Ultra-short term wind power forecasting method based on improved VMD

A CNN-GRU ultra-short term wind power prediction method based on improved VMD is proposed. Firstly, in the data processing part, according to the irregularity and non-stationary peculiarities of wind power, the variational mode decomposition algorithm is adopted to decompose the original sequence of wind power to reduce the non-stationarity. At the same time, considering that the parameter setting and selection in the variational modal decomposition algorithm are of great significance to the decomposition effect, the improved whale improvement method is adopted to find the best parameters in VMD to avoid the problem of poor decomposition effect caused by artificial setting. Finally, the CNN-GRU and GRU models are simulated and compared on the Matlab platform through the field wind farm data.


Introduction
Due to the rapid development of today's global economy, human dependence on energy is increasing.However, the excessive consumption of traditional energy not only leads to fewer reserves but also leads to different degrees of environmental problems.In order to solve environmental deterioration caused by traditional energy, vigorously developing non-polluting renewable energy has become the main measures taken by countries, among which wind power and photovoltaic gradually become hot research directions [1][2][3].However, wind power generation often has significant volatility and randomness, which will cause wind curtailment if it is integrated into the power system [4][5].The accurate prediction of power can avoid the appearance of wind curtailment phenomenon and improve economic benefits [6].In addition, it can provide guidance for dispatching and further ensure the safe operation of the system.This paper presents a CNN-GRU [7] forecasting tool based on improved VMD [8].
Firstly, the variational mode decomposition method is adopted to salvage the original sequence of wind power to reduce the non-stationarity.At the same time, considering that the parameter setting and selection in the variational modal decomposition algorithm are of great significance to the decomposition effect, the whale improvement method is adopted to find the best parameters in VMD to avoid the poor decomposition effect caused by artificial settings.Finally, the CNN-GRU hybrid network model for prediction and simulation comparison with the GRU model.

Variational mode decomposition algorithm
Variational mode decomposition (VMD), as a mature signal decomposition method, decomposed the original wind power sequence into multiple sets of inherent mode functions (IMF) to obtain a more stable signal sequence.At the same time, VMD can effectively avoid the mode aliasing and end effect generated during decomposition and reduce most of the noise of the original sequence.The original wind power data is often nonlinear and non-stationary, and the prediction result of the direct input prediction network is generally poor.Hence, the VMD is adopted to reduce the non-stationarity.It mainly consists of the following two steps: 1) The data signal decomposition problem is transformed into a constrained optimization problem.
where k u is each mode function; k f stands for corresponding center frequency; K refers to the total quantity of all modes; t  stands for gradient calculation;  stands for convolution symbol; () t  is a Dirac function; ()   wt is the input signal value; t is time.
2) The constrained problem is transformed into an unconstrained optimization problem by parameter introduction.The parameters introduced are Lagrange operator l and penalty factor, respectively      ( ) where L is Lagrange variation.

WOA optimizes VMD parameters
The selection of k and  usually has an important influence on the decomposition effect.Two parameters are usually selected according to experience, but they are often not accurate enough.Therefore, this article uses the whale improvement method (WOA) to find the best argument for VMD to avoid artificial settings leading to poor decomposition effects.The WOA algorithm flow is as follows: 1) Parameters such as the position of the individual and the number of iterations in the whale population are initialized.
2) For the target prey, each individual whale keeps getting closer to the optimal whale position by randomly encircling and contracting or spiraling.The formula for updating the whale position is where B is the distance between the location of each whale and the optimal whale location at the time of selection of the specified prey; B is the interval length between each whale orientation and the best whale orientation when the specified prey is surrounded by a spiral; A is wobble factor; ( ) Xt  is the optimal orientation for the whale at this time; t indicates how many times the current loop has gone through; ( ) Xt is the orientation of the individual whale at this time;  is a random number.The value ranges from -1 to 1; g is a constant of 0.4, where it represents the probability of tracking and capturing prey; b is the constant that determines the ascent path; d is the control coefficient, which decreases to 0 with iteration; max t is the maximum quantity of iterations; s is a random figure what ranges from 0 to 1.
3) When 1 D  , the position of an individual whale in the population is randomly selected to update the location of other whales and conduct a global search where B is the distance between an individual and the best coordinate when in a random search; o X is the coordinate of the individual unconsciously selected from the total quantity.4) Until all iterations are completed, the method is aborted, and the best parameters

Convolutional neural network
The convolutional network has the peculiarities of deep mining data potential features.The convolution layer mainly uses convolution to capture features.Pooling layers make feature extraction easier by reducing the difficulty of feature capture.The fully connected layer consolidates the data for subsequent regression operations.In this paper, CNN is used to mine the potential features of the data, and finally, the processed feature data is sent to the GRU network for calculation.

Gated recurrent neural networks
In order to better mine effective features, a gated recurrent neural network (GRU) is used to further process the output of a convolutional neural network.The GRU structure is shown in Figure 1.W is the weight parameter; t c indicates the hidden layer status at the current time; is the dot product;  is the sigmoid activation function; t c indicates the output value at the current time.

Example analysis
The data of a wind farm in the United States were selected to collect 2000 sets of data, and the collection interval was 5 minutes.The test set has 288 sample points.

Improved VMD decomposition of wind power data
When WOA optimizes the VMD algorithm, it first performs initialization.The population number is set to 30.The value range of the total number of modes k is set to (2,11).The value range of the secondary penalty factor is set to (1000,4000).The maximum number of iterations is set to 20.VMD parameters change with the number of iterations, as shown in Figure 2, and decomposition results are shown in Figure 3.

Model performance comparison
The CNN-GRU model and GRU model are simulated and compared on the Matlab platform.The correspondence analysis is presented in  2) Combining CNN and GRU networks to construct a prediction model, feature extraction ability and training speed can be improved.
3) Through the field wind farm data, CNN-GRU and GRU models are simulated and compared on the Matlab platform.The comparison results show that CNN-GRU has better prediction accuracy.

a
are update doors and reset doors, respectively; t x indicates the current input;

Figure 2 .
Figure 2. VMD hyperparameters vary with the number of iterations.

Figure 4 .
In comparison, the CNN-GRU model in the figure can better fit the real data.
GRU neural networks use update gates to replace memory gates and forget gates in LSTM network architecture.Compared with the two, the GRU network has a simpler structure, so it can greatly improve the training speed of neural networks and have higher efficiency.The GRU calculation formula is as follows: