CN112529282A - Wind power plant cluster short-term power prediction method based on space-time graph convolutional neural network - Google Patents

Abstract

The invention provides a wind power plant cluster short-term power prediction method based on a space-time graph convolutional neural network, which comprises the following steps of: acquiring historical power within a first target time period to obtain a historical power vector time sequence, and acquiring weather forecast parameter vectors within a second target time period to obtain a weather forecast parameter matrix time sequence; inputting the time sequence sequences of the historical power vector and the weather forecast parameter matrix into a prediction model, and outputting a predicted power vector time sequence in a third target time period; the prediction model is obtained by training a time sequence label based on a sample historical power vector, a sample weather forecast parameter matrix time sequence and a prediction power vector time sequence, and the neural network structure of the prediction model is formed based on a Bi-GRU network and a graph convolution network. The method provided by the invention realizes that the power prediction can jointly consider two factors of historical power and weather forecast parameters, and also improves the accuracy of the prediction.

Description

Short-term power prediction method of wind farm cluster based on spatiotemporal graph convolutional neural network

technical field

The invention relates to the technical field of wind farm cluster power prediction, in particular to a short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network.

Background technique

Renewable energy, especially wind energy, has become the key to alleviating the energy crisis. In recent years, the installed capacity of wind farms has been increasing, and the access to wind farms is mostly done in the form of clusters. However, due to the volatility and randomness of wind farms, wind power not only brings clean energy, but also poses a certain degree of threat to the safe and stable operation of the power grid. Therefore, accurate prediction of wind speed and power is an important guarantee to ensure the grid-connected operation of wind farms. The ultra-short-term power forecast of wind farm clusters requires forecasting the wind power in the next 4 hours at intervals of no less than 15 minutes, which is the basis for adjusting the power generation plan, arranging the dispatch plan and conducting intra-day trading, and it is of great significance. .

The difficulty in ultra-short-term power prediction of wind farm clusters lies in extracting features from the complex spatiotemporal correlation patterns within the wind farm clusters, and then predicting the power in the future. The purpose of the present invention is to design a spatiotemporal graph convolutional neural network based on multi-modal and multi-task learning, which can extract the complex spatiotemporal correlation patterns of the wind farm cluster, and can more accurately give the information of each wind farm in the wind farm cluster. Multi-step prediction power.

In general, the power prediction of a single wind farm can be divided into physical methods and statistical learning methods. Figure 1 is a classification diagram of wind farm prediction models provided by the prior art. The physical method mainly uses the atmospheric motion equation and the wind speed power conversion principle of the wind turbine to perform power prediction, and the statistical learning method includes the content shown in Figure 1.

As shown in Figure 1, the first type of forecasting method is a classic multivariate time series forecasting method, which has a good statistical theoretical basis and foundation. And adopt the LASSO method to extract the variable features. The advantage is that the mathematical meaning is clear, the operation is simple, and the model parameters are very easy to adjust, so it is suitable for online application scenarios. This method is often used in engineering, but the disadvantage is that the error is larger than the specially designed neural network method or other methods. The second type of prediction method mainly adopts the idea of probabilistic graphical model and Gaussian process, and regards the wind process as a combination of Gaussian process. Because the Gaussian process has the good property that it remains Gaussian after combining and superposing, it is easy to solve. The advantage is that the accuracy is high, and the disadvantage is that the amount of calculation is large. The third type of prediction method is the method of machine learning. There are many variants of different machine learning methods. This method also includes prediction methods based on deep learning. For example LSTM, CNN, LSTNet and many more. The advantage is that the accuracy of the model is higher when the model structure is well designed and the training data is sufficient. The disadvantage is that some deep learning methods take a long time to train and require GPU, which is suitable for offline training models and online application scenarios. The fourth category of forecasting methods is the hybrid method. The more typical method is to use the time series empirical mode decomposition and variational mode decomposition methods for decomposition, and use the machine learning prediction model on the subsequence after decomposition. It is also suitable for offline training and online application scenarios.

Wind farm cluster power forecasts also typically include three approaches. The first is the accumulation method, which predicts the power of each wind farm separately and then accumulates it. The second method is the statistical upscaling method. By selecting the representative wind farm in the cluster, the power of the representative wind farm is predicted, and then multiplied by a certain coefficient to obtain the power of the cluster wind farm. The third method is statistical extrapolation. By establishing a database of historical power and numerical weather forecast, and then comparing the current historical power and numerical weather forecast with the data in the historical database, the cluster prediction results are obtained. Of course, statistical downscaling methods have also been proposed in some studies. By downscaling the numerical weather forecast provided by the Meteorological Bureau in space, a numerical weather forecast with higher spatial accuracy can be obtained. Power forecast.

At present, there are also methods of using graph convolutional neural networks to predict wind speed or power. Some researchers have used the network structure of graph convolution combined with LSTM to predict the wind speed of cluster wind farms for ultra-short-term power. Fig. 2 is a schematic diagram of the framework of the graph convolution-based wind farm cluster wind speed prediction provided by the prior art. As shown in Fig. 2, it shows that the graph convolution combined with LSTM network structure is used to supervise the wind speed of the cluster wind farm. The relevant network structure of short-term power prediction, wherein, the network uses the wind speed measurement of each wind farm in the wind farm cluster as input, and the network structure of LSTM+GCN can realize the performance of each wind farm in the wind farm cluster. Wind speed forecast.

In the previous study, Fig. 3 is a schematic diagram of the framework for ultra-short-term power prediction of wind farm clusters based on graph convolution provided by the prior art, as shown in Fig. 3, which adopts graph convolution combined with multi-task and multi-modal learning method to forecast the ultra-short-term power of wind farm clusters, and Fig. 3 shows its related network structure.

Among the existing schemes, the first method uses graph convolution to predict the ultra-short-term wind speed of wind farms, which can be extended to ultra-short-term power prediction of wind farm clusters, but it does not consider how to use numerical weather prediction. Numerical weather forecasting based on atmospheric motion modeling contains relatively rich information on power trend changes. Reasonable consideration of weather forecasting during modeling is of great significance to improve the accuracy. Moreover, the network uses a one-way LSTM algorithm when extracting the time characteristics of historical wind speed, while wind power and wind speed actually have strong two-way characteristics, and the one-way LSTM algorithm also cannot be taken into account.

In the second method, although numerical weather forecasting is considered as the input of the model, it directly uses the historical power and the cubic wind speed in numerical weather forecasting as the feature of the node in the graph convolution as input, and there is no separate analysis of the time series. Features are extracted, so some hidden patterns in the time series cannot be extracted.

Therefore, how to avoid the inability to combine historical power and weather forecast information for future wind farm output power prediction in existing wind farm cluster power predictions, and the inability of extracting features from a single graph convolutional network to extract hidden information in time series, still It is an urgent problem to be solved by those skilled in the art.

SUMMARY OF THE INVENTION

The present invention provides a short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network, which is used to solve the problem that the existing wind farm cluster power prediction cannot be used to predict the output power of future wind farms by combining historical power and weather forecast information. , and the defect that a single graph convolutional network extraction feature cannot extract the hidden information in the time series. By setting the neural network structure used in the model training process, the bi-directional gated recurrent unit Bi-GRU (Bi-GatedRecurrent Unit) network and As the feature extraction module, the graph convolutional network uses the Bi-GRU network to mine the hidden information of the time series to upgrade the features in the time series, so that the power prediction can consider both historical power and weather forecast parameters, and can improve the prediction accuracy.

