Wind power prediction method based on weather forecast and convolution simplified LSTM

Abstract

The invention provides a wind power prediction method based on weather forecast and convolution simplified LSTM (Long Short Term Memory). The method comprises the following steps: analyzing the discrete degree of a historical wind power monitoring data sequence, verifying the integrity of historical wind power monitoring data, identifying abnormal data possibly influencing the accuracy of a model, repairing the abnormal data, and normalizing all data; meteorological factors highly related to the wind power are screened out; acquiring future forecast data of meteorological factors highly related to wind power, and combining the future forecast data with historical wind power monitoring data to serve as input features of the prediction model; on the basis of LSTM, local spatial-temporal characteristics of a wind power monitoring data sequence are extracted, a peephole structure and a novel gating structure are introduced, long-term dependence of time sequence data is captured by using an LSTM unit, ConvSLSTM is obtained, wind power prediction is performed, a prediction result is subjected to reverse normalization, and an evaluation index is calculated. The prediction precision of the method on an instance data set is superior to that of a traditional model, and the method has high application potential and theoretical value.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and in particular to a wind power prediction method based on weather forecast and convolution simplified LSTM. Background Art

[0002] As global energy demand continues to rise, wind power, a renewable energy source, is becoming a key force driving the structural transformation of the power system. According to relevant statistics, global installed renewable energy capacity reached 3,064 GW in 2021, of which wind power generation accounted for 842 GW. As an important green energy source, wind power generation plays a vital role in reducing greenhouse gas emissions and enhancing the sustainability and security of energy supply. However, wind power is characterized by volatility and uncertainty, and its integration with traditional power systems faces numerous challenges. In particular, the fluctuation of wind power puts significant pressure on the grid's load balancing, backup power scheduling, and system stability. For this reason, wind power forecasting, especially ultra-short-term wind power forecasting, has become a research hotspot of common concern in academia and industry in recent years.

[0003] In recent years, deep learning models such as long short-term memory (LSTM), convolutional neural networks (CNN), and graph convolutional networks (GCN) have gradually attracted widespread attention and in-depth research in the field of wind power forecasting. Deep learning methods can effectively capture nonlinear relationships in data through automatic learning and feature extraction from large amounts of data, significantly improving forecast accuracy. Research has shown that deep learning models have significant advantages in wind power forecasting, especially in handling complex features and capturing nonlinear relationships. Furthermore, wind power forecasting performance is often affected by the quality of input data and feature selection methods. Current research shows that meteorological forecast data not only contains key information such as future wind speed and direction but also reflects the underlying trends in wind power changes, providing a more comprehensive reference for models. Incorporating meteorological forecast data can effectively compensate for the information lag that can result from relying solely on historical data, further improving forecast accuracy. Therefore, the rational integration of meteorological forecast data with deep learning models to fully exploit the contribution of meteorological factors to wind power forecasting has become an important direction for improving model performance. Summary of the Invention

[0004] The purpose of the present invention is to solve the technical problems in the above background and propose a wind power prediction method based on weather forecast and convolutional simplified LSTM, which includes the following steps:

[0005] S1. Verify the integrity of historical wind power monitoring data by analyzing the discreteness of the historical wind power monitoring data series and identify abnormal data that may affect the accuracy of the model;

[0006] S2. Repair the detected abnormal data to ensure the continuity and integrity of the input data;

[0007] S3, normalizing the historical wind power monitoring data after the abnormal data is repaired;

[0008] S4. Screen out meteorological factors that are highly correlated with wind power among meteorological factors of wind farms through correlation analysis method;

[0009] S5. Obtain future forecast data of meteorological factors that are highly correlated with wind power and combine them with historical wind power monitoring data as input features for the prediction model to fully utilize historical trends and meteorological forecast information;

[0010] S6. Based on LSTM, a convolutional layer is used to extract the local spatiotemporal features of the wind power monitoring data sequence. A new gating structure with a peephole structure and cross-coupling is introduced to improve the model's ability to process high-dimensional complex data. At the same time, LSTM units are used to capture the long-term dependencies of time series data, resulting in ConvSLSTM.

[0011] S7. Use the obtained ConvSLSTM to predict wind power, denormalize the prediction results, and calculate the evaluation index.

[0012] In a preferred solution, step S1 further includes the following steps:

[0013] S11. Use a box plot to detect abnormal data. Arrange the historical wind power monitoring data from small to large and divide them into four parts. The three division points correspond to quartiles. The second quartile Q2 represents the median of the data. Its calculation formula is:

[0014]

[0015] The first and third quantiles correspond to the values ​​in the data and The value of the position, for different wind power data amounts m, is calculated as follows:

[0016] When m = 4k + 3 (k = 0, 1, ...), the calculation formulas for the first and third quantiles are:

[0017] Q1=x (m+1) / 4 ,Q3=x 3(m+1) / 4 .;

[0018] When m = 4k + 1 (k = 0, 1, ...), the calculation formulas for the first and third quantiles are:

[0019]

[0020] Based on the above calculation, the interquartile range I is obtained. i :

[0021] I i =Q3-Q1;

[0022] By setting the upper and lower limits W u,i 、W d,i Used to identify outliers, the formula is:

[0023] W u,i =Q3+1.5I i

[0024] W d,i =Q1-1.5I i ;

[0025] Defined in the range [W u,i ,W d,i ] are normal data, and those outside this range are considered abnormal data.

[0026] In a preferred embodiment, step S2 further includes the following steps:

[0027] Since the intermittent missing data are redundant and the data at adjacent moments will not mutate, the K nearest neighbor complementary method is used to reconstruct the intermittent missing data:

[0028] For a given wind power output data, the algorithm obtains the average value of the K nearest neighbor data near the missing data and replaces the missing data with the average value to repair the intermittent missing data. The formula is described as:

[0029]

[0030] Where, X i is the missing data location of the data sample, x i-K is the Kth data before the missing data, x i+K It is the Kth data after the missing data.

[0031] In a preferred embodiment, the correlation analysis method in step S4 includes the following steps:

[0032] S41. By placing the two sets of dominant component error sequences in the same coordinate system to form a scatter plot, and then gridding the coordinate system, the mutual information value is calculated based on the marginal probability and joint probability of the variable scatter points falling in the grid. Finally, the sum of the mutual information values ​​of all points is obtained as a measure of the correlation between the two sets of dominant component errors. Given two sets of dominant component error sequences, the mutual information I(X,Y) is defined as:

[0033]

[0034] Where p(x,y) is the joint probability of variables x and y, p(x) and p(y) are the marginal probabilities of variables x and y respectively;

[0035] S42. Place the given finite set in the same coordinate system and divide it into a grid G. By changing the number of columns a and the number of rows b of the grid, various grids can be generated, thereby obtaining different mutual information values. The maximum mutual information value is defined as:

[0036]

[0037] S43. Normalize the maximum mutual information values ​​on different grids G to obtain the feature matrix:

[0038]

[0039] On this basis, the maximum information coefficient of the error data set D is defined as:

[0040]

[0041] In a preferred solution, the correlation analysis method in step S4 includes:

[0042] S401, performing a preprocessing operation on historical wind power monitoring data and related meteorological data to eliminate noise and missing values ​​in the historical wind power monitoring data and related meteorological data;

[0043] S402, using the maximum information coefficient (MIC) to calculate the correlation coefficient between each meteorological factor and wind power, preliminarily screening out meteorological factors with a strong correlation with wind power, and obtaining a preliminary set of key meteorological factor data;

[0044] S403: Integrate the initially screened key meteorological factor data with historical wind power monitoring data, cluster and group wind farm units using a Gaussian mixture model, determine the optimal number of unit groups based on the Bayesian Information Criterion, and divide the wind farm into several groups of units with similar output characteristics;

[0045] S404: For each turbine group, the MIC value is combined with the characteristic weight determined based on the physical model or domain knowledge to calculate the fusion weight of each meteorological factor, and further determine the meteorological factors that are highly correlated with wind power.

[0046] In the preferred solution, step S404 specifically includes: combining the MIC value with the feature weight determined based on the physical model or domain knowledge, setting the MIC value to MIC ij , represents the correlation coefficient between the i-th characteristic factor and wind power in the j-th time window; the characteristic weight determined according to the physical model or domain knowledge is wi , then the feature correlation weight after fusion R i Expressed as:

[0047]

[0048] Among them, m is the number of feature factors and n is the number of time windows.

[0049] In a preferred solution, step S5 includes:

[0050] Obtain weather forecast data for the next N steps, and simultaneously extract M steps of historical wind power monitoring data containing meteorological characteristics and wind power. Then, combine the historical wind power monitoring data with the forecast data to establish a unified forecast input framework for predicting wind power for the next N steps.

[0051] In a preferred solution, step S5 further includes:

[0052] The historical wind power monitoring data and future forecast data are spliced ​​on the time axis; with the current time as the boundary, the historical wind power monitoring data of the past several hours and the weather forecast data of the next several hours are arranged and combined in the time dimension in chronological order to form a complete input feature sequence;

[0053] At each time point, several meteorological factor data are combined into a feature vector as the input feature of the prediction model. The number of selected meteorological factors is equal to the dimension of the feature vector at each time point.

[0054] In the preferred solution, in step S6, the construction process of ConvSLSTM includes:

[0055] To further reduce the number of LSTM parameters to be optimized and its training time, a new gating structure with a cross-coupling structure is used to replace the input gate, forget gate, and output gate in the LSTM network.

[0056] The introduction of convolution operations improves the ability to process spatiotemporal data and reduces model training time by sharing parameters of various gate structures;

[0057] A peephole structure is introduced to consider the impact of storage units on network output;

[0058] The forward propagation process of ConvSLSTM is shown as follows:

[0059]

[0060] In the formula, * is the convolution operation, is the Hadamard product operation, W x is the input weight, W h is the cycle weight, W cis the peephole weight, b i 、b c and b o is the input bias, σ(·) is the activation function, tanh(·) is the hyperbolic activation function, i t 、c t 、o t and h t Represent the input gate, memory cell state, output gate and network output of LSTM respectively, h t-1 Indicates the network output at the previous time, c t-1 Indicates the state of the memory unit at the previous time, x t Represents the input of the network.

[0061] In a preferred embodiment, step S7 includes the following steps:

[0062] S701. Denormalize the prediction result. The formula is:

[0063] X=(X norm *(X max -X min ))+X min ;

[0064] Where, X is the original wind power data, X norm is the normalized data; X max and X min are the maximum and minimum values ​​of the original data respectively;

[0065] S702: Root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) are used as evaluation indicators. The calculation formulas are:

[0066]

[0067]

[0068] Where y i represents the true value, represents the predicted value, n is the number of samples, Represents the mean of the true values.

[0069] The wind power prediction method based on weather forecast and convolutional simplified LSTM proposed in this invention has the following technical effects:

[0070] 1) To address the outlier problem in wind power data processing, the present invention improves data quality by verifying the integrity and rationality of historical data, identifying outliers using box plots, and reconstructing missing data through the K-nearest neighbor algorithm, thereby providing reliable data support for subsequent model training.

[0071] 2) Considering that high-dimensional redundant features in wind power prediction may affect model performance, the present invention introduces the maximum information coefficient. By calculating the correlation between meteorological factors and wind power, key features are screened out and redundant features are eliminated, thereby simplifying the model input and improving the model's generalization ability and prediction efficiency.

[0072] 3) This invention also incorporates weather forecast data to further enhance the model's predictive performance. Finally, this invention proposes a convolutional simplified long-short-term memory network method, combining the advantages of CNNs and LSTMs to enhance the model's ability to capture spatiotemporal relationships. Furthermore, the introduction of a novel cross-coupled gating structure and peephole architecture improves the model's memory capacity and predictive accuracy.

[0073] 4) By combining MIC and clustering methods, we can more effectively identify meteorological factors that have a long-term impact on wind power, more accurately grasp the changing patterns of wind power under different combinations of meteorological conditions, and improve the accuracy of correlation analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a box plot structure diagram of the present invention.

[0075] Figure 2 It is a framework diagram for fusing weather forecast data and historical data of the present invention.

[0076] Figure 3 1 is a structural diagram of the LSTM unit of the present invention.

[0077] Figure 4 4 is a structural diagram of the ConvLSTM unit of the present invention.

[0078] Figure 5 This is a structural diagram of the ConvSLSTM unit of the present invention.

[0079] Figure 6 It is a prediction flow chart of the present invention.

[0080] Figure 7 This is the maximum information coefficient correlation result diagram of the present invention

[0081] Figure 8 This is a comparison chart of prediction results of different models of the present invention. DETAILED DESCRIPTION

[0082] The present invention proposes a wind power prediction method based on weather forecast and convolution simplified LSTM. First, in order to address the common outlier problem in wind power data, the box plot method is used to detect outliers in the data, and the detected outliers are corrected in combination with the K-nearest neighbor complementary method, thereby improving the quality and credibility of the data. Secondly, in order to simplify the model input and improve the feature correlation, MIC is used to screen out meteorological factors that are strongly correlated with wind power, such as wind speed and hub wind speed. Then, based on LSTM, convolution operations are introduced to enhance the feature extraction capability, and a new cross-coupled gating structure and peephole structure are designed to reduce network parameters and improve the training efficiency of the model. Finally, combining historical data and numerical weather forecast data, the ConvSLSTM model is used to perform high-precision prediction of wind power series.

[0083] The wind power prediction method based on weather forecast and convolution simplified LSTM first uses the box plot method to detect outliers, and combines the K nearest neighbor complementary method to correct the detected outliers. The maximum information coefficient (MIC) is used to screen out meteorological factors that are highly correlated with wind power. Based on LSTM, a new gating structure and peephole mechanism with convolution operation and cross coupling are introduced to obtain ConvSLSTM. Finally, the ConvSLSTM model is used to make high-precision predictions of wind power series by combining historical data with numerical weather forecast data. The prediction process is as follows: Figure 6 shown.

