Short-term power prediction method for multiple wind farms considering time evolution and spatial correlation

Abstract

The application discloses a kind of short-term power prediction methods of multi-wind farm considering time evolution and space correlation, mainly divided into four modules: input module carries out the collection and pre-processing work of data, object is the historical power and meteorological prediction data of target area multi-wind farm;Time evolution mode tracking module extracts the time series and multi-periodicity time evolution mode of historical wind power data respectively through gate cycle unit and multi-core convolution layer;Space correlation mode attention module introduces time-varying mode attention mechanism to different time evolution modes of multiple space variables and gives correlation weight;Finally, the output module outputs multi-wind farm power day-ahead prediction scenario.The application builds a spatiotemporal fusion multi-wind farm short-term power prediction model with deep learning ability, fully considers the time evolution mode of wind power sequence, makes up for the defects of static extraction of spatial dependence of existing multiple models, and achieves the purpose of improving prediction accuracy and robustness.

Description

Technical Field

[0001] This invention belongs to the field of renewable energy power generation and integrated consumption, specifically involving a short-term power prediction method for multiple wind farms that takes into account time evolution and spatial correlation. Background Technology

[0002] In recent years, my country's installed wind power capacity and its proportion in the power generation sector have continued to grow. However, due to the inherent randomness and uncertainty of wind energy, the problems of wind power absorption and curtailment faced by large-scale wind power grid connection have become increasingly prominent. Accurate short-term wind power forecasting can effectively address the inherent randomness and volatility of wind power, providing a basis for power system dispatching, operation, and optimization decisions, thereby ensuring the safe and economical operation of large-scale wind power grid connection. Therefore, research on short-term wind power forecasting technology has significant practical implications.

[0003] Currently, research on wind power prediction is becoming increasingly in-depth both domestically and internationally, and the prediction methodology system is constantly being improved and expanded. Compared to single wind farms, the challenge of multi-wind farm power prediction lies in the multidimensionality of input variables and the complex and intertwined spatiotemporal relationships. Due to the spatiotemporal continuity of meteorological evolution, the wind speed or wind power among wind farms within a region is closely related. Multiple wind farms not only exhibit various different temporal evolution patterns, but these patterns also show spatial correlation. Utilizing the spatiotemporal correlation of multi-wind farm power to achieve information sharing and mutual benefit can effectively improve the accuracy and efficiency of multi-wind farm power prediction.

[0004] In existing technologies, deep learning methods applied to short-term power prediction for multiple wind farms mainly include convolutional neural networks (CNNs), recurrent neural networks (RNNs), and their variants, which can be used to extract temporal or spatial correlations of wind power sequences. For example, the paper "A Data-Driven and Deep Learning-Based Method for Ultra-Short-Term Wind Power Prediction" discloses the use of multi-scale convolutional neural networks and GRUs to extract the spatiotemporal features of wind farms and mine the spatiotemporal correlations between adjacent wind farms. However, in the problem of short-term wind power prediction, the factors affecting prediction accuracy include not only the temporal and spatial correlations of historical wind power data, but also the potential periodic features of wind power during the time evolution process. Furthermore, the combined effect of wind processes at different time scales may lead to wind power sequences containing multiple time scale periods, but few existing technical documents consider the multi-periodicity of wind farms, failing to extract periodic features or only extracting single-period features. In addition, many existing wind power prediction models, including this one, mainly extract the spatial correlations of multiple wind farms statically, failing to deeply explore the differences and dynamics of the temporal evolution patterns of power in multiple wind farms. Summary of the Invention

[0005] The purpose of this invention is to introduce GRU and multi-kernel convolutional layers to deeply mine the temporal information and potential multi-period characteristics of multiple wind farms, so as to extract the complete temporal evolution pattern of wind power sequence; at the same time, a time-varying pattern attention mechanism is adopted to dynamically pay attention to the correlation of different temporal evolution patterns of multiple spatial variables, so as to achieve accurate and effective prediction of short-term power of multiple wind farms.

[0006] The technical solution of the present invention is as follows:

[0007] A method for short-term power prediction of multiple wind farms that takes into account temporal evolution and spatial correlation includes the following steps:

[0008] Step 1: Collect forecast input data, including hourly historical power data and multi-dimensional meteorological forecast data from multiple wind farms in the target wind power base;

[0009] Step 2: Data preprocessing, which involves normalizing the various types of data used as input and output variables according to their characteristics;

[0010] Step 3: Split the dataset into training, validation, and test sets. First, based on the historical wind power data in the training set, use Fourier decomposition to extract the multi-periodic features of wind power, recording the more obvious cycle lengths T1, T2, ..., T u Determine the kernel size of the multi-kernel convolutional layer;

[0011] Step 4: Combine the normalized data of wind power at the d historical times before the prediction time of multiple wind farms and the meteorological data prediction values ​​at s times on the prediction day into multi-dimensional feature input variables, and use the wind power values ​​at s prediction times of multiple wind farms as output variables to establish a deep spatiotemporal fusion prediction model.

[0012] Step 5: Set the model hyperparameters, initialize the weights and biases, set the loss function, train the deep spatiotemporal fusion network to obtain the optimal weights and biases, and use grid search to optimize the model's optimal hyperparameters using validation set samples.

[0013] Step 6: Input the test samples into the deep spatiotemporal fusion model with the best hyperparameters, and perform inverse normalization on the output prediction results to obtain the power prediction results of multiple wind farms at each time of the prediction day.

[0014] In step 1, the collected input data consists of l-dimensional known features of n wind farms at w time points before the prediction time t, where X∈R n×w×l Each wind farm has an input vector x. t =[P t-d ,P t-d+1 ,...,P t-1 Q t Q t+1,...,Q t+s-1 ]∈R l , where P t-d ,P t-d+1 ,...,P t-1 The wind power at the previous d historical moments, Q t Q t+1 ,...,Q t+s-1 These are multidimensional meteorological forecast data for the last s forecast times.

[0015] Step 2 normalizes the wind power data and meteorological data collected in Step 1. The power data is normalized to the interval [0,1] based on the rated capacity of each wind farm. Wind speed and temperature are normalized using the maximum-minimum normalization method, and wind direction is normalized using the sin / cos trigonometric function method. Let the wind power before and after normalization be x1 and x2 respectively. Wind speed and temperature are x2 and Wind direction data is x3 and The maximum and minimum values ​​of the wind speed and temperature samples are x, respectively. max x min The rated capacity of wind power is x N The specific normalization formula is as follows:

[0016]

[0017] In step 3, based on the measured hourly average total wind power data of multiple wind farms in the target area from the training set, Fourier transforms are performed to plot amplitude-frequency curves. Considering the possible differences in the periodicity of wind power in different seasons, the wind turbine output characteristics are used to divide the season into high-wind and low-wind seasons, and their periodicity is statistically analyzed separately. The frequency components corresponding to the larger amplitude points in the amplitude-frequency curves have a larger proportion in the original sequence. These frequency points are converted into periods according to the sampling frequency to obtain the multi-period values ​​T1, T2, ..., T of the wind power. u This provides a reference for setting the hyperparameters of the model later.

[0018] Step 4: After determining the input and output data, begin building the deep spatiotemporal fusion model. Step 4 includes the following sub-steps:

[0019] Step 4.1: The deep spatiotemporal fusion model first feeds the input matrix into the gated recurrent unit (GRU) to extract the temporal information of historical wind power data. At time t, the GRU receives the current state x. t and the hidden state h from the previous moment t-1 The network output h t It is formed by the dynamic control of update gates and reset gates. A weight matrix W related to the input is defined. r Wu W z The weight matrix R related to the circular connection r ,R u ,R z Bias vector b r ,b u ,b z σ is the sigmoid activation function, tanh is the hyperbolic tangent function, and ⊙ is the dot product. GRU first uses the hidden state h from the previous time step... t-1 and the input x at the current time t To obtain two gating states, where r t To reset the door, z t For updating the gate:

[0020] r t =σ(W r x t +R r h t-1 +b r )

[0021] z t =σ(W z x t +R z h t-1 +b z )

[0022] After obtaining the gating signal, a reset gating is first used to obtain the reset hidden state. Then, this state is concatenated with the input and the activation function tanh is applied to make its data range [-1, 1] to obtain the result.

