Long-term wind power prediction method, device and equipment based on improved Informer and storage medium

Abstract

The invention relates to the technical field of wind power prediction, in particular to a long-term wind power prediction method, device and equipment based on improved Informer and a storage medium, and the method comprises the steps: processing lacking wind power data through employing a nearest neighbor interpolation method and a similar day substitution method, carrying out the correlation analysis of the wind power data, selecting out features which have obvious influences on the wind power, and carrying out the prediction of the wind power. Finally, making the data into a training set and a test set; establishing a wind power prediction model of the wind power plant, and combining a CNN model with an Informer model; selecting a proper loss function and a solver, and training the built wind power prediction model of the wind power plant; and inputting the historical data corresponding to the to-be-predicted time period into the trained wind power prediction model of the wind power plant, and outputting the corresponding predicted wind power of the wind power plant. According to the method, the prediction accuracy is improved, the number of models and the training time are reduced, the convergence speed of the models is increased, and the method has extremely high engineering application value and practical significance.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and discloses a long-term wind power prediction method, device, equipment and storage medium based on an improved Informer. Background Art

[0002] As traditional fossil energy sources are becoming increasingly depleted, the grid connection of renewable energy power generation will be one of the basic characteristics of my country's future power system. Wind energy, as the main force of renewable energy power generation, will also usher in a new period of growth.

[0003] Wind power faces challenges such as volatility and intermittency. Accurate long-term forecasts of wind energy directly impact operational decisions within the power system. However, traditional forecasting methods are ineffective for long-term series. Filling missing values ​​often involves large amounts of continuous missing data within a vast dataset, which doesn't adequately reflect the original data's characteristics. Furthermore, it's difficult to determine whether wind power, wind speed, and wind direction follow a specific distribution pattern.

[0004] Therefore, in order to formulate efficient scheduling plans and control strategies for the power system and reduce the difficulty of wind power grid integration, a long-term wind power prediction method is urgently needed to solve the above technical problems. Summary of the Invention

[0005] The present invention provides a long-term wind power forecasting method, device, equipment and storage medium based on an improved Informer, which overcomes the shortcomings of the above-mentioned existing technologies and can effectively solve the problem that traditional forecasting methods are not ideal for long time series.

[0006] One of the technical solutions of the present invention is achieved by the following measures: a long-term wind power prediction method based on an improved Informer, comprising the following steps:

[0007] Step S1: Process the missing wind power data, perform correlation analysis on the wind power data, select the features that have a significant impact on wind power, and make the data into training sets and test sets;

[0008] Step S2: Establish a wind power prediction model for a wind farm by combining the CNN and Informer models;

[0009] Step S3: Select a suitable loss function and solver to train the wind power prediction model of the wind farm established in step S2;

[0010] Step S4: input the historical data corresponding to the time period to be predicted into the trained wind power prediction model of the wind farm, and output the corresponding predicted wind power of the wind farm.

[0011] The following is a further optimization and / or improvement of one of the above-mentioned technical solutions:

[0012] The above step S1 may specifically include the following steps:

[0013] Step S1.1: Use the nearest neighbor average interpolation method and the similar day replacement method to process the missing wind power data to ensure the continuity of the wind power data;

[0014] Step S1.2: Based on the wind power data processed in step S1.1, use the Spearman correlation coefficient to analyze the factors affecting wind power output and select high-quality features with high correlation.

[0015] The above step S2 may specifically include the following steps:

[0016] Step S2.1: Use the Kalman filter to smooth the optimized data features. By fusing the observed variable with the state variable to obtain approximate true data, the observed variable is then used as the next state variable for further fusion, and this process is repeated, so that the abnormal data is theoretically closer to the true value.

[0017] The prediction process of the Kalman filter is calculated as follows:

[0018]

[0019]

[0020] in, is the prior estimate of time k; A is the state transition matrix; w k-1 is a Gaussian white noise with a covariance matrix Q; is the prior estimated covariance at time k; P k-1 is the posterior estimated covariance at time k-1.

[0021] The update process of the Kalman filter is calculated as follows:

[0022]

[0023]

[0024]

[0025] Among them, K k is the Kalman gain; H is the observation matrix; R is the range of the predicted target value; H T is the transpose of H; z k is the observed value at time k;

[0026] Step S2.2: constructing a wind power prediction model for a wind farm, wherein the wind power prediction model for a wind farm includes a CNN feature extraction sub-model and an Informer prediction sub-model;

[0027] Among them, the CNN feature extraction sub-model is used to capture local features to more comprehensively obtain the change pattern of wind power data;

[0028] Among them, the Informer prediction sub-model is used to improve the computational efficiency of the self-attention mechanism, multi-layer network stacking and step-by-step decoding method; the Informer prediction sub-model includes an encoder and a decoder. The encoder receives a large amount of long sequence input and uses a multi-head sparse self-attention mechanism instead of the conventional self-attention mechanism. The distillation operation is used to extract the main attention to reduce the network size. The decoder accepts long sequence input and the target elements are filled with 0 values.

[0029] The above step S3 may specifically include the following steps:

[0030] Step S3.1: Use MSE as the loss function. The MSE calculation formula is as follows:

[0031]

[0032] Where n is the total number of samples; y i is the true value, is the predicted value output by the neural network;

[0033] Step S3.2: Using the Adam solver, which has high computational efficiency, requires little memory, and is suitable for optimizing a large number of parameters;

[0034] Step S3.3: Complete model training based on pytorch and save the trained model parameters.

[0035] The above step S4 may specifically include the following steps:

[0036] Step S4.1: Read the model parameters trained in step S3.3;

[0037] Step S4.2: Input the historical data corresponding to the time period to be predicted into the wind power prediction model of the wind farm, and use the neural network to calculate the predicted value of the wind power of the wind farm.

[0038] Residual connections can be introduced in the encoder of the above Informer model.

[0039] The second technical solution of the present invention is achieved by the following measures: a long-term wind power prediction device based on an improved informer, comprising:

[0040] The data preprocessing module is used to process missing wind power data, perform correlation analysis on the wind power data, select features that significantly affect wind power, and create training and test sets based on the data.

[0041] The model building module is used to build a wind power prediction model for wind farms, combining the CNN and Informer models;

[0042] The model training module is used to select appropriate loss functions and solvers to train the wind power prediction model of the wind farm;

[0043] The input and output module is used to input the historical data corresponding to the forecast period into the trained wind power forecast model of the wind farm and output the corresponding forecasted wind power of the wind farm.

[0044] The third technical solution of the present invention is achieved through the following measures: a long-term wind power prediction device based on an improved informer, including a memory and a processor, the memory stores a program that can be run on the processor, and when the processor executes the program, the above-mentioned long-term wind power prediction method based on the improved informer is implemented.

[0045] The fourth technical solution of the present invention is achieved through the following measures: a storage medium, the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the above-mentioned long-term wind power prediction method based on the improved Informer.

[0046] The present invention preprocesses the relevant data of the wind power plant, including selecting high-quality features through correlation analysis, processing missing data by nearest neighbor mean and similar day filling method, and forming a feature library and training samples. The present invention constructs a wind power prediction model for the wind power plant, and uses a Kalman filter to smooth the data features after optimization so that they can be input into the CNN and Informer models. The model adopts a codec structure. The difference from the traditional Informer model is that, in order to accelerate the convergence of the model, the present invention introduces a residual connection in the encoder part of the Informer model so that the model can better obtain the information of the input data; compared with the conventional long short-term memory neural network, the accuracy of the prediction is significantly improved. The present invention takes into account multiple features including meteorological data and wind power data, and learns the correlation of the output of each feature through Informer, which effectively improves the accuracy of wind power generation prediction of the wind power plant. Compared with traditional wind power prediction methods, the present invention directly predicts the wind power of the wind power plant, improves the accuracy of the prediction, reduces the number of models and training time, and has extremely high engineering application value and practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a long-term wind power prediction method based on an improved Informer according to Example 1 of the present invention.

[0048] Figure 2 This is a graph showing the correlation analysis results between wind power and various characteristics according to Example 3 of the present invention.

[0049] Figure 3 This is a structural framework diagram of the CNN and Informer models of Example 3 of the present invention.

[0050] Figure 4 This is a diagram showing the effect of long-term power prediction of a wind power plant based on the improved Informer according to Example 3 of the present invention.

[0051] Figure 5 This is a comparison diagram of the convergence effect of using residual connection before and after the prediction step size is 24 in Example 3 of the present invention.