[0084] S1, the box plot method is used to detect outliers in wind power data. The data are arranged from small to large and divided into four parts. The three split points correspond to the quartiles, e.g. Figure 1 As shown. Among them, the second quantile Q2 represents the median of the data, and its calculation formula is:

[0085]

[0086] The first and third quantiles correspond to the values ​​at the 25% and 75% positions in the data, respectively. For different wind power data amounts m, the calculation method is as follows:

[0087] When m = 4k + 3 (k = 0, 1, ...), the calculation formulas for the first and third quantiles are:

[0088] Q1=x (m+1) / 4 ,Q3=x 3(m+1) / 4 .

[0089] When m = 4k + 1 (k = 0, 1, ...), the calculation formulas for the first and third quantiles are:

[0090]

[0091] Based on the above calculation, the interquartile range Ii can be obtained:

[0092] I i =Q3-Q1

[0093] By setting the upper and lower limits Wu,i and Wd,i to identify outliers, the formulas for these two values ​​are:

[0094] W u,i =Q3+1.5I i

[0095] W d,i =Q1-1.5I i

[0096] The data within the range [Wd,I,Wu,i] are defined as normal values, and the data outside this range are defined as abnormal values.

[0097] S2 fills in the missing values ​​after S1 removes the outliers. Because intermittent missing data is redundant and the data at adjacent moments do not undergo sudden changes, the K-nearest neighbor complementary method is used to reconstruct the intermittent missing data. For a given wind power output data, the algorithm obtains the average value of the K nearest neighbor data near the missing value and replaces the missing value with this average value to achieve intermittent missing data repair. The formula is described as follows:

[0098]

[0099] Where Xi is the missing value position of the data sample, xi-K is the Kth data before the missing value, and xi+K is the Kth data after the missing value.

[0100] S3, normalize the wind power data to the interval [0,1]. The normalization formula is:

[0101]

[0102] Where, X is the original wind power data; X norm is the normalized data; X max and X min are the maximum and minimum values ​​of the original data respectively;

[0103] S4, considering that the historical information of the wind farm contains meteorological information such as wind speed, temperature, air pressure and historical power data, in order to avoid input redundancy, MIC is used to filter out the relevant meteorological data. The heat map of the correlation coefficient between each factor and power is shown as follows: Figure 7 As shown by Figure 7 It can be seen that among the main factors affecting wind power, wind speed has the greatest correlation, so the wind speed at different heights of the wind tower and the wind speed at the generator hub are selected as multi-factor variables.

[0104] MIC is a scatter plot method that places two sets of dominant component error sequences in the same coordinate system. This coordinate system is then gridded and the mutual information (MI) value is calculated based on the marginal and joint probabilities of the variable's scatter plot falling within the grid. The sum of the mutual information values ​​for all points is used as a measure of the correlation between the two sets of dominant component errors. Mutual information is one of the most effective metrics for measuring nonlinear correlations between variables. Given two sets of dominant component error sequences, the mutual information (I(X,Y)) is defined as follows:

[0105]

[0106] Where p(x,y) is the joint probability of variables x and y, and p(x) and p(y) are the marginal probabilities of variables x and y, respectively.

[0107] The given finite set is placed in the same coordinate system and divided by the grid G. By changing the number of columns a and rows b of the variable grid, various grids can be generated, thereby obtaining different mutual information values, where the maximum mutual information value is defined as:

[0108]

[0109] In order to facilitate comparison between different dimensions, it is necessary to normalize the maximum mutual information values ​​on different grids G to obtain the feature matrix:

[0110]

[0111] On this basis, the maximum information coefficient of the error data set D is defined as:

[0112]

[0113] Where n is the number of elements in the error sequence.

[0114] S5, obtain the weather forecast data for the next N steps, and extract the M steps of historical data containing weather characteristics and wind power. Then combine the historical data with the forecast data to establish a unified forecast input framework for predicting the wind power for the next N steps. The specific combination process is as follows Figure 2 shown.

[0115] S6, uses ConvsLSTM to predict the input data obtained in S5.

[0116] S7, denormalize the prediction results to obtain the final ultra-short-term wind power prediction results. The denormalization formula is:

[0117] X=(X norm *(X max -X min ))+X min

[0118] Where, X is the power data after denormalization, X norm is the normalized data, X max and X min are the maximum and minimum values ​​of the original power data respectively.