[0023]

[0024] Next, we selectively forget the hidden state passed from the previous time step and selectively remember the hidden state information containing the current time step, to obtain h. t The update expression is:

[0025]

[0026] Step 4.2: Process the hidden state matrix [h] obtained after GRU processing using multi-kernel convolutional layers. t-w+1 ,h t-w+2 ,...,h t-1 The row vectors of the hidden state matrix h are processed (w represents the time window length) to extract multi-periodic features from multiple wind farms. First, convolution operations are performed on the row vectors of the hidden state matrix h using convolution kernels of u different sizes in k channels to obtain k different feature maps f. map The calculation formula is as follows:

[0027]

[0028] Where * denotes the convolution operation, i represents the i-th row vector of matrix h, Concat denotes the feature concatenation operation, and K denotes the convolution kernel, with its subscripts indicating different kernel sizes (T1, T2, ..., T). u The kernel length (1 represents the kernel width) and channels are then used. A sliding window is used to downsample the feature map into several sub-blocks, resulting in a new feature map f′. map (k). Its calculation formula is as follows:

[0029]

[0030] To fuse feature maps from multi-channel and multi-size convolutional operations, the output vectors of the pooling layers are concatenated along the channel direction. The calculation formula is as follows:

[0031]

[0032] The final step of the multi-kernel convolutional layer is to linearly map the concatenated feature maps into row vectors H of the new state matrix H. i The calculation formula is as follows:

[0033]

[0034] W f and b f This represents the weights and biases in the mapping process.

[0035] Step 4.3: After obtaining the temporal evolution pattern information containing the time series and multi-periodicity of wind power sequences in steps 4.1 and 4.2, a time-varying pattern attention mechanism is used to process the new hidden state matrix H obtained by GRU and multi-kernel convolutional layers to extract relevant information of spatial variables with different temporal evolution patterns. First, a scoring function f for evaluating the correlation is defined, and the obtained attention weights are normalized. The calculation formula is as follows:

[0036] f(H i ,h t )=(H i ) T W a h t

[0037] a i =σ(f(H) i ,h t ))

[0038] Among them W aThis is the weight matrix, which needs to be obtained through neural network training. σ represents the sigmoid activation function. Then, the i-th row vector (containing temporal evolution pattern information) in the hidden state matrix H is combined with the obtained attention weights a. i The weighted summation is calculated as follows:

[0039]

[0040] m represents the number of neurons in the hidden layer.

[0041] Step 4.4: Finally, integrate the feature vector v through a dropout layer and a fully connected layer. t and h t To obtain the final prediction result. The calculation formula is as follows:

[0042] h′ t =W h h t +W v dropout(v t )

[0043]

[0044] W h W v W y The corresponding weight matrix is ​​also obtained through neural network training.

[0045] Step 5: After establishing the deep spatiotemporal fusion model in Step 4, set the hyperparameters, such as the number of neurons m, sample time window length w, number of GRU layers in the temporal network g, number of channels in the multi-kernel convolutional layer k, and convolutional kernel size. Then, initialize the weights and biases, select training samples, use the mean squared error as the loss function, and use the Adam optimization algorithm to train the model to obtain the optimal weight and bias parameters. Then, input the validation set samples into the trained deep spatiotemporal fusion model, and use grid search to optimize the model's optimal hyperparameters based on the validation error. The optimization range of some key hyperparameters is as follows: number of neurons m: {16, 32, 64, 100, 128, 200, 300}; number of convolutional channels k: {16, 24, 32, 48}; number of temporal network layers g: {1, 2, 3}.

[0046] Step 6 involves inputting the test samples into a deep spatiotemporal fusion model with optimal hyperparameters, and then inversely normalizing the output prediction results to obtain the power prediction results for each wind farm at each time point on the prediction day. Where n is the number of wind farms and s is the time step for future prediction.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] 1) This invention considers the temporal evolution and spatial information of multiple wind farms, and outputs short-term power prediction results for multiple wind farms and multiple time steps, which is conducive to the grid-connected operation of large-scale wind power data and has a positive impact on the safe and stable operation of the power grid.

[0049] 2) This invention delves into the potential multi-period characteristics of multiple wind farms and combines multi-kernel convolutional layers that extract multi-period characteristics with GRU units that capture temporal dependencies, enabling the model to learn the complete temporal evolution of multiple wind farms.