[0052] Figure 6 This is a comparison diagram of the convergence effect of using residual connection before and after the prediction step size is 48 in Example 3 of the present invention.

[0053] Figure 7 This is a comparison diagram of the convergence effect of using residual connection before and after the prediction step size is 96 in Example 3 of the present invention.

[0054] Figure 8 This is a comparison chart of the prediction effects of the improved Informer and LSTM methods in Example 3 of the present invention. DETAILED DESCRIPTION

[0055] The present invention is not limited to the following embodiments, and specific implementation methods can be determined based on the technical solutions of the present invention and actual conditions.

[0056] The present invention will be further described below in conjunction with the embodiments:

[0057] Example 1: As shown in the attached Figure 1 As shown, the long-term wind power prediction method based on the improved Informer includes the following steps:

[0058] Step S1: Use the nearest neighbor average interpolation method and the similar day substitution method to process the missing wind power data, perform correlation analysis on the wind power data, select the features that have a significant impact on wind power, and finally make the data into training sets and test sets.

[0059] Step S2: Establish a wind farm wind power prediction model, combining the CNN and Informer models. In this embodiment of the present invention, the long-term wind power prediction method based on the improved Informer is a wind farm wind power prediction method based on the CNN and Informer models. In the improved Informer model, the CNN and Informer models are combined, and a residual connection is introduced in the encoder of the Informer model to accelerate the convergence of the model.

[0060] Step S3: Select a suitable loss function and solver to train the wind power prediction model of the wind farm established in step S2.

[0061] Step S4: input the historical data corresponding to the time period to be predicted into the trained wind power prediction model of the wind farm, and output the corresponding predicted wind power of the wind farm.

[0062] In the embodiment of the present invention, step S1 specifically includes the following steps:

[0063] Step S1.1: Use the nearest neighbor average interpolation method and the similar day replacement method to process the missing wind power data to ensure the continuity of the wind power data;

[0064] Step S1.2: Based on the wind power data processed in step S1.1, use the Spearman correlation coefficient to analyze the factors affecting wind power output and select high-quality features with high correlation.

[0065] In this embodiment of the present invention, step S2 specifically includes the following steps:

[0066] Step S2.1: Use a Kalman filter to smooth the optimized data features. This process fuses the observed variables with the state variables to obtain approximate true data, which is then used as the next state variable for further fusion. This process repeats, making the abnormal data theoretically closer to the true value for input into the wind power prediction model of the wind farm.

[0067] In an embodiment of the present invention, the prediction process of the Kalman filter is calculated by the following formula:

[0068]

[0069]

[0070] in, is the prior estimate of time k; A is the state transition matrix; w k-1 is a Gaussian white noise with a covariance matrix Q; is the prior estimated covariance at time k; P k-1is the posterior estimated covariance at time k-1.

[0071] In an embodiment of the present invention, the update process of the Kalman filter is calculated by the following formula:

[0072]

[0073]

[0074]

[0075] Among them, K k is the Kalman gain; H is the observation matrix; R is the range of the predicted target value; H T is the transpose of H; z k is the observed value at time k.

[0076] Step S2.2: constructing a wind power prediction model for a wind farm, wherein the wind power prediction model for a wind farm includes a CNN feature extraction sub-model and an Informer prediction sub-model;

[0077] Among them, the CNN feature extraction sub-model is used to capture local features to more comprehensively obtain the change pattern of wind power data;

[0078] Among them, the Informer prediction sub-model is a network structure model based on the attention mechanism, which is used to improve the computational efficiency of the self-attention mechanism, multi-layer network stacking and step-by-step decoding method; the Informer prediction sub-model includes an encoder and a decoder. The encoder receives a large amount of long sequence input and uses a multi-head sparse self-attention mechanism instead of the conventional self-attention mechanism. The distillation operation is used to extract the main attention to reduce the network size. The decoder accepts long sequence input and the target elements are filled with 0 values.

[0079] The encoder consists of a stack of multi-head sparse self-attention and distillation. The multi-head sparse self-attention mechanism is a variant of the self-attention mechanism. It is designed to capture dependencies between features. Its advantage lies in reducing the computational complexity of each attention layer. The multi-head sparse self-attention mechanism achieves O(LlogL) time complexity and memory usage, and has comparable performance for sequence dependency alignment.