The present invention provides a short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network, the method comprising:

Collect the historical power of each wind farm in the wind farm cluster in the first target time period until the current moment to obtain the historical power vector time series, and collect the weather forecast parameter vector of each wind farm in the second target time period from the current moment to obtain the weather Forecast parameter matrix time series;

Inputting the historical power vector time series and the weather forecast parameter matrix time series into a prediction model, and outputting the predicted power vector time series of the wind farm cluster in the third target time period in the future from the current moment;

The prediction model is obtained after training based on the sample historical power vector time series, the sample weather forecast parameter matrix time series and the corresponding future third target time period predicted power vector time series labels. The neural network structure used is based on a bidirectional gated recurrent unit GRU network and a graph convolutional network.

According to a short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network provided by the present invention, the method further includes:

The neural network structure used during the training of the prediction model also includes a feature fusion network and a multi-task learning network located behind the graph convolution network;

Wherein, the feature fusion network is used to perform feature fusion of the processed historical power features and the processed weather forecast output by the graph convolution network using a multi-modal learning feature splicing method to obtain the power splicing weather features of each wind farm, The multi-task learning network is used to design different task layer structures and loss functions for each wind farm.

According to a short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network provided by the present invention, the multi-task learning network is used to design different task layer structures and loss functions for each wind farm, specifically including:

Inputting the power splicing weather features of each wind farm into the task layer network corresponding to each wind farm, wherein the task layer network is a fully connected layer network designed according to the power requirements of each wind farm;

The loss function L of the prediction model is determined based on the prediction loss of each task layer network by the following formula:

n表示所述风电场集群中风电场的总个数，

是第i个风电场的预测功率，H为所述预测功率向量时序序列中的时步个数，

是所述风电场集群中第i个风电场的功率拼接天气特征，d_out为所述功率拼接天气特征的维度，W_oi是所述风电场集群中第i个风电场的全连接层网络的待调参数。in,

n represents the total number of wind farms in the wind farm cluster,

is the predicted power of the i-th wind farm, H is the number of time steps in the predicted power vector time series,

is the power splicing weather feature of the ith wind farm in the wind farm cluster, d _out is the dimension of the power splicing weather feature, and W _oi is the fully connected layer network of the ith wind farm in the wind farm cluster Parameters to be adjusted.

According to a short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network provided by the present invention, the neural network structure used in the training of the prediction model is composed of a bidirectional gated recurrent unit GRU network and a graph convolutional network. include:

The feature extraction module in the neural network structure used during the training of the prediction model includes a bidirectional gated recurrent unit GRU network and a graph convolution network;

During the training of the prediction model, the sample historical power vector time series and the sample weather forecast parameter matrix time series are input into the two-way gated cyclic unit GRU network to output the power feature after the dimension increase and the weather forecast feature after the dimension increase. Post-power features and post-dimension-upgraded weather forecast features are input into a graph convolutional network to output post-dimension-reduced power features and post-dimension-reduced weather forecast features.

According to a short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network provided by the present invention, the power feature after dimension-up and the weather forecast feature after dimension-up are input into the graph convolution network and output after dimension-reduction and the power feature and Features of weather forecast after dimensionality reduction, including:

The feature transfer process of the graph convolutional network is represented by the following formula:

为邻接矩阵

的度矩阵，X^(l)∈R^n×d，d是升维后功率特征的维度或升维后天气预报特征的维度，X^(l)是第l层图卷积的输入特征，X^(l+1)∈R^n×h是所述图卷积网络中第l层图卷积的输出特征和l+1层图卷积的输入特征，l＝1,2,…,N,所述图卷积网络中的图卷积总层数为N+1，I_n为维度为n的单位矩阵。Among them, A∈Rn ^×n , A represents the adjacency matrix of the wind farm cluster, n represents the total number of wind farms in the wind farm cluster,

is the adjacency matrix

The degree matrix of , X ^(l) ∈ R ^n×d , d is the dimension of the power feature after up-dimension or the dimension of the weather forecast feature after up-dimension, X ^(l) is the input feature of the l-th layer graph convolution, X ^{( l+1)} ∈R ^n×h is the output feature of the lth layer graph convolution and the input feature of the l+1 layer graph convolution in the graph convolutional network, l=1,2,...,N, the The total number of graph convolution layers in a graph convolutional network is N+1, and In is an identity matrix with dimension _n .

According to a short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network provided by the present invention, the time series in the historical power vector time series, the weather forecast parameter matrix time series, and the predicted power vector time series The step sizes are all equal, and the duration of the third target time period is lower than the preset threshold.

According to a short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network provided by the present invention, the time step is 15 minutes, and the third target time period is 4 hours.

The present invention also provides a short-term power prediction device for wind farm clusters based on a spatiotemporal graph convolutional neural network, including:

The collection unit is used to collect the historical power of each wind farm in the wind farm cluster in the first target time period until the current moment to obtain the historical power vector time series, and collect the weather of each wind farm in the second target time period from the current moment The forecast parameter vector obtains the weather forecast parameter matrix time series;

a prediction unit, configured to input the historical power vector time series and the weather forecast parameter matrix time series into a prediction model, and output the predicted power vector time series of the wind farm cluster in the third target time period in the future from the current moment;

The prediction model is obtained after training based on the sample historical power vector time series, the sample weather forecast parameter matrix time series and the corresponding future third target time period predicted power vector time series labels. The neural network structure used is based on a bidirectional gated recurrent unit GRU network and a graph convolutional network.

The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, when the processor executes the program, the processor implements the space-time based method described in any of the above Steps of a graph convolutional neural network short-term power prediction method for wind farm clusters.

The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, implements the wind farm cluster based on any one of the above-mentioned spatiotemporal graph convolutional neural networks Steps of a short-term power prediction method.

The method and device for short-term power prediction of a wind farm cluster based on a spatiotemporal graph convolutional neural network provided by the present invention obtain the historical power vector time sequence sequence by collecting the historical power of each wind farm in the wind farm cluster within the first target time period up to the current moment, Collecting the weather forecast parameter vectors of the wind farms in the second target time period from the current moment to obtain a weather forecast parameter matrix time series; inputting the historical power vector time series and the weather forecast parameter matrix time series into a prediction model, Output the predicted power vector time series of the wind farm cluster in the third target time period in the future starting from the current moment; wherein, the prediction model is based on the sample historical power vector time series, the sample weather forecast parameter matrix time series and the corresponding future The predicted power vector time series labels in the third target time period are obtained after training, and the neural network structure used in the training of the prediction model is based on a bidirectional gated recurrent unit GRU network and a graph convolution network. Since the prediction model is obtained after training based on the sample power vector time series, the sample weather forecast parameter matrix time series and the corresponding future third target time period predicted power vector time series labels, the power prediction no longer only considers the historical power Instead, the historical power factor and the weather forecast factor are considered together. At the same time, since the bi-directional gated recurrent unit GRU network and the graph convolutional network are set in the neural network used in the training of the prediction model, the Bi-GRU network is very sensitive to the input time series. Mining the hidden information in the sequence is to increase the dimension of the features in the time series sequence, so that the feature extraction module in the neural network extracts more hidden information for input to the subsequent prediction module to improve the accuracy of model training. Therefore, the method and device provided by the present invention realize that the power prediction can jointly consider two factors of historical power and weather forecast parameters, and also improves the accuracy of the prediction.