[0050] 3) This invention uses a time-varying model attention mechanism to dynamically extract the spatial correlation of multiple time-varying models of multiple wind farms over time, which makes up for the shortcomings of existing models that statically extract spatial dependencies and enhances the stability of prediction.

[0051] 4) Actual calculation results show that the method of the present invention has a reasonable structure and performs well on actual wind power datasets. It also has better prediction accuracy and robustness compared with other popular prediction models.

[0052] 5) The method of the present invention not only fully considers the temporal and periodic evolution patterns of wind power sequences, but also takes into account the spatial correlation information of multiple temporal evolution patterns of different spatial variables, and dynamically considers the spatial correlation of multiple wind farms; finally, it outputs the predicted power values ​​of multiple wind farms at each time of the predicted day that simultaneously meet the requirements of temporal, periodic and spatial correlation. Attached Figure Description

[0053] Figure 1 This is a process for short-term power prediction of multiple wind farms based on a deep spatiotemporal fusion model.

[0054] Figure 2 This is the Fourier transform amplitude-frequency diagram.

[0055] Figure 3 This is a schematic diagram of the gated loop unit.

[0056] Figure 4 This is a schematic diagram of multi-kernel convolution operations.

[0057] Figure 5 This is a comparison chart of the predicted curve and the actual curve of wind farm No. 3 in this embodiment of the invention.

[0058] Figure 6 This is a comparison chart (%) of the MAPE error of prediction results for 18 wind farms using different prediction methods according to an embodiment of the present invention.

[0059] Figure 7 This is a comparison chart (MW) of the RMSE error of prediction results from different prediction methods for 18 wind farms in this embodiment of the invention.

[0060] Figure 8 This is a comparison chart of prediction errors under different periodic characteristics according to an embodiment of the present invention. Detailed Implementation Plan

[0061] This invention provides a short-term power prediction framework for multiple wind farms based on a deep spatiotemporal fusion network, comprising four parts: an input module, a temporal evolution pattern tracking module, a spatial correlation pattern attention module, and an output module. The input module collects and preprocesses data, including historical power data and meteorological forecast data for multiple wind farms in the target area. The temporal evolution pattern tracking module extracts the temporal and multi-periodic temporal evolution patterns of historical wind power data using gated recurrent units and multi-kernel convolutional layers. The spatial correlation pattern attention module introduces a time-varying pattern attention mechanism to assign correlation weights to different temporal evolution patterns of multiple spatial variables, enabling both longitudinal tracking of temporal evolution patterns and horizontal comparison of different temporal evolution patterns of multiple spatial variables. Finally, the output module outputs a day-ahead power prediction scenario for multiple wind farms that simultaneously meets the requirements of temporal sequence, periodicity, and spatial correlation.

[0062] like Figure 1 As shown, the short-term power prediction method for multiple wind farms based on a deep spatiotemporal fusion model includes the following steps:

[0063] Step 1: Collect hourly historical power data and external meteorological information data for multiple target wind farms in the target area. The raw input data consists of l-dimensional known features of n wind farms at w time points before the prediction time t, where X∈R n×w×l Each wind farm has an input vector x. t =[P t-d ,P t-d+1 ,...,P t-1 Q t Q t+1 ,...,Q t+s-1 ]∈R l , where P t-d ,P t-d+1 ,...,P t-1 The wind power at the previous d historical moments, Q t Q t+1 ,...,Q t+s-1 These are multidimensional meteorological forecast data (such as wind speed, wind direction, and temperature) for the last s forecast times.

[0064] Step 2: Normalize the wind power data and meteorological data collected in Step 1. The power data is normalized to the interval [0,1] based on the rated capacity of each wind farm. Wind speed and temperature are normalized using the maximum-minimum normalization method, and wind direction is normalized using the sin / cos trigonometric function method. Let the wind power before and after normalization be x1 and x2. Wind speed and temperature are x2 and Wind direction data is x3 and The maximum and minimum values ​​of the wind speed and temperature samples are x, respectively. max x min The rated capacity of wind power is x N The specific normalization formula is as follows:

[0065]

[0066]

[0067]

[0068] Step 3: Divide the input data into training, validation, and test sets according to a ratio of 80%, 10%, and 10%, respectively. Based on the measured hourly average total wind power data of multiple wind farms in the target area from the training set, perform Fourier transform on the data and plot the amplitude-frequency curve as shown below. Figure 2 As shown. Considering the potential differences in wind power periodicity across different seasons, the wind turbine output characteristics are used to divide the season into high-wind and low-wind seasons, and their periodicity patterns are statistically analyzed separately. The frequency components corresponding to points with larger amplitudes in the amplitude-frequency curves have a larger proportion in the original sequence. Converting these frequency points into periods based on the sampling frequency yields the multi-period values ​​T1, T2, ..., T of the wind power. u This provides a reference for setting the hyperparameters (kernel size of multi-kernel convolution) of the model later;

[0069] Step 4: After determining the input and output data, begin building the deep spatiotemporal fusion model. Step 4 includes the following sub-steps:

[0070] Step 4.1: The deep spatiotemporal fusion model first feeds the input matrix into the GRU to extract the temporal information of historical wind power data. At time t, the GRU receives the current state x. t and the hidden state h from the previous moment t-1 The network output h t It is formed by the dynamic control of update gates and reset gates. A weight matrix W related to the input is defined. r W u W z The weight matrix R related to the circular connection r ,R u ,R z Bias vector b r ,b u ,b z σ is the sigmoid activation function, tanh is the hyperbolic tangent function, and ⊙ is the dot product. GRU first uses the hidden state h from the previous time step... t-1 and the input x at the current timet To obtain two gating states, where r t To reset the door, z t For updating the gate:

[0071] r t =σ(W r x t +R r h t-1 +b r )

[0072] z t =σ(W z x t +R z h t-1 +b z )

[0073] After obtaining the gating signal, a reset gating is first used to obtain the reset hidden state. Then, this state is concatenated with the input and the activation function tanh is applied to make its data range [-1, 1] to obtain the result.

[0074]

[0075] Next, we selectively forget the hidden state passed from the previous time step and selectively remember the hidden state information containing the current time step, to obtain h. t The update expression is:

[0076]

[0077] Step 4.2: Process the hidden state matrix [h] obtained after GRU processing using multi-kernel convolutional layers. t-w+1 ,h t-w+2 ,...,h t-1 The row vectors of the hidden state matrix h are processed (w represents the time window length) to extract multi-periodic features from multiple wind farms. First, convolution operations are performed on the row vectors of the hidden state matrix h using convolution kernels of u different sizes in k channels to obtain k different feature maps f. map The calculation formula is as follows:

[0078]

[0079] Where * denotes the convolution operation, i represents the i-th row vector of matrix h, Concat denotes the feature concatenation operation, and K denotes the convolution kernel, with its subscripts indicating different kernel sizes (T1, T2, ..., T). u The kernel length (1 represents the kernel width) and channels are then used. A sliding window is used to downsample the feature map into several sub-blocks, resulting in a new feature map f′. map(k). Its calculation formula is as follows:

[0080]

[0081] To fuse feature maps from multi-channel and multi-size convolutional operations, the output vectors of the pooling layers are concatenated along the channel direction. The calculation formula is as follows:

[0082]

[0083] The final step of the multi-kernel convolutional layer is to linearly map the concatenated feature maps into row vectors H of the new state matrix H. i The calculation formula is as follows:

[0084]

[0085] W f and b f This represents the weights and biases in the mapping process.

[0086] Step 4.3: After obtaining the temporal evolution pattern information of the wind power sequence in steps 4.1 and 4.2, a time-varying pattern attention mechanism is used to process the new hidden state matrix H obtained by GRU and multi-kernel convolutional layers to extract relevant information of spatial variables with different temporal evolution patterns. First, a scoring function f for evaluating the correlation is defined, and the obtained attention weights are normalized. The calculation formula is as follows:

[0087] f(H i ,h t )=(H i ) T W a h t

[0088] a i =σ(f(H) i ,h t ))

[0089] Among them W a This is the weight matrix, which needs to be obtained through neural network training. σ represents the sigmoid activation function. Then, the i-th row vector (containing temporal evolution pattern information) in the hidden state matrix H is combined with the obtained attention weights a. i The weighted summation is calculated as follows:

[0090]

[0091] m represents the number of neurons in the hidden layer.