[0119] S8, in order to verify the superiority of the model proposed in the present invention, the LSTM, Transformer, NS-Transformer, and time series Transformer models combined with weather forecast data are compared with the ConvSLSTM model without forecast data to verify the performance in 16 steps, 12 steps, 8 steps, and 4 steps respectively. All experiments were completed on the Intel(R) i7-10750H CPU@2.60GHz, GTX1650Ti, and 16GB RAM experimental platform. The number of iterations of all models is 300 times, the number of input steps is twice the number of prediction steps, the number of neurons is 64, and the batch size is 16. The number of attention heads of the transformer model is 4 and the number of encoding layers is 3. The prediction results and evaluation indicators are as follows: Figure 8 and as shown in Table 1.

[0120] from Figure 8 As can be seen from Table 1, the prediction results of the six prediction models all fluctuate around the actual wind power curve, but the forecast-ConvSLSTM model has the best performance at each prediction step. In the multi-step prediction task, the ConvSLSTM model based on meteorological forecast data shows obvious advantages, with the lowest RMSE, low MAE, and R 2 The value is the highest, which fully demonstrates the significant advantages of the proposed model in processing time series data and multi-step prediction. In the 16-step prediction, the RMSE of the Forecast-ConvSLSTM model is reduced from 13.821 to 10.753, the MAE is reduced from 10.472 to 7.634, and the R 2The accuracies of the predictions for wind power generation were significantly improved from 0.532 to 0.786, demonstrating that the inclusion of weather forecast information plays a crucial role in improving forecast accuracy. Furthermore, the prediction results of the Forecast-ConvSLSTM model significantly improve compared to the No-Forecast-ConvSLSTM model at all other forecast steps, further validating the role of weather forecast data in promoting wind power forecasting. The Forecast-ConvSLSTM model also demonstrates superior forecasting performance compared to the Forecast-LSTM model. This demonstrates that the combined design of the peephole architecture and the novel gating structure not only helps to more accurately extract dynamic features in time series but also better captures spatial features, significantly improving the overall model performance. Based on the above analysis, the ConvSLSTM model, due to its fusion of CNN and LSTM, can effectively capture both spatial features and temporal dependencies in wind power series. Consequently, it provides more accurate forecasts at all forecast steps, demonstrating its strong potential in wind power forecasting tasks.

[0121] Table 1 Evaluation indicators of prediction results of different models

[0122]

[0123]

[0124] In the present invention, by calculating the MIC value between different meteorological factors (such as wind speed, wind direction, temperature, etc.) and wind power, the impact of these factors on wind power can be quantified. Meteorological factors with higher MIC values ​​indicate that they have a stronger correlation with wind power, and therefore can be given priority for use in the wind power prediction model. Conversely, factors with lower MIC values ​​may have less impact on power, and perhaps their weight can be reduced or excluded in the model. In this way, MIC can screen out the most influential meteorological factors, thereby optimizing the prediction model and improving the prediction accuracy and robustness. This process ensures that the wind power prediction model can make full use of the information of key meteorological factors and improve the operating efficiency and prediction ability of the wind farm.

[0125] Preferably, in step S5, weather forecast data for the next N steps is obtained, and M steps of historical data including weather characteristics and wind power are extracted. The historical data and forecast data are then combined to establish a unified prediction input framework for predicting wind power for the next N steps. The specific combination process is as follows: Figure 2 shown.

[0126] However, when MIC alone is used for correlation analysis, wind power data containing long-term trends (seasonality, periodicity, etc.) and short-term fluctuations are easily disturbed by short-term fluctuations, making it difficult to accurately capture the long-term trend correlation. Therefore, in one embodiment, a combination of MIC and GMM clustering is used to start from multiple levels of the macro structure and micro details of the time series to reduce the impact of short-term random noise and grasp the long-term variation pattern of wind power.

[0127] Preferably, the correlation analysis method includes:

[0128] Preprocess historical wind power data and related meteorological data to eliminate noise and missing values ​​in the data and make data with different characteristics comparable;

[0129] S401. Use the maximum information coefficient MIC to calculate the correlation coefficient between each meteorological factor and wind power, preliminarily screen out meteorological factors with a strong correlation with wind power, and obtain a preliminary set of key meteorological factors.

[0130] S402: Integrate the initially screened key meteorological factor data with historical wind power data, cluster and group wind farm units using a Gaussian mixture model, determine the optimal number of unit groups based on the Bayesian Information Criterion, and divide the wind farm into several groups of units with similar output characteristics;

[0131] Different turbines in a wind farm may be affected by different factors or respond differently to the same factors. Grouping turbines makes the characteristics of turbines within the same group more similar, reducing data heterogeneity. This allows for more accurate capture of the relationship between meteorological factors and wind power during correlation analysis and model building, improving the accuracy of analysis and prediction. Some turbines may perform better in high wind speed conditions, while others may perform better in low wind speed conditions. By grouping, the correlation between each group and meteorological factors can be analyzed separately, capturing specific correlation patterns within the group, and thus establishing more accurate prediction models for each turbine group.