[0080] Multi-head sparse self-attention is composed of several stacked sparse self-attention layers. Generally, the encoder contains multiple multi-head sparse self-attention layers. A distillation operation is performed when the input data flows from the i-th multi-head sparse self-attention layer to the i+1-th multi-head sparse self-attention layer. The purpose of distillation is to reduce model complexity and computational complexity by assigning high weights to key features, simplifying the data according to the following formula, and accelerating computational speed. After encoding the input sequence, the encoder generates a concatenated feature map and sends it to the multi-head self-attention layer of the decoder.

[0081]

[0082] where [·] is the attention module in multi-head sparse self-attention, Conv1d represents the regularized convolution operation on the time series, and ELU is the activation function.

[0083] The decoder consists of a masked multi-head sparse self-attention block and a multi-head attention block. Its attention mechanism is designed to prevent each position from focusing on future information, thus avoiding autoregression. The masked multi-head sparse self-attention block receives the label vector of equation (4), and the decoder uses a generative approach for decoding. The decoder uses the intermediate output of the encoder to perform a multi-head attention operation, adjusts the dimension of the output data through a fully connected layer, and finally outputs the predicted sequence length in one go.

[0084]

[0085] in is the t-th input sequence of the decoder, The start marker of the tth sequence is the 0 value combination mentioned above. is a placeholder for the target sequence of the t-th sequence.

[0086] The encoder and decoder of the Informer model receive the output of the convolutional module. The encoder and decoder receive different data. After entering the encoder, the data undergoes multiple operations in the multi-head probabilistic sparse self-attention module and the "distillation" mechanism, outputting an intermediate result. The decoder's input data is first masked and subjected to the multi-head probabilistic sparse self-attention operation. Then, the decoder's intermediate result is subjected to the multi-head self-attention operation. Finally, a fully connected layer adjusts the data output dimension to obtain the predicted result. Backward gradient propagation is performed on the output prediction result to continuously optimize the model. To address the network degradation caused by stacking multiple layers in the Informer model, residual connections are introduced in the encoder of the Informer model to enable the Informer model to better capture information about the input data and accelerate model convergence. The entire residual block consists of two convolutional pooling layers and three attention mechanism blocks. Therefore, the encoder output not only includes the output information of the last layer but also incorporates the output information that has undergone nonlinear transformations. This ensures that the output layer of the model can obtain as much information as possible from the input layer, accelerating model convergence.

[0087] In this embodiment of the present invention, step S3 specifically includes the following steps:

[0088] Step S3.1: Use MSE as the loss function. The MSE calculation formula is as follows:

[0089]

[0090] Where n is the total number of samples; y i is the true value, is the predicted value output by the neural network;

[0091] Step S3.2: Using the Adam solver, which has high computational efficiency, requires little memory, and is suitable for optimizing a large number of parameters;

[0092] Step S3.3: Complete model training based on pytorch and save the trained model parameters.

[0093] In this embodiment of the present invention, step S4 specifically includes the following steps:

[0094] Step S4.1: Read the model parameters trained in step S3.3;

[0095] Step S4.2: Input the historical data corresponding to the time period to be predicted into the wind power prediction model of the wind farm, and use the neural network to calculate the predicted value of the wind power of the wind farm.

[0096] This embodiment of the present invention studies the filling of missing data in wind power series and the identification and correction of abnormal data. It uses the nearest neighbor mean and similar day filling methods to fill missing wind power data. First, the K nearest valid values ​​before and after the missing value are found. These K values ​​are then summed and averaged to serve as the target data for the missing value. This completes the missing value filling process. While this method can fill missing values, it does not effectively reflect the original data features when large amounts of continuous missing data are present in large datasets. Therefore, the similar day filling method is employed to address this large number of continuous missing values, thus completing the missing value filling process. The Spearman correlation coefficient method is then used to perform quantitative correlation analysis. Since wind power, wind speed, and wind direction are difficult to determine to follow a specific distribution pattern, the Spearman correlation coefficient method is suitable for quantitative correlation analysis. This quantitative correlation analysis aims to identify the most relevant features, effectively preventing the impact of redundant data on the model and providing relatively accurate data for the subsequent construction and training of the prediction model.