[0092] Step 4.4: Finally, integrate the feature vector v through a dropout layer and a fully connected layer.t and h t To obtain the final prediction result. The calculation formula is as follows:

[0093] h′ t =W h h t +W v dropout(v t )

[0094]

[0095] W h W v W y The corresponding weight matrix is ​​also obtained through neural network training.

[0096] Step 5: After establishing the deep spatiotemporal fusion model, set the hyperparameters, such as the number of neurons m, sample time window length w, number of GRU layers in the temporal network g, number of channels in the multi-kernel convolutional layer k, and convolutional kernel size; then initialize the weights and biases, select training samples, use the mean squared error as the loss function, and use the Adam optimization algorithm to train the model to obtain the optimal weight and bias parameters; then input the validation set samples into the trained deep spatiotemporal fusion model, and use grid search to optimize the model's hyperparameters based on the validation error. The optimization ranges for some key hyperparameters are: number of neurons m: {16, 32, 64, 100, 128, 200, 300}; number of convolutional channels k: {16, 24, 32, 48}; number of temporal network layers g: {1, 2, 3}.

[0097] Step 6: Input the test samples into the deep spatiotemporal fusion model with optimal hyperparameters, and inversely normalize the output prediction results to obtain the power prediction results for each wind farm at each time of the prediction day. Where n is the number of wind farms and s is the time step for future prediction.

[0098] In this embodiment, hourly measured power data from 18 wind farms in a domestic target wind power base in 2016 are used as the analysis object. The input to the prediction model is the hourly historical wind power data from one week in advance and the meteorological forecast data for the prediction day (including hourly wind speed, wind direction, and temperature forecasts). The output is the multi-wind farm power prediction results for the next 24 hours. This embodiment uses Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) to evaluate the prediction accuracy. Their calculation formulas are as follows:

[0099]

[0100]

[0101] In the formula: n is the number of test samples; y i and These are the actual and predicted wind power values ​​at the i-th sampling point on the prediction day, respectively.

[0102] Figure 5 The image shows a comparison between the predicted curve and the actual curve for wind farm No. 9 in this embodiment of the invention. It can be seen that the method of the present invention performs best in wind farms, accurately simulating the trend of wind power fluctuations. It also exhibits smaller overall error and smaller single-point prediction error. Figure 6 and Figure 7 The figures show a comparison of MAPE and RMSE errors for the prediction results of 18 wind farms using different prediction methods according to embodiments of the present invention. It can be seen that the method of the present invention has the best prediction accuracy for day-ahead wind power. Compared with temporal neural networks such as TCN, GRU, and LSTM, the average MAPE improvement rate (Imp) of DSTFM is 9.99%, 8.80%, and 7.87%, respectively, indicating that considering only the temporal characteristics of wind power is insufficient; the complete temporal evolution pattern and the spatial correlation between multiple wind farms must also be taken into account. GRU-CNN incorporates convolutional kernels that can extract input spatial features, enabling the model to extract spatiotemporal information. Compared with these models, the average MAPE of DSTFM is reduced by 5.17%, demonstrating the effectiveness of applying a time-varying pattern attention mechanism to dynamically capture the spatial correlation of multiple wind farms with multiple time-varying patterns. Figure 6 It can also be seen that due to the different geographical locations of the wind farms, the MAPE of each model varies across different wind farms. The MAPE of all models is relatively high for wind farms 4 and 8, indicating that the power predictability of these two wind farms is poor. Nevertheless, DSTFM still maintains strong robustness to multiple wind farms with varying predictability.

[0103] Figure 8 This chart compares the impact of different periodic characteristics of wind power on prediction results, comparing three schemes: ignoring periodic characteristics, considering only a single period, and considering a combination of multiple periods. The multi-period combined DSTFM model has the smallest prediction error, with a 0.75% reduction in MAPE error compared to the model without considering periodicity, demonstrating the necessity of incorporating the periodicity of wind power data. Compared to the three models considering only single-period characteristics, taking 24h, 38h, and 80h as examples, the MAPE improvement rates of the proposed model are 0.51%, 0.40%, and 0.57%, respectively. Although the prediction error of the single-period model is lower than that of the model without incorporating periodic characteristics, the improvement effect is limited. Furthermore, the stability of the model decreases with the length of the period, indicating that the model's sensitivity to changes in the temporal evolution pattern within the period decreases with longer periods. The multi-period combined model can take into account the sensitivity to different periodic characteristics, thus achieving higher prediction accuracy.