[0132] S403. For each turbine group, the MIC value is combined with the characteristic weights determined based on physical models or domain knowledge to calculate the fusion weights of each meteorological factor. For example, based on the physical generation principle of wind power, the influence of wind speed on wind power may have a higher weight. At the same time, the MIC value is referenced and adjusted to obtain the final fusion weights of each meteorological factor, further identifying meteorological factors that are highly correlated with wind power.

[0133] When calculating the correlation between each characteristic factor and wind power, the MIC value is combined with the characteristic weight determined based on the physical model or domain knowledge, and the MIC value is set to MIC ij, represents the correlation coefficient between the i-th characteristic factor and wind power in the j-th time window; the characteristic weight determined according to the physical model or domain knowledge is w i , then the feature correlation weight after fusion R i Expressed as:

[0134]

[0135] Among them, m is the number of feature factors and n is the number of time windows.

[0136] The historical monitoring data and future forecast data are spliced ​​on the time axis; with the current time as the boundary, the historical monitoring data of the past few hours and the weather forecast data of the next 24 hours are arranged and combined in the time dimension in chronological order to form a complete input feature sequence;

[0137] At each time point, data from multiple meteorological factors are combined into a feature vector, which serves as the input feature for the prediction model. This can include meteorological factor values ​​from historical monitoring data and predicted values ​​from future forecast data. Assuming the meteorological factors selected include wind speed, wind direction, air pressure, temperature, and historical wind power values, the dimension of the feature vector at each time point is 5.

[0138] Preferably, step S6 includes the following steps,

[0139] S601, LSTM is a neural network that efficiently processes time series by introducing memory units. Each memory unit contains an input gate, a forget gate, and an output gate. Its main function is to track the amount of information flow to determine whether certain information should be forgotten or updated and retained for a long time. LSTM can effectively avoid the gradient vanishing problem in traditional recurrent neural networks (RNNs), thereby maintaining the memory of important information in longer sequences. The network unit structure of LSTM is as follows Figure 3 As shown, the specific update state and network output formulas are as follows:

[0140] f t =σ(W xf x t +W hf h t-1 +b f )

[0141] i t =σ(W xi x t +W hi h t-1 +b i )

[0142] c t =f t ct-1 +i t tanh(W xc x t +W hc h t-1 +b c )

[0143] o t =σ(W xo x t +W ho h t-1 +b o )

[0144] h t =o t tanh(c t )

[0145] Where Wxf, Wxi, Wxc, and Wxo represent input weights, Whf, Whi, Whc, and Who are recurrent weights, and bi, bf, bc, and bo are bias terms. σ(·) represents the activation function, tanh(·) is the hyperbolic tangent activation function, and ft, it, ct, ot, and ht represent the LSTM's forget gate, input gate, memory cell state, output gate, and network output.

[0146] S602, in order to further enhance the ability of LSTM to process spatiotemporal data, Shi et al. proposed a convolutional long short-term memory network (ConvLSTM), whose network unit structure is as follows Figure 4 As shown in Figure 2, in ConvLSTM, convolution operations are introduced to replace the input-state and state-state conversion processes in LSTM to improve the network's ability to extract features from spatial data. Similar to LSTM, the forward propagation process of ConvLSTM is as follows:

[0147] f t =σ(W xf *x t +W hf *h t-1 +b f )

[0148] i t =σ(W xi *x t +W hi *h t-1 +b i )

[0149] c t =f t c t-1 +i t tanh(W xc *xt +W hc *h t-1 +b c )

[0150] o t =σ(W xo *x t +W ho *h t-1 +b o )

[0151] h t =o t tanh(c t )

[0152] Here, * denotes a convolution operation, ensuring that the dimensions of the input and state tensors match. Before applying the convolution operation, the input and state tensors are typically padded, typically with zeros, to avoid edge effects.