[0097] The embodiment of the present invention pre-processes the relevant data of the wind power plant, including correlation analysis to select high-quality features, nearest neighbor mean and similar day filling method to process missing data, and form a feature library and training samples. The embodiment of the present invention constructs a wind power prediction model for a wind power plant. First, the Kalman filter is used to smooth the data features after optimization so that they can be input into the CNN and Informer models. The model adopts an encoder-decoder structure. The difference from the traditional Informer model is that, in order to speed up the convergence of the model, the embodiment of the present invention introduces a residual connection in the encoder part of the Informer model so that the model can better obtain the information of the input data; compared with the conventional long short-term memory (LSTM) neural network, the prediction accuracy is significantly improved. The embodiment of the present invention takes into account a variety of features including meteorological data and wind power data, and learns the correlation between the output of each feature through Informer, which effectively improves the accuracy of wind power generation prediction of the wind power plant. To demonstrate its long-term forecasting capabilities, this embodiment of the present invention uses an LSTM method to perform long-term wind power forecasts. The prediction accuracy is compared with that of the embodiment of the present invention, confirming its superiority. Compared to traditional wind power forecasting methods, this embodiment of the present invention directly forecasts wind power for wind farms, improving forecast accuracy while reducing the number of models and training time. This approach has significant engineering application value and practical significance.

[0098] Example 2: The embodiment of the present invention discloses a long-term wind power prediction device based on an improved informer, comprising:

[0099] The data preprocessing module handles missing wind power data and performs correlation analysis on it, identifying features that significantly influence wind power and creating training and test sets. Nearest neighbor average interpolation and similar day substitution are used to process missing wind power data to ensure data continuity. After processing, the Spearman correlation coefficient is used to analyze factors influencing wind power output and select high-quality features with high correlation.

[0100] The model building module is used to establish a wind power prediction model for a wind farm, combining the CNN and Informer models. When constructing the wind power prediction model, the model includes a CNN feature extraction sub-model and an Informer prediction sub-model. The CNN feature extraction sub-model is used to capture local features to more comprehensively capture the changing patterns of wind power data. The Informer prediction sub-model is a network structure model based on the attention mechanism, which is used to improve the computational efficiency of the self-attention mechanism, multi-layer network stacking, and step-by-step decoding methods. The Informer prediction sub-model includes an encoder and a decoder. The encoder receives a large number of long sequence inputs and uses a multi-head sparse self-attention mechanism instead of a conventional self-attention mechanism. A distillation operation is used to extract the main attention and reduce the network size. The decoder accepts long sequence inputs and fills the target elements with zero values. Residual connections can be introduced in the encoder of the Informer model to accelerate model convergence.

[0101] The model training module selects an appropriate loss function and solver to train the wind power prediction model for wind farms. It uses the Adam solver, which is computationally efficient, requires little memory, and is suitable for optimizing a large number of parameters. It trains the model based on PyTorch and saves the trained model parameters.

[0102] The input / output module is used to input the historical data corresponding to the forecast period into the trained wind farm power forecast model and output the corresponding predicted wind farm power. The trained model parameters are first read, and then the historical data corresponding to the forecast period is input into the wind farm power forecast model. The neural network then calculates the predicted wind farm power value.

[0103] This embodiment of the present invention constructs a wind power prediction model for a wind farm. Using a Kalman filter, it smoothes the optimized data features for input into CNN and informer models. This model utilizes a codec structure. Unlike traditional informer models, this embodiment introduces residual connections in the encoder portion of the informer model to accelerate model convergence, enabling the model to better capture input data information. This embodiment improves prediction accuracy, reduces the number of models, and reduces training time, resulting in high engineering application value and practical significance.

[0104] Example 3: The present invention discloses a specific example of the present invention to further illustrate the present invention. Figures 2 to 8 As shown, an embodiment of the present invention uses the improved informer-based long-term wind power prediction method of the present invention to perform long-term power prediction on a wind power data set provided by a wind farm in southern Australia, including the following steps:

[0105] Step S1: Use the nearest neighbor average interpolation method and the similar day replacement method to process the missing wind power data to ensure the continuity of the wind power data;

[0106] In actual measurement systems, measurement data collected will inevitably contain missing items. Therefore, in order to maintain the continuity of wind power data and facilitate the generation of data sets, the data needs to be interpolated. The working process of the nearest neighbor mean interpolation method using the nearest neighbor mean interpolation method and the similar day replacement method to handle missing items is as follows:

[0107] Step S1.1: Get the original dataset containing missing values;

[0108] Step S1.2: Query the missing values ​​in the data set and obtain the m valid values ​​X1, X2, X3, ..., Xm that are closest to the missing value.