Claims

1. A method for short-term power prediction of a multi-wind farm taking into account temporal evolution and spatial correlation, characterized in that, Comprising the following steps: Step 1: Collecting prediction input data, collecting historical power data of multiple wind farms at hourly level and multi-dimensional meteorological prediction data of a target wind power base; Step 2: Data preprocessing, normalizing various data as input variables and output variables according to their characteristics respectively; Step 3: The data set is divided into a training set, a validation set and a test set. First, the Fourier decomposition is used to extract the multi-periodicity characteristics of wind power based on the historical wind power data of the training set, and the more obvious period lengths are recorded T 1, T 2,..., T u , determine the convolution kernel size of the multi-core convolution layer; Step 4: The wind power of multiple wind farms at a historical time point before the prediction time point is predicted d s The normalized data set of the predicted value of the meteorological data at the prediction day is composed into a multi-dimensional feature input variable, and the wind power value at the prediction time point of the multiple wind farms is taken as an output variable to establish a deep spatio-temporal fusion prediction model s The normalized data set of the predicted value of the meteorological data at the prediction day is composed into a multi-dimensional feature input variable, and the wind power value at the prediction time point of the multiple wind farms is taken as an output variable to establish a deep spatio-temporal fusion prediction model​ Step 5: Setting model hyperparameters, initializing weights and biases, setting loss function, training deep spatio-temporal fusion network to obtain optimal weight and bias parameters, and using grid search to optimize the best hyperparameters of the model through validation set samples; Step 6: Inputting test samples into the deep spatio-temporal fusion model with the best hyperparameters, and performing inverse normalization on the output prediction results to obtain the power prediction results of multiple wind farms at each time of the prediction day; Step 4 includes the following sub-steps: Step 4.1: The deep spatiotemporal fusion model first sends the input matrix into the gated recurrent unit (GRU) to extract the temporal information of the historical wind power data. At each time point, the GRU receives the current state t and the hidden state x of the previous time point t , h t-1 , the output of the network h t is formed by the dynamic control of the update gate and the reset gate, and the definition of the weight matrix related to the input , the weight matrix related to the recurrent connection R r , R u , R z , the bias vector b r , b u , b z , is the sigmoid activation function, tanh is the hyperbolic tangent function, is the dot product, the GRU first obtains two gate states through the hidden state h t-1 of the previous time point and the input x t of the current time point, wherein r t is the reset gate, z t is the update gate: ； ； After obtaining the gating signal, the reset gating is used to obtain the reset hidden state, and then it is spliced with the input and passed through the activation function tanh to make the data range [-1, 1], and the following formula is obtained : ； Then, the hidden state delivered at the last time is selectively forgotten and the hidden state information containing the current time is selectively remembered, to obtain The update expression is: ； Step 4.2: Process the row vectors of the hidden state matrix h t-w+1 , h t-w+2 ,..., h t-1 ] processed by GRU through multi-kernel convolution layer, where w represents the length of the time window, in order to extract the multi-periodic characteristics of multiple wind farms, first use k kinds of convolution kernels with size u to perform convolution operation on the row vectors of the hidden state matrix h , and get k different feature maps f map , the calculation formula is as follows: ； wherein is a convolution operation, i denotes a matrix The i th is a row vector, Concat denotes a feature concatenation operation, K denotes a convolution kernel, the subscript of which denotes different sizes and channels of the convolution kernel, specifically T 1, T 2, …, T u denotes the length of the convolution kernel, 1 denotes the width of the convolution kernel; then a down-sampling operation is performed on a plurality of sub-blocks divided by the sliding window on the feature map, i.e. the down operation in the following formula, to obtain a new feature map , the calculation formula of which is as follows: ； In order to fuse the multi-channel and multi-size convolution operation feature maps, the output vectors of the pooling layer are spliced in the channel direction to obtain The calculation formula is as follows: ； The last step of the multi-core convolution layer is to linearly map the concatenated feature maps to a new state matrix H a row vector of H i whose computational formula is as follows: ； W f and b f represent weights and biases of the mapping process; Step 4.3: After the time evolution pattern information of wind power sequence time series and multi-periodicity is obtained by the two-step processing of steps 4.1 and 4.2, the new hidden state matrix obtained by the GRU and multi-core convolution layer is processed by the time-varying pattern attention mechanism H The processing is performed to extract the relevant information of the spatial variable with different time evolution patterns, and first, a score function for evaluating the correlation is defined f , and the obtained attention weight is normalized, and the calculation formula is as follows: ； ； wherein W a is a weight matrix, which needs to be obtained through neural network training, and σ represents a sigmoid activation function; and then the hidden state matrix H The first column vector in the matrix i contains time evolution pattern information, and the obtained attention weight a i is weighted and summed, and the calculation formula is as follows: ； m represents the number of hidden layer neurons; Step 4.4: The feature vector is finally integrated through a dropout layer and a fully connected layer v t and h t to obtain the final prediction result whose calculation formula is as follows: ； ； W h , W v , W y denote the corresponding weight matrices, which are also obtained by training the neural network.