Description of drawings

In order to explain the present invention or the technical solutions in the prior art more clearly, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are the For some embodiments of the invention, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without any creative effort.

Fig. 1 is a classification diagram of wind farm prediction models provided by the prior art;

FIG. 2 is a schematic diagram of a framework for predicting the wind speed of a cluster of wind farms based on graph convolution provided by the prior art;

FIG. 3 is a schematic diagram of a framework for ultra-short-term power prediction of wind farm clusters based on graph convolution provided by the prior art;

4 is a schematic flowchart of a short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network provided by the present invention;

5 is a schematic diagram of multi-task learning provided by the present invention;

6 is a schematic structural diagram of a short-term power prediction device for wind farm clusters based on a spatiotemporal graph convolutional neural network provided by the present invention;

7 is a schematic diagram of the framework of the improved multimodal multitasking spatiotemporal graph convolution cluster power prediction method provided by the present invention;

8 is a schematic diagram showing the comparison of cluster power prediction performance based on improved multimodal multitasking spatiotemporal graph convolution provided by the present invention;

FIG. 9 is a schematic diagram of a statistical result of power prediction error of each wind farm in a wind farm cluster provided by the present invention;

FIG. 10 is a schematic diagram of the physical structure of an electronic device provided by the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention. , not all examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Power forecasting of wind farm clusters is a typical time series forecasting problem. Given the historical power of M time steps before and the numerical weather forecast of N time steps, it can give the most promising future H time steps. possible power output. Cluster power prediction can be described as:

P _t+1 ,...,P _t+H =f(P _t-M+1 ,...,P _t ,V _t+1 ,...,V _t+N )

where P _t ∈ R ^n×1 represents the wind power vector of the wind farms in the wind farm cluster, V _t ∈ R ^n×k is the matrix of numerical weather forecast, n is the number of wind farms in the wind farm cluster, and k is the numerical weather The number of variables in the forecast, the process of wind farm power forecasting is to reasonably and effectively approximate the mapping by finding the most suitable f. In ultra-short-term power forecasting, because it is necessary to forecast the power of the next 4 hours, the time interval is 15min, So here H is generally 16;

The prediction error is mainly measured by root mean square error (RMSE) and mean absolute error (MAE). The calculation method is as follows:

是所述任一风电场在ti时刻的预测功率值，i＝1,2,3,…,K，对应时序序列中的t1,t2,t3,…,tK一共K个时步，K为时序序列中时步总数。where x _ti is the power value of any wind farm at time ti,

is the predicted power value of any wind farm at time ti, i=1, 2, 3,...,K, corresponding to t1, t2, t3,..., tK in the time series with a total of K time steps, K is the time series The total number of time steps in the sequence.

Existing wind farm cluster power forecasts generally have problems that the historical power and weather forecast information cannot be combined for future wind farm output power forecasts, and that a single graph convolutional network extraction feature cannot extract the implicit information in the time series. The following describes a short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network of the present invention with reference to FIGS. 4-5 . FIG. 4 is a schematic flowchart of a short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network provided by the present invention. As shown in FIG. 4 , the method includes:

Step 410: Collect the historical power of each wind farm in the wind farm cluster in the first target time period until the current moment to obtain a power vector time series, and collect the weather forecast parameter vector of each wind farm in the second target time period from the current moment Obtain the weather forecast parameter matrix time series.

Specifically, in order to predict the power generation of each wind farm in the wind farm cluster for a period of time in the future from the current moment, it is first necessary to collect the historical power of each wind farm in the most recent period of time, that is, until the current moment. The first target time period, for example, the first target time period is 24 hours, then the collected historical power of each wind farm in the wind farm cluster is the current time pushed forward 24 hours as the start time, and the current time as the end time, For the historical power of each wind farm from the start time to the end time, a preset time step is used as the collection time interval when collecting the historical power, and the historical power is collected every short period of the preset time, and then The power vector of the wind farm cluster is composed of the historical power of each wind farm, and the historical power vector sequence is composed of the power vector of the wind farm cluster collected once every preset time step within the first target time period until the current moment. sequence. In the same way, the weather forecast parameter vector of each wind farm in the second target time period from the current moment is collected to obtain the weather forecast parameter matrix time series sequence, that is, the weather forecast parameter of each wind farm is used to form the weather forecast parameter of the wind farm cluster. Matrix, because the weather forecast parameters are usually a set of physical parameters, such as wind speed level, temperature, humidity and fog level, etc., and then collected at every predetermined time step within the second target time period until the current moment The weather forecast parameter matrix of the wind farm cluster constitutes a time series of weather forecast parameter matrix.

Step 420: Input the historical power vector time series and the weather forecast parameter matrix time series into a prediction model, and output the predicted power vector time series of the wind farm cluster in the third target time period in the future from the current moment;

The prediction model is obtained after training based on the sample historical power vector time series, the sample weather forecast parameter matrix time series and the corresponding future third target time period predicted power vector time series labels. The neural network structure used is based on a bidirectional gated recurrent unit GRU network and a graph convolutional network.