[0153] S603, as described in step S601, ConvLSTM enhances the spatial data processing capability of LSTM by introducing the convolution operator. However, ConvLSTM has more parameters to be optimized and longer training time. In order to reduce training time and improve computational efficiency, a convolutional simplified long short-term memory network (ConvSLSTM) is proposed. ConvSLSTM reduces the number of parameters to be optimized by simplifying the gate structure within the network while retaining the advantages of convolution operations. In addition, ConvSLSTM replaces the input gate, forget gate, and output gate in the original ConvLSTM with a cross-coupling structure, and also introduces a peephole structure to consider the impact of storage units on network output. The structure of ConvSLSTM is as follows: Figure 5 As shown, the forward propagation process is as follows:

[0154] i t =σ(W x *x t +W h *h t-1 +W c *c t-1 +b i )

[0155]

[0156] o t =σ(W x *x t +W h *h t-1 +W c *c t +b o)

[0157]

[0158] In the formula, * is the convolution operation, is the Hadamard product operation, W x is the input weight, W h is the cycle weight, W c is the peephole weight, b i 、b c and b o is the input bias, σ(·) is the activation function, tanh(·) is the hyperbolic activation function, i t 、c t 、o t and h t They represent the input gate, memory cell state, output gate and network output of LSTM respectively.

[0159] Step S7 includes the following steps:

[0160] S701, the formula for denormalizing the prediction results is:

[0161] X=(X norm *(X max -X min ))+X min

[0162] Where, X is the original wind power data, X norm is the normalized data; X max and X min are the maximum and minimum values ​​of the original data respectively.

[0163] S702, the present invention uses root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R2) as evaluation indicators, and the calculation formula is as follows:

[0164]

[0165] Where y i represents the true value, represents the predicted value, n is the number of samples, Represents the mean of the true values.

[0166] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A wind power prediction method based on weather forecast and convolutional simplified LSTM, characterized by: The following steps are involved: S1. Verify the integrity of historical wind power monitoring data by analyzing the discreteness of the historical wind power monitoring data series and identify abnormal data that may affect the accuracy of the model; S2. Repair the detected abnormal data to ensure the continuity and integrity of the input data; S3, normalizing the historical wind power monitoring data after the abnormal data is repaired; S4. Screen out meteorological factors that are highly correlated with wind power among meteorological factors of wind farms through correlation analysis method; S5. Obtain future forecast data of meteorological factors that are highly correlated with wind power and combine them with historical wind power monitoring data as input features for the prediction model to fully utilize historical trends and meteorological forecast information; S6. Based on LSTM, a convolutional layer is used to extract the local spatiotemporal features of the wind power monitoring data sequence. A new gating structure with a peephole structure and cross-coupling is introduced to improve the model's ability to process high-dimensional complex data. At the same time, LSTM units are used to capture the long-term dependencies of time series data, resulting in ConvSLSTM. S7. Use the obtained ConvSLSTM to predict wind power, denormalize the prediction results, and calculate the evaluation index.

2. The wind power prediction method based on weather forecast and convolutional simplified LSTM according to claim 1 is characterized by: Step S1 also includes the following steps: S11. Use a box plot to detect abnormal data. Arrange the historical wind power monitoring data from small to large and divide them into four parts. The three division points correspond to quartiles. The second quartile Q2 represents the median of the data. Its calculation formula is: The first and third quantiles correspond to the values ​​in the data and The value of the position, for different wind power data amounts m, is calculated as follows: When m = 4k + 3 (k = 0, 1, ...), the calculation formulas for the first and third quantiles are: When m = 4k + 1 (k = 0, 1, ...), the calculation formulas for the first and third quantiles are: Based on the above calculation, the interquartile range I is obtained. i : I i =Q3-Q1; By setting the upper and lower limits W u,i 、W d,i Used to identify outliers, the formula is: <h2 style=";text-align:left;direction:ltr">W<h2 style=";text-align:left;direction:ltr"> u,i <h2 style=";text-align:left;direction:ltr"> =Q3+1.5I<h2 style=";text-align:left;direction:ltr"> i W d,i =Q1-1.5I i ; Defined in the range [W u,i ,W d,i ] are normal data, and those outside this range are considered abnormal data.

3. The wind power prediction method based on weather forecast and convolutional simplified LSTM according to claim 1 is characterized by: Step S2 also includes the following steps: Since the intermittent missing data are redundant and the data at adjacent moments will not mutate, the K nearest neighbor complementary method is used to reconstruct the intermittent missing data: For a given wind power output data, the algorithm obtains the average value of the K nearest neighbor data near the missing data and replaces the missing data with the average value to repair the intermittent missing data. The formula is described as: Where, X i is the missing data location of the data sample, x i-K is the Kth data before the missing data, x i+K It is the Kth data after the missing data.