2. The method of claim 1, wherein: In step 1, the collected input data is processed by n The wind farm at the prediction time t The w The l dimensional known features compositions, X ∈ R n×w×l ; for each wind farm there is an input vector x t [ P t-d ,P t-d+1 ,...,P t-1 ,Q t ,Q t+1 ,...,Q t+s-1 ]∈ R l wherein P t-d , P t-d+1 ,...,P t-1 are the wind power at the previous d historical time instants, Q t, Q t+1 ,...,Q t+s-1 are the multi-dimensional weather forecast data at the following s forecast time instants.

3. The method of claim 1, wherein: Step 2: Normalization is performed on the wind power data and meteorological data collected in step 1, respectively. The power data is normalized to the interval [0, 1] based on the rated capacity of each wind farm. The wind speed and temperature are normalized using the maximum and minimum normalization method, and the wind direction is normalized using the sin / cos trigonometric function normalization method. Let the normalized wind power be x 1 and , the wind speed and temperature be x 2 and , and the wind direction data be x 3 and . The maximum and minimum values of the wind speed and temperature samples are x max , x min , and the rated capacity of the wind power is x N . The specific normalization formula is as follows: 。 4. The method of claim 1, wherein: In step 3, based on the measured data of the target area of the training set of multiple wind power plants at the hourly level, the Fourier transform is carried out, and the amplitude-frequency curve is drawn. Considering that there may be differences in the periodicity of wind power in different seasons, the wind power is divided into seasons with more wind and seasons with less wind according to the output characteristics of the wind turbine, and the periodicity of each season is counted. The frequency components corresponding to the larger amplitude points in the amplitude-frequency curve have a larger proportion in the original sequence. These frequency points are converted into periods according to the sampling frequency to obtain the multi-period value of the wind power T 1, T 2, …, T u , which provides a reference for the setting of the model hyperparameters.

5. The method of claim 1, wherein: Step 5: After establishing the deep spatiotemporal fusion model in step 4, the hyperparameters are set, which are the number of neurons m , sample time window length w , number of GRU layers in the time sequence network g , number of channels in the multi-core convolutional layer k , and convolution kernel size; then initialize the weights and biases, select the training samples, use mean square error as the loss function, and use the Adam optimization algorithm to train the model to obtain the optimal weight and bias parameters; The verification set samples are input into the trained deep spatio-temporal fusion model, and the best super parameter of the model is optimized according to the verification error by using a grid search, and the optimization range of the super parameter is as follows: the number of neurons m : 16, 32, 64, 100, 128, 200, 300; the number of convolution channels k : 16, 24, 32, 48; the number of time sequence network layers g : 1, 2, 3.

6. The method of claim 1, wherein: Step 6: input the test sample into the deep spatio-temporal fusion model with the best hyperparameters, and perform inverse normalization on the output prediction result to obtain the power prediction result of each wind farm at each time of the prediction day wherein n is the number of wind farms, s is the time step of the prediction.

Images