Specifically, input the historical power vector time series and the weather forecast parameter matrix time series collected in step 410 into the prediction model, and output the predicted power vector time series of the wind farm cluster in the third target time period in the future from the current moment, That is, the output is the predicted power vector time series of the wind farm cluster in the third target time period in the future starting from the current moment. For example, the third target time period is 4h, and the time step corresponding to the predicted power vector time series is 15min. The output is a time series containing 16 elements of the predicted power vector of the wind farm cluster, wherein each element in turn corresponds to a moment in the third target time period. The prediction model is obtained after training based on the sample historical power vector time series, the sample weather forecast parameter matrix time series and the corresponding future third target time period predicted power vector time series labels. The sample historical power vector time series The sequence, the sample weather forecast parameter matrix time series, and the corresponding future third target time period predicted power vector time series labels are all extracted based on the historical data relative to the current moment. For example, the sample historical power vector time series is derived from the historical data. Select a large number of historical powers of each wind farm in the wind farm cluster in the first target time period to obtain a large number of sample historical power vector time series, and the construction of sample historical power vector time series also needs to ensure that the collection is carried out at equal intervals, and the interval time is also It is set to the preset time step, and at the same time, to construct any sample weather forecast parameter matrix time series, it is necessary to find the first target time period when the corresponding sample historical power vector time series is collected, and then use the first target time period. The ending time is used as the starting time to search for the weather forecast parameter matrix time sequence sequence in the second target time period as the any sample weather forecast parameter matrix time sequence sequence, and any sample weather forecast parameter matrix time sequence sequence also needs to ensure equal intervals Time to collect, and the interval time is also set to the predetermined time step, and finally, to construct the predicted power vector time series label in any third target time period in the future, it is necessary to find the corresponding sample historical power vector time series when the time series is collected. a target time period, and then use the end time of the first target time period as the starting time to search for the historical power vector time series in the third target time period as the predicted power vector time series in any future third target time period Label. During the training of the prediction model, the samples include the sample historical power vector time series and the sample weather forecast parameter matrix time series, and the label corresponding to the sample is the predicted power vector time series label in the third target time period in the future. It should be noted here that since the prediction model is trained first, and the accurate prediction model formed after the training is used for prediction later, the actual input and output specification parameters have been determined before the model training, for example, The sample historical power vector time series includes the number of elements, the time interval between each element, the sample weather forecast parameter matrix time series includes the number of elements, the time interval between each element, and the predicted power vector time series index. Including the number of elements, the time interval between each element is determined before training, and the training set is constructed with the above-determined specifications. The historical power vector time series and weather forecast parameter matrix time series that need to be input have the same specifications as the corresponding samples in the training set, that is, the historical power vector time series input during use is consistent with the number of elements included in the sample historical power vector time series. The time interval between each element is also the same, the input weather forecast parameter matrix time series sequence is consistent with the number of elements included in the sample weather forecast parameter matrix time series sequence, and the time interval between each element is also the same. Similarly, The output predicted power vector time series is also consistent with the specifications of the predicted power vector time series label used in the training of the prediction model, that is, the output predicted power vector time series and the predicted power vector time series label include the same number of elements and The time interval between each element is also consistent. It is defined here that the neural network structure used in the training of the prediction model is based on the bidirectional gated recurrent unit GRU network and the graph convolutional network, that is, only the graph convolutional network in the prior art is used as a feature extraction module. The feature extraction module includes Bi-GRU network and graph convolution network. Among them, the newly added Bi-GRU network can extract the hidden information in the input sample time series, that is, the input time series features are dimensionally upgraded to increase the dimension of the features. , extract more hidden information.

It should be noted that the input used in this neural network includes historical power and the cube of wind speed at different heights in numerical weather forecasting, because the cube of wind speed contains information that can better reflect the power fluctuations of wind farms than the wind speed itself. The specific reason can be obtained by analyzing the output principle formula of the following wind farm:

Among them, P _wind is the wind power, C _p is the wind power conversion rate of the wind farm, v is the wind speed, ρ is the air density, and A is the air coefficient. It can be seen from the above formula that the power and the cube of wind speed have a closer relationship, so using the cube of wind speed in numerical weather forecasting as the input of the model may have better performance. And the input of the model is the wind speed at different heights, because the wind speed at different heights may contain information such as air pressure and temperature.

The bidirectional GRU is explained here. This part of the network is mainly used to extract the time series features of the historical power and the cubic wind speed of numerical weather forecast. The inputs are historical power of length M, and numerical weather forecast wind speed of length N.

The GRU network is a variant of the LSTM network, which is simpler in structure than the LSTM network, and generally works better because it has fewer parameters and is easier to converge. Because GRU replaces the forget gate and input gate in LSTM with update gate. The cell state and hidden state in GRU are merged, and the method of calculating the current moment information is different from that of LSTM. GRU includes four units: reset gate, update gate, candidate memory unit and current memory unit, and their calculations are shown in the following equations.

r _t =σ(W _r x _t-1 +U _r h _t-1 )

z _t =σ(W _z x _t +U _z h _t-1 )

h_t分别为重置门、更新门、候选记忆单元和当前时刻记忆单元向量；W_r和U_r分别为重置门参数；W_z和U_z分别为更新门参数；W_k和U_k分别为候选记忆单元的参数；σ(·)表示激活函数，“⊙”表示数组元素依次相乘。where x _t is the input vector; r _i , z _i ,

h _t are reset gate, update gate, candidate memory unit and current memory unit _vector respectively; W _r and Ur are reset gate parameters; W _z and U _z are update gate parameters; W _k and U _k are respectively is the parameter of the candidate memory unit; σ(·) represents the activation function, and “⊙” represents the array elements are multiplied in turn.

The short-term power prediction method for a wind farm cluster based on a spatiotemporal graph convolutional neural network provided by the present invention obtains the historical power vector time series by collecting the historical power of each wind farm in the wind farm cluster within the first target time period up to the current moment, and collects a sequence from The weather forecast parameter vector of each wind farm in the second target time period from the current moment to obtain the weather forecast parameter matrix time series; input the historical power vector time series and the weather forecast parameter matrix time series into the prediction model, and output from The current moment starts the predicted power vector time series of the wind farm cluster in the third target time period in the future; wherein, the prediction model is based on the sample historical power vector time series, the sample weather forecast parameter matrix time series and the corresponding future third The predicted power vector time series labels in the target time period are obtained after training, and the neural network structure used in the training of the prediction model is based on a bidirectional gated recurrent unit GRU network and a graph convolutional network. Since the prediction model is obtained after training based on the sample power vector time series, the sample weather forecast parameter matrix time series and the corresponding future third target time period predicted power vector time series labels, the power prediction no longer only considers the historical power Instead, the historical power factor and the weather forecast factor are considered together. At the same time, since the bi-directional gated recurrent unit GRU network and the graph convolutional network are set in the neural network used in the training of the prediction model, the Bi-GRU network is very sensitive to the input time series. Mining the hidden information in the sequence is to increase the dimension of the features in the time series sequence, so that the feature extraction module in the neural network extracts more hidden information for input to the subsequent prediction module to improve the accuracy of model training. Therefore, the method provided by the present invention realizes that two factors of historical power and weather forecast parameters can be jointly considered in power prediction, and the accuracy of prediction is also improved.

On the basis of the foregoing embodiment, the method further includes:

The neural network structure used during the training of the prediction model also includes a feature fusion network and a multi-task learning network located behind the graph convolution network;

Wherein, the feature fusion network is used to perform feature fusion of the processed historical power features and the processed weather forecast output by the graph convolution network using a multi-modal learning feature splicing method to obtain the power splicing weather features of each wind farm, The multi-task learning network is used to design different task layer structures and loss functions for each wind farm.

Specifically, the neural network structure used in the training of the prediction model includes four parts, which are Bi-GRU network, graph convolution network, feature fusion network and multi-task learning network in turn. In order to extract the features of the sample time series series to extract the hidden information in the time series series, the feature fusion adopts the multi-modal learning method to perform feature fusion on the processed historical power features and the processed weather forecast output by each wind farm from the graph convolution network. The historical power feature and the numerical weather forecast feature are fused. This feature fusion mainly adopts the feature splicing method in the multi-modal learning to perform feature fusion. The multi-task learning network is used to design different task layer structures and loss functions for each wind farm. That is, the multi-task learning network mainly uses the spatiotemporal features extracted by the feature fusion network (that is, the power splicing weather features) to predict the power of each wind farm, wherein each wind farm adopts a fully connected model, and the output is each wind farm. The power of the wind farm, that is, the multi-task learning network can be designed separately according to the specific requirements of each wind farm. The multi-task learning network usually adopts the same part of the network in the fixed task, and then designs the loss function and the task layer separately for different tasks.

