more accurate prediction of wind power generation. On the other hand, a well-adjusted SVR seems to give very satisfactory results even for the prediction of the previous day (99.60% R2, 98.99% consistency index and 334kW mean absolute error).

Description of the drawings:

fig. 1 is a graph of prediction accuracy of ANNs of the present invention for different prediction views.

Fig. 2 is a graph showing the predicted results of ANNs of the present invention.

FIG. 3 is a graph of the prediction accuracy for 3 hours for the test method of the present invention.

The specific implementation mode is as follows:

example 1:

the prediction tool is an automatic AI platform based on an Encog ML framework, and can solve regression problems such as prediction of wind power generation. The available range of ML methods is limited to only two ML methods (ANNs and SVR), while for hyper-parametric optimization of models, random or grid searches can be performed. Grid searching refers to the cyclic use of a fixed set of parameter values (defined in a user interface). Random searching refers to updating parameter values based on the results of a previous trial. In this case, different feature prediction combinations can be used as inputs to form a set of "data" cases that also automatically expand according to the desired prediction range (length and step size). The partitions are tested and verified, as well as the historical data (reverse data) for each property. Likewise, for each "data" case, the tool generates a set of "model" cases according to preferences defined on the user interface. In this way, the performance of different ML methods and the impact of each hyper-parameter involved can be explored.

Example 2:

the ML methods employed are support vector machines (SVR) and neural networks (ann), the latter taking the form of feed-forward or recursive networks. The main idea of SVR is to find a function f (x) with the largest deviation from the actual training target epsilon in all training patterns [23 ]. The range of the tolerance s is set to 0.1. To handle the capability and precision epsilon of the non-existent function, all training of the approximation adapts the "soft edge" penalty function concept to the SVR model by using the constraint optimization problem of the relaxation variables (i.e., find a hyperplane, maximize the distance between the data support vector and the hyperplane). In addition, a constant C is introduced in the objective function to penalize the use of those slack variables. The constant C is the superparameter influence function f (x), the value of which is limited by the fact that the experiment uses logarithmic steps and the appropriate values are training sets that are highly dependent on the training pattern, in case of separating methods that are non-linear in the N-dimensional space, rather than finding a higher degree curve, a possible increase in N, since this fact makes only the dot product of the test data needed, N, and even a transfer function, redundant. To this end, the dot product can be directly replaced by a kernel function. In the developed tool, radial basis function kernels, or gaussian kernels, are used:

where x (test vector) and x' (support vector) are vectors to be projected into the new vector space, but are hyper-parameters, representing the inverse of the radius of influence of the selected support vector.

The main difference between feed forward networks and recursive networks (Elman networks in our example) for ANNs is that the former allows information to be propagated from input to output, while the latter allows information to be propagated in both directions. The latter is achieved by an additional layer of neurons, called the context layer, which stores the values of the hidden cells. The content of the context layer is fed back to the hidden layer in the input of the next stage, providing "memory" for the network.

Case study and training data set

To apply the described method and the internal prediction tool of SEALAB. The wind farm studied used a total of 10 750 kw wind turbines. A detailed data set of operations of up to 2.5 years has been provided, including measurements of the actual wind energy production of all 10 wind turbines, as well as wind speed and wind direction at hub altitude. Therefore, the lower graph provides a representative wind speed and wind power generation measurement sample, and the long-term average wind speed of the wind field is estimated to be 9.3m/s, so that the average capacity factor exceeds 38%.

The particular measurement set is processed to produce a clean data set for training the different predictive models used. As previously described, different prediction ranges were tested, i.e. 1 to 24 hours in advance, while training on the model used 41 features and 60.000 records. These are measurements (i.e. wind speed, wind direction, power generation, ambient temperature) and a time indicator registered to each wind turbine.

For wind power generation predictions using the ML method, as expected, the performance of the model gradually decays with increasing prediction horizon. For this reason, in the current case study, the Elman model was found to be sufficiently accurate up to 6 hours, with the result that 24 hours ahead was an R2 protocol indexed by 87% and greatly improved over previous studies based on the use of artificial neural networks. The generated prediction results are respectively presented in a time series (static and mobile) form as shown in the following graph, and the prediction results can be directly compared with the original wind power output value (ground real value).

At the same time, the application of support vector regression proved to be more efficient as the applied model performed better in predicting the peak and minimum values of power generation. However, even though the performance of SVR is generally better than that of ANNs, the combined results (ensemble model) of ANNs and SVR appear to be the best approach in a given power generation region. One example is to introduce a combination of two approaches by clustering, i.e., average predicted usage between 1500 kilowatts and 3800 kilowatts, as shown in the following figure. Finally, as the predicted horizon increases, the artificial neural network provides a more accurate prediction of wind power generation. On the other hand, a well-adjusted SVR seems to give very satisfactory results even for the prediction of the previous day (99.60% R2, 98.99% consistency index and 334kW mean absolute error).

Example 3:

case study and training data set:

Claims

1. A method for predicting wind power generation using artificial neural network and support vector regression model, the method comprises the steps:

(1) Process and use actual data from an operating wind farm in order to test the forecasting model to predict the level of wind power generation at the wind farm;

(2) Use forecasting software tools to train ANN and SVR models, in the context of non-exhaustive training, taking into account the forecast range 24 hours ago every hour;

(3) Use typical evaluation indicators to evaluate the case, and evaluate the individual performance of the neural network and support vector regression model;

(4) Combining neural network with support vector machine through clustering, under the assumption of non-exhaustive model training, the optimal wind power forecasting result can be obtained.

2. The method for predicting wind power generation using artificial neural network and support vector regression model according to claim 1, the method comprises the steps:

The forecast tool is an automated AI platform based on the Encog ML framework, capable of solving regression problems, such as wind power forecasting, where the availability of ML methods is limited to two ML methods;

In order to deal with non-existent functions with capability and accuracy ε, approximate all trained pairs of SVR models by adapting the soft-edge loss function concept by using slack variables for the constrained optimization problem;

where x (test vector) and x' (support vector) are the vectors to be projected into the new vector space, but are hyperparameters representing the inverse of the radius of influence of the selected support vector;

For ANNs, the main difference between feedforward networks and recurrent networks is that the former allows information to flow from input to output, while the latter allows information to propagate in both directions;

The latter is achieved by an additional layer of neurons, called the context layer, which stores the values of the hidden units;

The content of the context layer is fed back to the hidden layer in the input of the next stage, providing the "memory" for the network.