4. The wind power prediction method based on weather forecast and convolutional simplified LSTM according to claim 1 is characterized by: The correlation analysis method in step S4 includes the following steps: S41. By placing the two sets of dominant component error sequences in the same coordinate system to form a scatter plot, and then gridding the coordinate system, the mutual information value is calculated based on the marginal probability and joint probability of the variable scatter points falling in the grid. Finally, the sum of the mutual information values ​​of all points is obtained as a measure of the correlation between the two sets of dominant component errors. Given two sets of dominant component error sequences, the mutual information I(X,Y) is defined as: Where p(x,y) is the joint probability of variables x and y, p(x) and p(y) are the marginal probabilities of variables x and y respectively; S42. Place the given finite set in the same coordinate system and divide it into a grid G. By changing the number of columns a and the number of rows b of the grid, various grids can be generated, thereby obtaining different mutual information values. The maximum mutual information value is defined as: S43. Normalize the maximum mutual information values ​​on different grids G to obtain the feature matrix: On this basis, the maximum information coefficient of the error data set D is defined as:

5. The wind power prediction method based on weather forecast and convolutional simplified LSTM according to claim 1 is characterized by: The correlation analysis method in step S4 includes: S401, performing a preprocessing operation on historical wind power monitoring data and related meteorological data to eliminate noise and missing values ​​in the historical wind power monitoring data and related meteorological data; S402, using the maximum information coefficient (MIC) to calculate the correlation coefficient between each meteorological factor and wind power, preliminarily screening out meteorological factors with a strong correlation with wind power, and obtaining a preliminary set of key meteorological factor data; S403: Integrate the initially screened key meteorological factor data with historical wind power monitoring data, cluster and group wind farm units using a Gaussian mixture model, determine the optimal number of unit groups based on the Bayesian Information Criterion, and divide the wind farm into several groups of units with similar output characteristics; S404: For each turbine group, the MIC value is combined with the characteristic weight determined based on the physical model or domain knowledge to calculate the fusion weight of each meteorological factor, and further determine the meteorological factors that are highly correlated with wind power.

6. The wind power prediction method based on weather forecast and convolutional simplified LSTM according to claim 5 is characterized by: Step S4 04 Specifically includes: combining the MIC value with the feature weight determined based on the physical model or domain knowledge, and setting the MIC value to MIC ij , represents the correlation coefficient between the i-th characteristic factor and wind power in the j-th time window; the characteristic weight determined according to the physical model or domain knowledge is w i , then the feature correlation weight after fusion R i Expressed as: Among them, m is the number of feature factors and n is the number of time windows.

7. The wind power prediction method based on weather forecast and convolutional simplified LSTM according to claim 1 is characterized by: Step S5 includes: Obtain weather forecast data for the next N steps, and simultaneously extract M steps of historical wind power monitoring data containing meteorological characteristics and wind power. Then, combine the historical wind power monitoring data with the forecast data to establish a unified forecast input framework for predicting wind power for the next N steps.

8. The wind power prediction method based on weather forecast and convolutional simplified LSTM according to claim 1 is characterized by: Step S5 further includes: The historical wind power monitoring data and future forecast data are spliced ​​on the time axis; with the current time as the boundary, the historical wind power monitoring data of the past several hours and the weather forecast data of the next several hours are arranged and combined in the time dimension in chronological order to form a complete input feature sequence; At each time point, several meteorological factor data are combined into a feature vector as the input feature of the prediction model. The number of selected meteorological factors is equal to the dimension of the feature vector at each time point.

9. The wind power prediction method based on weather forecast and convolutional simplified LSTM according to claim 1 is characterized by: Step S6 In [1], the construction process of ConvSLSTM includes: To further reduce the number of LSTM parameters to be optimized and its training time, a new gating structure with a cross-coupling structure is used to replace the input gate, forget gate, and output gate in the LSTM network. The introduction of convolution operations improves the ability to process spatiotemporal data and reduces model training time by sharing parameters of various gate structures; A peephole structure is introduced to consider the impact of storage units on network output; The forward propagation process of ConvSLSTM is shown as follows: In the formula, * is the convolution operation, is the Hadamard product operation, W x is the input weight, W h is the cycle weight, W c is the peephole weight, b i 、b c and b o is the input bias, σ(·) is the activation function, tanh(·) is the hyperbolic activation function, i t 、c t 、o t and h t Represent the input gate, memory cell state, output gate and network output of LSTM respectively, h t-1 Indicates the network output at the previous time, c t-1 Indicates the state of the memory unit at the previous time, x t Represents the input of the network.

10. The wind power prediction method based on weather forecast and convolutional simplified LSTM according to claim 1 is characterized by: Step S7 includes the following steps: S701. Denormalize the prediction result. The formula is: X=(X norm *(X max -X min ))+X min ; Where, X is the original wind power data, X norm is the normalized data; X max and X min are the maximum and minimum values ​​of the original data respectively; S702: Root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) are used as evaluation indicators. The calculation formulas are: Where y i represents the true value, represents the predicted value, n is the number of samples, Represents the mean of the true values.

Images