In the method provided by the present invention, by setting up a multi-task learning network, different wind farms can set different task layer structures and their respective methods for calculating the error of their respective label values according to the requirements of the application scenarios, so as to ensure the flexibility of the overall prediction, and according to the scenarios Application requirements achieve more accurate forecasts.

Based on the above embodiment, in this method, the multi-task learning network is used to design different task layer structures and loss functions for each wind farm, specifically including:

Inputting the power splicing weather features of each wind farm into the task layer network corresponding to each wind farm, wherein the task layer network is a fully connected layer network designed according to the power requirements of each wind farm;

The loss function L of the prediction model is determined based on the prediction loss of each task layer network by the following formula:

n表示所述风电场集群中风电场的总个数，

是第i个风电场的预测功率，H为所述预测功率向量时序序列中的时步个数，

Y_i是所述风电场集群中第i个风电场的功率拼接天气特征，d_out为所述功率拼接天气特征的维度，W_oi是所述风电场集群中第i个风电场的全连接层网络的待调参数。in,

n represents the total number of wind farms in the wind farm cluster,

is the predicted power of the i-th wind farm, H is the number of time steps in the predicted power vector time series,

Yi is the power splicing weather feature of the _ith wind farm in the wind farm cluster, d _out is the dimension of the power splicing weather feature, and W _oi is the fully connected layer of the ith wind farm in the wind farm cluster The parameters to be tuned for the network.

是第n个风电场的预测功率，

表示所有风电场的功率拼接天气特征形成的矩阵。每个风电场根据各自的任务层网络求各自的预测值，并根据各自的标签值确定各自的误差，但是预测模型的神经网络整体的损失函数是将所有风电场预测功率的误差进行求和。Specifically, FIG. 5 is a schematic diagram of multi-task learning provided by the present invention. As shown in FIG. 5 , for n wind farms in a wind farm cluster, each uses its own task layer network, for example, the last one in FIG. 5 Wind farm, namely the nth wind farm, its power splicing weather feature is Y _n ,

is the predicted power of the nth wind farm,

A matrix representing the power splicing weather characteristics of all wind farms. Each wind farm obtains its own prediction value according to its own task layer network, and determines its own error according to its own label value, but the overall loss function of the neural network of the prediction model is to sum the errors of the predicted power of all wind farms.

The loss function L of the prediction model is determined based on the prediction loss of each task layer network by the following formula:

n表示所述风电场集群中风电场的总个数，

是第i个风电场的预测功率，H为所述预测功率向量时序序列中的时步个数，

Y_i是所述风电场集群中第i个风电场的功率拼接天气特征，d_out为所述功率拼接天气特征的维度，W_oi是所述风电场集群中第i个风电场的全连接层网络的待调参数。这样，可以实现对每个风电场分别进行预测和计算损失函数。in,

n represents the total number of wind farms in the wind farm cluster,

is the predicted power of the i-th wind farm, H is the number of time steps in the predicted power vector time series,

Yi is the power splicing weather feature of the _ith wind farm in the wind farm cluster, d _out is the dimension of the power splicing weather feature, and W _oi is the fully connected layer of the ith wind farm in the wind farm cluster The parameters to be tuned for the network. In this way, it is possible to predict and calculate the loss function separately for each wind farm.

On the basis of the above embodiment, the neural network structure used in the training of the prediction model is based on the bidirectional gated recurrent unit GRU network and the graph convolutional network, and specifically includes:

The feature extraction module in the neural network structure used during the training of the prediction model includes a bidirectional gated recurrent unit GRU network and a graph convolution network;

During the training of the prediction model, the sample historical power vector time series and the sample weather forecast parameter matrix time series are input into the two-way gated cyclic unit GRU network to output the power feature after the dimension increase and the weather forecast feature after the dimension increase. Post-power features and post-dimension-upgraded weather forecast features are input into a graph convolutional network to output post-dimension-reduced power features and post-dimension-reduced weather forecast features.

Specifically, the feature extraction module in the neural network structure used in the training of the prediction model is composed of a bidirectional gated recurrent unit GRU network and a graph convolutional network connected in sequence. The sequence and the sample weather forecast parameter matrix time series are input into the bidirectional gated cyclic unit GRU network to output the power feature after the dimensional upgrade and the weather forecast feature after the dimensional upgrade, and the power feature after the dimensional upgrade and the weather forecast feature after the dimensional upgrade are input to the map volume The product network outputs the power features after dimensionality reduction and the weather forecast features after dimensionality reduction.

On the basis of the above-mentioned embodiment, the power feature after the dimension increase and the weather forecast feature after the dimension increase are input to the graph convolution network to output the power feature after the dimension reduction and the weather forecast feature after the dimension reduction, specifically including:

The feature transfer process of the graph convolutional network is represented by the following formula:

为邻接矩阵

的度矩阵，X^(l)∈R^n×d，d是升维后功率特征的维度或升维后天气预报特征的维度，X^(l)是第l层图卷积的输入特征，X^(l+1)∈R^n×h是所述图卷积网络中第l层图卷积的输出特征和l+1层图卷积的输入特征，l＝1,2,…,N,所述图卷积网络中的图卷积总层数为N+1，I_n为维度为n的单位矩阵。Among them, A∈Rn ^×n , A represents the adjacency matrix of the wind farm cluster, n represents the total number of wind farms in the wind farm cluster,

is the adjacency matrix

The degree matrix of , X ^(l) ∈ R ^n×d , d is the dimension of the power feature after up-dimension or the dimension of the weather forecast feature after up-dimension, X ^(l) is the input feature of the l-th layer graph convolution, X ^{( l+1)} ∈R ^n×h is the output feature of the lth layer graph convolution and the input feature of the l+1 layer graph convolution in the graph convolutional network, l=1,2,...,N, the The total number of graph convolution layers in a graph convolutional network is N+1, and In is an identity matrix with dimension _n .

Specifically, the graph convolution network mainly uses graph convolution to extract the spatial features of historical power and numerical weather forecast wind speed, and the input is the time series feature extracted by the bidirectional GRU (that is, the power feature after the dimension increase and the weather forecast feature after the dimension increase). ), and the output is the spatiotemporal features of historical power and numerical weather forecast (that is, power features after re-dimension reduction and weather forecast features after re-dimension reduction).

The graph convolution used in the present invention is mainly spectral domain graph convolution. The first-order approximation of the Chebyshev spectral domain filter is adopted, where the feature transfer process of the graph convolutional layer is as follows:

为邻接矩阵

的度矩阵，X^(l)∈R^n×d，d是升维后功率特征的维度或升维后天气预报特征的维度，X^(l)是第l层图卷积的输入特征，X^(l+1)∈R^n×h是所述图卷积网络中第l层图卷积的输出特征和l+1层图卷积的输入特征，l＝1,2,…,N,所述图卷积网络中的图卷积总层数为N+1，I_n为维度为n的单位矩阵。Among them, A∈Rn ^×n , A represents the adjacency matrix of the wind farm cluster, n represents the total number of wind farms in the wind farm cluster,

is the adjacency matrix

The degree matrix of , X ^(l) ∈ R ^n×d , d is the dimension of the power feature after up-dimension or the dimension of the weather forecast feature after up-dimension, X ^(l) is the input feature of the l-th layer graph convolution, X ^{( l+1)} ∈R ^n×h is the output feature of the lth layer graph convolution and the input feature of the l+1 layer graph convolution in the graph convolutional network, l=1,2,...,N, the The total number of graph convolution layers in a graph convolutional network is N+1, and In is an identity matrix with dimension _n .