[0109] Step S1.3: Perform a sum operation on the m valid values ​​according to the following formula, i.e., Sum;

[0110]

[0111] Step S1.4: Then calculate the mean M = Sum / m and fill the missing value position with the calculated M;

[0112] Step S1.5: For multiple missing value problems in the dataset, repeat steps S1.2 to S1.4 until all missing values ​​in the dataset are filled.

[0113] Compared with the nearest neighbor mean missing value filling method, the data filling method using the similar day missing value filling method can more effectively reflect the original characteristic information of the data, and is particularly suitable for data sets with a large number of consecutive missing values.

[0114] The output of a wind farm is related to many factors. The embodiment of the present invention uses the Spearman correlation coefficient to analyze the correlation between wind power and many factors. The calculation formula of the Spearman correlation coefficient is:

[0115]

[0116] R in the above formula i and S i Respectively represent the rank of the observation value i after sorting vectors x and y; and Represents the rank of vectors x and y respectively; N represents the number of observations; d i =R i -S i , which means R i and S i The difference between the i-th observation series of two variables after rearranging them in descending order. ρ s It represents the correlation between two variables, and its value range is a real number between [-1,1].

[0117] The dataset used in the embodiment of the present invention is the weather and power data of the second phase of the Hornsdale wind farm in South Australia. The hourly weather data is provided by DarkSkyAPI in his local time zone and converted into UTC format for subsequent processing. The power data is collected from the API provided by the Australian Renewable Energy Mapping Infrastructure Project. The time span of the entire dataset is from June 30, 2018 to April 30, 2020, a total of 20 months. The dataset includes features such as wind speed, wind direction, gust level, wind power, relative humidity, air pressure, temperature, precipitation intensity, etc. The sampling time interval is 1h, and the total data volume is 15,384 data. The correlation coefficient is analyzed to analyze the relevant factors affecting wind power and obtain a heat map, such as Figure 2 shown.

[0118] Depend on Figure 2 The three factors most strongly correlated with wind power are wind speed, wind direction, and gust, with Spearman correlation coefficients of 0.848, 0.837, and 0.753, respectively, further confirming the significant impact of wind speed on power generation. At the same time, some meteorological factors have a relatively small impact on power generation. For example, the correlation coefficient between ozone and power generation is 0.002, indicating the weakest relationship between the two. Visibility is second, with a correlation coefficient of 0.031, indicating a weak correlation.

[0119] Step S2: Using a Kalman filter, the optimized data features are smoothed to establish a wind power prediction model for the wind farm;

[0120] In the embodiment of the present invention, the CNN and Informer model structures are as follows: Figure 3 As shown in Table 1.

[0121] Table 1 Network structure and parameters of CNN and Informer models

[0122] parameter Parameter value Convolution kernel size 3 The number of multi-head attention heads 8 Encoder input step size 96 Decoder input step size 48 Prediction step length {24，48，96，192} Number of encoder layers 3、2、1 Number of decoder layers 2 Batch size 32 Activation Function gelu Loss Function mse The learning rate of the model 0.0001 discard rate 0.05

[0123] Step S3: Select a suitable loss function and solver to train the wind power prediction model of the wind farm established in step S2;

[0124] Step S4: input the historical data corresponding to the time period to be predicted into the trained wind power prediction model of the wind farm, and output the corresponding predicted wind power of the wind farm.

[0125] As an example, in this embodiment of the present invention, a dataset consisting of 640 days of historical wind farm data was used. During model training, the training set to test set ratio was 7:3, and long-term wind power forecasting was performed with a 1-hour time step. The predicted values ​​for 96 future time points were compared with the actual power curve, and two evaluation metrics, mean absolute error (MAE) and mean square error (MSE), were calculated. The following are detailed experimental results:

[0126] (1) Informer effect analysis

[0127] In order to verify the validity of the model proposed in this invention, verify whether it can be applied to the long-term forecast of wind power, and verify whether its prediction results can well conform to the future trend of wind power, the predicted values ​​of the next 170 time points are compared with the actual power curve, as shown in Figure 2. Figure 4 As shown in the figure, by comparing the curves, we can see that when we predict the wind power of a specified time period of nearly a week using the dataset obtained from a real wind turbine, the wind turbine power curve of the model proposed by the present invention is close to the actual wind turbine power curve. It can also be seen that the error between the predicted value of this model and the actual value is not large.