The adjacency matrix is explained here. The adjacency matrix is a parameter used to show the geographical dispersion of wind farms. It is very important for the modeling of spatiotemporal fusion and there are many ways to construct it. Since the wind farms in the wind farm cluster are all located in the same area, their geographical dispersion is usually described based on distance. Therefore, in the present invention, the adjacency matrix is determined based on the Gaussian kernel threshold distance function of each wind farm, and the specific formula is as follows:

Among them, A _i,j is the element value of the ith row and jth column of the adjacency matrix A of the wind farm cluster, and dist(i,j) represents the ith wind farm and the jth wind farm in the wind farm cluster The geographic distance (geographical distance) of , std represents the standard deviation of the distances between all n wind farms in the wind farm cluster, and ε is a defined threshold. Here, the size of ε is preferably set to be half of the average distance. If the distance between the two wind farms is less than the threshold, it is determined that there is no correlation between the two wind farms to ensure the sparsity of the adjacency matrix.

On the basis of the above embodiment, the time step lengths in the historical power vector time series, the weather forecast parameter matrix time series, and the predicted power vector time series are all equal, and the third target time period is less than Preset threshold.

Specifically, the time steps in the historical power vector time series, the weather forecast parameter matrix time series, and the predicted power vector time series are usually set to the same value to facilitate subsequent calculation and recording, and the third target time The duration of the segment is lower than the preset threshold, because the present invention is most suitable for predicting power in the short term. Therefore, the duration of the third target duration has an upper limit, which is adjusted according to different requirements of application scenarios.

On the basis of the above embodiment, the time step is 15 minutes, and the third target time period is 4 hours.

Specifically, the time step is usually set to 15min, and the duration of the third target time period is set to 4h.

The short-term power prediction device for wind farm clusters based on spatiotemporal graph convolutional neural networks provided by the present invention will be described below. The short-term power prediction device for wind farm clusters based on spatiotemporal graph convolutional neural networks The short-term power prediction methods of wind farm clusters based on spatiotemporal graph convolutional neural networks can be referred to each other.

FIG. 6 is a schematic structural diagram of a short-term power prediction device for wind farm clusters based on a spatiotemporal graph convolutional neural network provided by the present invention. As shown in FIG. 6 , the device includes a collection unit 610 and a prediction unit 620, wherein,

The collecting unit 610 is configured to collect the historical power of each wind farm in the wind farm cluster in the first target time period until the current moment to obtain a historical power vector time series, and collect the wind power in the second target time period starting from the current moment. The weather forecast parameter vector of the field is used to obtain the weather forecast parameter matrix time series;

The prediction unit 620 is configured to input the historical power vector time series and the weather forecast parameter matrix time series into a prediction model, and output the predicted power vector of the wind farm cluster in the third target time period in the future from the current moment time series;

The prediction model is obtained after training based on the sample historical power vector time series, the sample weather forecast parameter matrix time series and the corresponding future third target time period predicted power vector time series labels. The neural network structure used is based on a bidirectional gated recurrent unit GRU network and a graph convolutional network.

The short-term power prediction device for wind farm clusters based on the spatiotemporal graph convolutional neural network provided by the present invention obtains the historical power vector time series by collecting the historical power of each wind farm in the wind farm cluster within the first target time period up to the current moment. The weather forecast parameter vector of each wind farm in the second target time period from the current moment to obtain the weather forecast parameter matrix time series; input the historical power vector time series and the weather forecast parameter matrix time series into the prediction model, and output from The current moment starts the predicted power vector time series of the wind farm cluster in the third target time period in the future; wherein, the prediction model is based on the sample historical power vector time series, the sample weather forecast parameter matrix time series and the corresponding future third The predicted power vector time series labels in the target time period are obtained after training, and the neural network structure used in the training of the prediction model is based on a bidirectional gated recurrent unit GRU network and a graph convolutional network. Since the prediction model is obtained after training based on the sample power vector time series, the sample weather forecast parameter matrix time series and the corresponding future third target time period predicted power vector time series labels, the power prediction no longer only considers the historical power Instead, the historical power factor and the weather forecast factor are considered together. At the same time, since the bi-directional gated recurrent unit GRU network and the graph convolutional network are set in the neural network used in the training of the prediction model, the Bi-GRU network is very sensitive to the input time series. Mining the hidden information in the sequence is to increase the dimension of the features in the time series sequence, so that the feature extraction module in the neural network extracts more hidden information for input to the subsequent prediction module to improve the accuracy of model training. Therefore, the device provided by the present invention realizes that two factors, historical power and weather forecast parameters, can be considered jointly in power prediction, and the accuracy of prediction is also improved.

On the basis of the above-mentioned embodiment, in this device,

The neural network structure used in the training of the prediction model also includes a feature fusion network and a multi-task learning network located behind the graph convolution network;

Wherein, the feature fusion network is used to perform feature fusion of the processed historical power features and the processed weather forecast output by the graph convolution network using a multi-modal learning feature splicing method to obtain the power splicing weather features of each wind farm, The multi-task learning network is used to design different task layer structures and loss functions for each wind farm.

In the method provided by the present invention, by setting up a multi-task learning network, different wind farms can set different task layer structures and their respective methods for calculating the error of their respective label values according to the requirements of the application scenarios, so as to ensure the flexibility of the overall prediction, and according to the scenarios Application requirements achieve more accurate forecasts.

On the basis of the above embodiment, in the device, the multi-task learning network is used to design different task layer structures and loss functions for each wind farm, specifically including:

Inputting the power splicing weather features of each wind farm into the task layer network corresponding to each wind farm, wherein the task layer network is a fully connected layer network designed according to the power requirements of each wind farm;

The loss function L of the prediction model is determined based on the prediction loss of each task layer network by the following formula:

n表示所述风电场集群中风电场的总个数，

是第i个风电场的预测功率，H为所述预测功率向量时序序列中的时步个数，

是所述风电场集群中第i个风电场的功率拼接天气特征，d_out为所述功率拼接天气特征的维度，W_oi是所述风电场集群中第i个风电场的全连接层网络的待调参数。in,

n represents the total number of wind farms in the wind farm cluster,

is the predicted power of the i-th wind farm, H is the number of time steps in the predicted power vector time series,

is the power splicing weather feature of the ith wind farm in the wind farm cluster, d _out is the dimension of the power splicing weather feature, and W _oi is the fully connected layer network of the ith wind farm in the wind farm cluster Parameters to be adjusted.

On the basis of the above embodiment, in the device, the neural network structure used in the training of the prediction model is based on a bidirectional gated recurrent unit GRU network and a graph convolutional network, and specifically includes:

The feature extraction module in the neural network structure used during the training of the prediction model includes a bidirectional gated recurrent unit GRU network and a graph convolution network;

During the training of the prediction model, the sample historical power vector time series and the sample weather forecast parameter matrix time series are input into the two-way gated cyclic unit GRU network to output the power feature after the dimension increase and the weather forecast feature after the dimension increase. Post-power features and post-dimension-upgraded weather forecast features are input into a graph convolutional network and output post-dimension-reduced power features and post-dimension-reduced weather forecast features.

On the basis of the above-mentioned embodiment, in the device, the power feature after dimension increase and the weather forecast feature after dimension increase are input into a graph convolution network to output the power feature after dimension reduction and the weather forecast feature after dimension reduction, specifically including:

The feature transfer process of the graph convolutional network is represented by the following formula:

为邻接矩阵

的度矩阵，X^(l)∈R^n×d，d是升维后功率特征的维度或升维后天气预报特征的维度，X^(l)是第l层图卷积的输入特征，X^(l+1)∈R^n×h是所述图卷积网络中第l层图卷积的输出特征和l+1层图卷积的输入特征，l＝1,2,…,N,所述图卷积网络中的图卷积总层数为N+1，I_n为维度为n的单位矩阵。Among them, A∈Rn ^×n , A represents the adjacency matrix of the wind farm cluster, n represents the total number of wind farms in the wind farm cluster,

is the adjacency matrix

The degree matrix of , X ^(l) ∈ R ^n×d , d is the dimension of the power feature after up-dimension or the dimension of the weather forecast feature after up-dimension, X ^(l) is the input feature of the l-th layer graph convolution, X ^{( l+1)} ∈R ^n×h is the output feature of the lth layer graph convolution and the input feature of the l+1 layer graph convolution in the graph convolutional network, l=1,2,...,N, the The total number of graph convolution layers in a graph convolutional network is N+1, and In is an identity matrix with dimension _n .

On the basis of the above embodiment, in the device, the time step size in the historical power vector time series, the weather forecast parameter matrix time series, and the predicted power vector time series are all equal, and the third target time The segment duration is below the preset threshold.

On the basis of the above embodiment, in the device, the time step is 15 minutes, and the third target time period is 4 hours.

The present invention also provides an improved multi-modal multi-task spatio-temporal graph convolution cluster power prediction method. FIG. 7 is a schematic diagram of the framework of the improved multi-modal and multi-task spatio-temporal graph convolution cluster power prediction method provided by the present invention. As shown in Figure 7, the historical power vector time series (Wind power in Figure 7) and the weather forecast parameter matrix time series (Cubic NWP Windspeed in Figure 7) are respectively processed by the Bi-GRU network to extract the hidden hidden in the time series. The features are upgraded with information, and then the obtained temporary power features and temporary NWP features are input into the graph convolution network GCN. The output value is processed by a multi-modal feature fusion network (Multi-modal Feature Fusion), and the fusion of each wind farm is output. Then, the fusion features of each wind farm are processed using the task layer network of each wind farm (Multi Task Learning), that is, the predicted power vector time series of the final wind farm cluster is output.

By constructing a combined method of bidirectional GRU and graph convolution, the present invention can effectively extract complex spatiotemporal features of wind farm clusters, which are not available in other algorithms. In addition, the present invention is combined with multi-modal learning and multi-task learning theory, wherein multi-modal learning can effectively integrate historical power characteristics of wind farms and numerical weather forecast characteristics, and multi-task learning can greatly improve calculation efficiency. Through the verification of relevant examples, the effectiveness of this method for power prediction of wind farm clusters is proved. Specifically, for the embodiments provided in the above content, they respectively implement the following functions:

1. Since the present invention introduces a method of extracting spatiotemporal features of wind farm clusters in combination with bidirectional GRU and graph convolution, the extracted features can better reflect the complex spatiotemporal correlations of wind farm clusters, thereby improving prediction accuracy.

2. Since the wind speed cube in the numerical weather forecast can effectively reflect the power fluctuation of the wind farm cluster, the present invention also adopts the multimodal fusion method to effectively fuse the characteristics of historical power and the characteristics of numerical weather forecast.

3. The traditional method predicts the wind farms individually, or directly predicts the power of the cluster. The present invention adopts the multi-task learning method, which can predict the power of multiple wind farms at the same time, and provides different settings for different wind farms. The flexibility of the task layer structure can effectively improve the universality of the predicted scenarios.

Experimental results: Taking a provincial wind farm cluster including 20 wind farms as an example, the present invention focuses on comparing the prediction of the network after the bidirectional GRU is used to extract the time series features and the network that does not use the bidirectional GRU to extract the time series features. performance and the performance of several other classical prediction methods, Table 1 is a comparison table of the prediction performance of different methods, and the experimental results are shown in Table 1 below:

Table 1 Comparison of prediction performance of different methods

Figure 8 is a schematic diagram showing the performance comparison of cluster power prediction based on the improved multi-modal multi-task spatiotemporal graph convolution provided by the present invention. As shown in Figure 8, it counts the error (RMSE) of power prediction at intervals of 15 minutes within 4 hours .

At the same time, because the multi-task learning method is adopted, the present invention can not only provide the power prediction result of the wind farm cluster, but also the power prediction result of each wind farm. FIG. 9 is a schematic diagram of the statistical result of power prediction error of each wind farm in the wind farm cluster provided by the present invention. As shown in FIG. 9 , the RMSE of 20 wind farms in the wind farm cluster within 4 hours is respectively counted.

FIG. 10 illustrates a schematic diagram of the physical structure of an electronic device. As shown in FIG. 10 , the electronic device may include: a processor (processor) 1010, a communication interface (Communications Interface) 1020, a memory (memory) 1030, and a communication bus 1040, The processor 1010 , the communication interface 1020 , and the memory 1030 communicate with each other through the communication bus 1040 . The processor 1010 may invoke the logic instructions in the memory 1030 to execute a method for short-term power prediction of a wind farm cluster based on a spatiotemporal graph convolutional neural network, the method comprising: collecting the wind power in the wind farm cluster within a first target time period up to the current moment. The historical power of the wind farm is used to obtain a time series of historical power vectors, and the weather forecast parameter vectors of the wind farms in the second target time period from the current moment are collected to obtain a time series of weather forecast parameter matrices; The weather forecast parameter matrix time series is input into the prediction model, and the predicted power vector time series of the wind farm cluster in the third target time period in the future from the current moment is output; wherein, the prediction model is based on the sample historical power vector time series, The sample weather forecast parameter matrix time series and the corresponding future third target time period predicted power vector time series labels are obtained after training, and the neural network structure used in the training of the prediction model is based on the bidirectional gated recurrent unit GRU network and graph Convolutional network composition.

In addition, the above-mentioned logic instructions in the memory 1030 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product in essence, or the part that contributes to the prior art or the part of the technical solution. The computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, Read-Only Memory (ROM, Read-Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes .

In another aspect, the present invention also provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, when the program instructions are executed by a computer During execution, the computer can execute the method for short-term power prediction of a wind farm cluster based on a spatiotemporal graph convolutional neural network provided by the above methods. Obtain the historical power vector time series from the historical power, collect the weather forecast parameter vectors of the wind farms in the second target time period from the current moment to obtain the weather forecast parameter matrix time series; combine the historical power vector time series with the weather The forecast parameter matrix time series is input to the forecast model, and the forecast power vector time series of the wind farm cluster in the third target time period in the future from the current moment is output; wherein, the forecast model is based on the sample historical power vector time series, the sample weather The prediction parameter matrix time series and the corresponding future third target time period predicted power vector time series labels are obtained after training, and the neural network structure used in the training of the prediction model is based on the bidirectional gated recurrent unit GRU network and graph convolution network composition.

In yet another aspect, the present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored, and the computer program is implemented when executed by a processor to execute the spatiotemporal graph convolutional neural network-based method provided by the above methods. A short-term power prediction method for a wind farm cluster, the method comprising: collecting the historical power of each wind farm in the wind farm cluster within a first target time period up to the current moment to obtain a historical power vector time series, and collecting the second target time period from the current moment The weather forecast parameter vector of each wind farm obtains the weather forecast parameter matrix time series; the historical power vector time series and the weather forecast parameter matrix time series are input into the prediction model, and the output starts from the current moment in the future third target time period The predicted power vector time series of the wind farm clusters in After the labels are trained, the neural network structure used in the training of the prediction model is based on a bidirectional gated recurrent unit GRU network and a graph convolutional network.

The device embodiments described above are only illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in One place, or it can be distributed over multiple network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.

From the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and certainly can also be implemented by hardware. Based on this understanding, the above-mentioned technical solutions can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products can be stored in computer-readable storage media, such as ROM/RAM, magnetic A disc, an optical disc, etc., includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform the methods described in various embodiments or some parts of the embodiments.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be The technical solutions described in the foregoing embodiments are modified, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. a wind farm cluster short-term power prediction method based on spatiotemporal graph convolutional neural network, is characterized in that, comprises: Collect the historical power of each wind farm in the wind farm cluster in the first target time period until the current moment to obtain the historical power vector time series, and collect the weather forecast parameter vector of each wind farm in the second target time period from the current moment to obtain the weather Forecast parameter matrix time series; Inputting the historical power vector time series and the weather forecast parameter matrix time series into a prediction model, and outputting the predicted power vector time series of the wind farm cluster in the third target time period in the future from the current moment; The prediction model is obtained after training based on the sample historical power vector time series, the sample weather forecast parameter matrix time series and the corresponding future third target time period predicted power vector time series labels. The neural network structure used is based on a bidirectional gated recurrent unit GRU network and a graph convolutional network. 2. The short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network according to claim 1, further comprising: The neural network structure used during the training of the prediction model also includes a feature fusion network and a multi-task learning network located behind the graph convolution network; Wherein, the feature fusion network is used to perform feature fusion of the processed historical power features and the processed weather forecast output by the graph convolution network using a multi-modal learning feature splicing method to obtain the power splicing weather features of each wind farm, The multi-task learning network is used to design different task layer structures and loss functions for each wind farm. 3. The short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network according to claim 2, wherein the multi-task learning network is used to design different task layer structures and loss functions for each wind farm, Specifically include: Inputting the power splicing weather features of each wind farm into the task layer network corresponding to each wind farm, wherein the task layer network is a fully connected layer network designed according to the power requirements of each wind farm; The loss function L of the prediction model is determined based on the prediction loss of each task layer network by the following formula:

in,

n represents the total number of wind farms in the wind farm cluster,

is the predicted power of the i-th wind farm, H is the number of time steps in the predicted power vector time series,

is the power splicing weather feature of the ith wind farm in the wind farm cluster, d _out is the dimension of the power splicing weather feature, and W _oi is the fully connected layer network of the ith wind farm in the wind farm cluster Parameters to be adjusted. 4. The short-term power prediction method for wind farm clusters based on spatiotemporal graph convolutional neural networks according to claims 1-3, wherein the neural network structure used during the training of the prediction model is based on a bidirectional gated recurrent unit GRU network and graph convolutional network, including: The feature extraction module in the neural network structure used during the training of the prediction model includes a bidirectional gated recurrent unit GRU network and a graph convolution network; During the training of the prediction model, the sample historical power vector time series and the sample weather forecast parameter matrix time series are input into the two-way gated cyclic unit GRU network to output the power feature after the dimension increase and the weather forecast feature after the dimension increase. Post-power features and post-dimension-upgraded weather forecast features are input into a graph convolutional network to output post-dimension-reduced power features and post-dimension-reduced weather forecast features. 5 . The short-term power prediction method for wind farm clusters based on spatiotemporal graph convolutional neural network according to claim 4 , wherein the power feature after the dimension-up and the weather forecast feature after the dimension-up are input into the graph convolution network and output again. 6 . Power characteristics after dimension reduction and weather forecast characteristics after dimension reduction, including: The feature transfer process of the graph convolutional network is represented by the following formula:

Among them, A∈Rn ^×n , A represents the adjacency matrix of the wind farm cluster, n represents the total number of wind farms in the wind farm cluster,

is the adjacency matrix

The degree matrix of , X ^(l) ∈ R ^n×d , d is the dimension of the power feature after up-dimension or the dimension of the weather forecast feature after up-dimension, X ^(l) is the input feature of the l-th layer graph convolution, X ^{( l+1)} ∈R ^n×h is the output feature of the lth layer graph convolution and the input feature of the l+1 layer graph convolution in the graph convolutional network, l=1,2,...,N, the The total number of graph convolution layers in a graph convolutional network is N+1, and In is an identity matrix with dimension _n . 6 . The short-term power prediction method for wind farm clusters based on spatiotemporal graph convolutional neural network according to claim 5 , wherein the historical power vector time series, the weather forecast parameter matrix time series and the predicted power The time steps in the vector time series are all equal, and the duration of the third target time period is lower than the preset threshold. 7 . The short-term power prediction method for wind farm clusters based on a spatiotemporal graph convolutional neural network according to claim 6 , wherein the time step is 15 minutes, and the third target time period is 4 hours. 8 . 8. A short-term power prediction device for wind farm clusters based on a spatiotemporal graph convolutional neural network, comprising: The collection unit is used to collect the historical power of each wind farm in the wind farm cluster in the first target time period until the current moment to obtain the historical power vector time series, and collect the weather of each wind farm in the second target time period from the current moment The forecast parameter vector obtains the weather forecast parameter matrix time series; a prediction unit, configured to input the historical power vector time series and the weather forecast parameter matrix time series into a prediction model, and output the predicted power vector time series of the wind farm cluster in the third target time period in the future from the current moment; The prediction model is obtained after training based on the sample historical power vector time series, the sample weather forecast parameter matrix time series and the corresponding future third target time period predicted power vector time series labels. The neural network structure used is based on a bidirectional gated recurrent unit GRU network and a graph convolutional network. 9. An electronic device comprising a memory, a processor and a computer program stored on the memory and running on the processor, wherein the processor implements any one of claims 1 to 7 when the processor executes the program. Steps of a described method for short-term power prediction of wind farm clusters based on spatiotemporal graph convolutional neural networks. 10. A non-transitory computer-readable storage medium on which a computer program is stored, characterized in that, when the computer program is executed by a processor, the space-time graph-based volume according to any one of claims 1 to 7 is implemented Steps of a short-term power prediction method for wind farm clusters by integrating neural networks.

Images