[0128] (2) Analysis of model convergence effect

[0129] Due to the introduction of the residual connection module in the Informer encoder part, the improved Informer model has a certain improvement in convergence speed compared with the Informer model. Under the same input and environment configuration, Figures 5 to 7The results compare the convergence speed of the Improved Informer and Informer models for decoder prediction step sizes of {24, 48, 96}. Observing the changes in the graph, we can see that in the early stages of training, the loss values ​​of both models are high and unstable. This is because the model network requires time to gradually adjust weights and optimize prediction results. As the number of training times increases, the loss value gradually stabilizes. We also find that the Improved Informer model converges significantly faster than the Informer model at different prediction step sizes, and the convergence difference between the two models increases with increasing prediction step size.

[0130] (3) Analysis of model prediction effect

[0131] In order to verify the performance and comparative advantages of the prediction model proposed in this paper, this section also constructs the LSTM neural network, CNN, Transformer and Informer commonly used in traditional time series prediction and compares them with them. Using the same input features, the same loss function, learning rate and number of training times, the predicted values ​​​​at 96 time points in the future are predicted and compared with the true value power curve, as shown in the figure below. Figure 8 The comparison of prediction errors of each neural network is shown in Table 2.

[0132] Table 2 Comparison of the effects based on the improved Informer model

[0133]

[0134] As can be seen from Table 2, among all the compared methods, the model of the embodiment of the present invention achieved the best experimental results. This shows that the improved Informer model has significant advantages in extracting time series features for long-term prediction. The improved Informer model extracts data features that are closely related to wind power and reduces network degradation by introducing convolutional neural networks and residual connection modules. Because the Informer model mainly focuses on global features and is not good at capturing local features, a convolutional neural network is introduced to better capture local features in the input data set, so that the model can more fully include input information and features, and perform better in the long-term prediction process. Therefore, the present invention is better than LSTM, CNN, Transformer and Informer in both mean absolute error and mean square error, significantly improving the prediction accuracy, which proves the superiority of the present invention.

[0135] Example 4: An embodiment of the present invention provides a long-term wind power prediction device based on an improved Informer, which includes a memory, a processor, a communication interface, and a communication bus. The memory stores a program that can be run on the processor. When the processor executes the program, the long-term wind power prediction method based on the improved Informer in the above embodiment is implemented.

[0136] The processor may be a central processing unit, or may be other general-purpose processors, digital signal processors, application-specific integrated circuits, field programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components or other chips, or a combination of the above chips.

[0137] The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and units, such as the corresponding program units in the above-described method embodiments of the present invention. The processor executes the non-transitory software programs, instructions, and modules stored in the memory to perform various processor functions and process data, thereby implementing the long-term wind power forecasting method based on the improved informer in the above-described embodiment.

[0138] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. The memory may optionally include a memory remotely located relative to the processor, and these remote memories may be connected to the processor via a network. One or more programs are stored in the memory, and when executed by the processor, the long-term wind power prediction method based on the improved Informer in the above embodiment is executed.

[0139] Example 5: An embodiment of the present invention provides a storage medium for computer-readable storage, wherein the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the long-term wind power prediction method based on the improved Informer as provided in the above method embodiment.

[0140] The storage medium may be an internal storage unit of the device, such as the device's hard drive or memory. The storage medium may also be an external storage device, such as a plug-in hard drive, smart memory card, secure digital card, flash memory card, etc.

[0141] The above technical features constitute the embodiments of the present invention, which have strong adaptability and implementation effect. Non-essential technical features can be added or removed according to actual needs to meet the requirements of different situations.

Claims

1. A long-term wind power forecasting method based on improved Informer, characterized by The following steps are involved: Step S1: Process the missing wind power data, perform correlation analysis on the wind power data, select the features that have a significant impact on wind power, and make the data into training sets and test sets; Step S2: Establish a wind power prediction model for a wind farm by combining the CNN and Informer models; Step S3: Select a suitable loss function and solver to train the wind power prediction model of the wind farm established in step S2; Step S4: input the historical data corresponding to the time period to be predicted into the trained wind power prediction model of the wind farm, and output the corresponding predicted wind power of the wind farm.

2. The long-term wind power forecasting method based on the improved Informer according to claim 1 is characterized in that Step S1 specifically includes the following steps: Step S1.1: Use the nearest neighbor average interpolation method and the similar day replacement method to process the missing wind power data to ensure the continuity of the wind power data; Step S1.2: Based on the wind power data processed in step S1.1, use the Spearman correlation coefficient to analyze the factors affecting wind power output and select high-quality features with high correlation.

3. The long-term wind power forecasting method based on the improved Informer according to claim 1 or 2 is characterized in that Step S2 specifically includes the following steps: Step S2.1: Use the Kalman filter to smooth the optimized data features. By fusing the observed variable with the state variable to obtain approximate true data, the observed variable is then used as the next state variable for further fusion, and this process is repeated, so that the abnormal data is theoretically closer to the true value. The prediction process of the Kalman filter is calculated as follows: in, is the prior estimate of time k; A is the state transition matrix; w k-1 is a Gaussian white noise with a covariance matrix Q; is the prior estimated covariance at time k; P k-1 is the posterior estimated covariance at time k-1. The update process of the Kalman filter is calculated as follows: Among them, K k is the Kalman gain; H is the observation matrix; R is the range of the predicted target value; H T is the transpose of H; z k is the observed value at time k; Step S2.2: constructing a wind power prediction model for a wind farm, wherein the wind power prediction model for a wind farm includes a CNN feature extraction sub-model and an Informer prediction sub-model; Among them, the CNN feature extraction sub-model is used to capture local features to more comprehensively obtain the change pattern of wind power data; Among them, the Informer prediction sub-model is used to improve the computational efficiency of the self-attention mechanism, multi-layer network stacking and step-by-step decoding method; the Informer prediction sub-model includes an encoder and a decoder. The encoder receives a large amount of long sequence input and uses a multi-head sparse self-attention mechanism instead of the conventional self-attention mechanism. The distillation operation is used to extract the main attention to reduce the network size. The decoder accepts long sequence input and the target elements are filled with 0 values.

4. The long-term wind power forecasting method based on the improved Informer according to claim 3 is characterized in that Step S3 specifically includes the following steps: Step S3.1: Use MSE as the loss function. The MSE calculation formula is as follows: Where n is the total number of samples; y i is the true value, is the predicted value output by the neural network; Step S3.2: Using the Adam solver, which has high computational efficiency, requires little memory, and is suitable for optimizing a large number of parameters; Step S3.3: Complete model training based on pytorch and save the trained model parameters.

5. The long-term wind power forecasting method based on the improved Informer according to claim 4 is characterized in that Step S4 specifically includes the following steps: Step S4.1: Read the model parameters trained in step S3.3; Step S4.2: Input the historical data corresponding to the time period to be predicted into the wind power prediction model of the wind farm, and use the neural network to calculate the predicted value of the wind power of the wind farm.

6. The long-term wind power forecasting method based on the improved Informer according to claim 4 is characterized in that Introducing residual connections in the encoder of the Informer model.

7. The long-term wind power forecasting method based on improved Informer according to claim 5 is characterized in that Introducing residual connections in the encoder of the Informer model.

8. A long-term wind power prediction device based on improved Informer, characterized by include: The data preprocessing module is used to process missing wind power data, perform correlation analysis on the wind power data, select features that significantly affect wind power, and create training and test sets based on the data. The model building module is used to build a wind power prediction model for wind farms, combining the CNN and Informer models; The model training module is used to select appropriate loss functions and solvers to train the wind power prediction model of the wind farm; The input and output module is used to input the historical data corresponding to the forecast period into the trained wind power forecast model of the wind farm and output the corresponding forecasted wind power of the wind farm.

9. A long-term wind power prediction device based on an improved informer, comprising a memory and a processor, wherein the memory stores a program that can be run on the processor, characterized in that: When the processor executes the program, the long-term wind power prediction method based on the improved Informer according to any one of claims 1 to 7 is implemented.

10. A storage medium, characterized in that: The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the long-term wind power prediction method based on the improved Informer as claimed in any one of claims 1 to 7.

Images