A wind speed prediction correction method based on artificial neural network
By using an LSTM-based neural network model, historical data and real-time observations are used to generate feature vectors and train a wind speed forecast correction model. This solves the problem of error accumulation in long-term predictions of existing wind speed forecasting methods and achieves high-precision, purely objective wind speed forecasts.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING INST OF GREEN & INTELLIGENT TECH CHINESE ACAD OF SCI
- Filing Date
- 2022-10-25
- Publication Date
- 2026-05-12
AI Technical Summary
Existing wind speed forecasting methods suffer from error accumulation in long-term forecasts, and their physical models are complex. Statistical models have good accuracy in short-term forecasts but poor performance in long-term forecasts, making it difficult to achieve high-precision wind speed forecasts.
A wind speed forecast correction model is trained using an LSTM-based neural network model and historical numerical weather prediction results and actual observations. By using cluster analysis and time series prediction methods, feature vectors are generated to reduce prediction errors and achieve high-precision wind speed forecasts.
It achieves high-precision, purely objective wind speed forecasting, reduces human subjective intervention, simplifies the workflow, improves forecast accuracy, and is suitable for real-time wind speed forecasting at the station level.
Smart Images
Figure CN115600498B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence and relates to a wind speed forecast correction method based on artificial neural networks. Background Technology
[0002] With the increasing consumption of traditional energy sources, the research and development of clean and renewable energy is becoming increasingly urgent and important. Among them, wind energy is green, environmentally friendly, and renewable, attracting more and more countries to invest in building wind power plants. Wind speed is a key factor affecting the continuous and stable operation of wind power generation, which places higher demands on wind speed forecasting.
[0003] In recent decades, with the rapid development of numerical weather prediction (NWP), the wind speed forecasting process has shifted from the traditional subjective forecasting based on weather map analysis to a combined subjective and objective model, namely, the "experience-based forecast + NWP result" model. Although numerical weather prediction (NWP) is an objective quantitative calculation, its calculation is based on a grid network, representing the average value of weather elements within a rectangular area. It cannot directly predict elements for a single observation station, requiring forecasters to manually correct based on their experience "from grid to station." When encountering complex weather conditions, forecasters also need to consult with multiple people to determine the final weather element forecast results. Therefore, it inevitably carries a significant degree of subjectivity and introduces human errors that are difficult to quantify.
[0004] Currently, there are many methods suitable for wind speed forecast correction, which can be broadly categorized into two types: physical models and statistical models. Numerical weather prediction (NWP) systems are commonly used physical models for both weather and wind speed forecasts. However, these models are based on large, complex, and experimental meteorological systems, requiring extensive meteorological information and physical mechanisms, which complicates their application and modeling. Statistical models require historical data such as wind speed and timestamps for prediction. Various statistical learning algorithms are also applied in the wind speed forecasting industry, such as the Integrated Moving Average Autoregressive Model (ARIMA), Artificial Neural Networks (ANN), and some hybrid methods. However, while these statistical model-based wind speed forecasts achieve good accuracy in short-term predictions, they exhibit poor performance in long-term predictions due to error accumulation. Although physical models like NWP can achieve large-scale predictions, the results are coarse and have low accuracy. Summary of the Invention
[0005] In view of this, the purpose of this invention is to provide a wind speed forecast correction method based on artificial neural networks. This method uses historical numerical weather prediction results and actual observations to train an artificial neural network model, which is then applied to the daily real-time numerical weather prediction results to obtain station-level wind speed forecast corrections. This achieves better utilization of existing forecast products, inherits past forecast experience, and completely eliminates uncertainties caused by subjective human intervention. This model can minimize prediction errors, providing a high-precision, purely objective, and real-time wind speed forecast post-processing method.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A wind speed forecast correction method based on artificial neural networks includes the following steps:
[0008] S1: Organize the historical forecast grid data of the 10-meter wind field in the target area and the corresponding real data, and perform data preprocessing;
[0009] S2: Use clustering algorithms to perform cluster analysis on the data and generate category features;
[0010] S3: Use time series forecasting methods to analyze the characteristics of 10-meter wind field data, generate feature vectors, and then add the category features generated in step S2 and the forecast lead time of the samples to the feature vectors. Use the sliding window method to process historical data to generate sequence samples.
[0011] S4: Design and training of LSTM-based wind speed forecast correction model;
[0012] S5: Correct the wind speed forecast using the trained LSTM-based wind speed forecast correction model.
[0013] Furthermore, in step S1, the historical forecast grid data containing 10-meter wind field data in the target area is organized, and the corresponding real-time grid data is collected. The historical data spans more than one year, including hourly forecasts reported daily from 00:00 UTC and 12:00 UTC, with a forecast lead time of 48 hours. The organized wind field data is parsed, and then the parsed and processed data is stored in the MongoDB database.
[0014] Further, in step S2, the data preprocessed in S1 is subjected to cluster analysis using the K-means clustering algorithm to generate category features, which are used as one of the features in the next step. The features include statistical features and location features. The statistical features are general features that represent historical NWP wind speed errors and additional related information. The location features are groups clustered by the K-means unsupervised learning algorithm using geographic location data.
[0015] Furthermore, in step S3, eight feature vectors are generated, including ori_diff, Max, Min, mean, median, Q_25, Q_75, and std. Each generated sequence has a length of 30, contains 8 features, is spaced 12 hours apart, and spans 15 days, meaning each sample has a shape of 30. 8. The label corresponding to each sample is the next ori_diff of that sequence.
[0016] Furthermore, in step S3, the wind speed prediction process is defined as... = (.),in It is the predicted wind speed of 10 meters per second. It is a numerical weather prediction model; when actual measurements are collected, the prediction error... for:
[0017] = -
[0018] In the formula, Let be the NWP wind speed at time t. The wind speed at 10 meters per second was measured at time t.
[0019] The analysis of wind field data using the Time Series Prediction (TSP) method specifically includes: developing a model based on previous data and applying it to predict future values over a period of time; learning and modeling historical errors to predict future errors; and correcting the NWP wind speed by adding the predicted errors, as shown in the following formula:
[0020] =
[0021] = +
[0022] in It is the prediction error for time t. This is the corrected NWP wind speed.
[0023] Furthermore, in step S4, a generalized model is trained using location patterns and prediction cycles as input features, specifically including:
[0024] This paper utilizes a Long Short-Term Memory (LSTM) artificial neural network to model time series data with cyclic structures, modeling the functional relationship between input features and the target variable from the most recent past to the future. The LSTM incorporates one unit state c and three gates. Each gate consists of a sigmoid neural network layer and a pointwise multiplication operation. The input gate processes the current input... And the previous output Weighting:
[0025] =σ
[0026] in and These are the weight matrix and bias vector of the input gate, respectively. σ() is a sigmoid function with an output range of (0, 1). When it is close to 1, it tends to select all information in the cell state, while when it is close to 0, it tends to discard input information.
[0027] Then, the information selection is multiplied by the input vector:
[0028] :=
[0029] in and These are the weight matrix and bias vector of the input layer, respectively;
[0030] The selected information is represented as follows:
[0031] := ⊙
[0032] Where ⊙ represents element-wise multiplication. It is the input information composed of the current input and the previous output. It is the selection information of the input layer stored in the cell state;
[0033] The forget gate determines which part of the information stored in the cell state should be forgotten, as shown in the following formula:
[0034] = ,
[0035] in and These are the weight matrix and bias vector of the forget gate, respectively;
[0036] The new cell state is:
[0037]
[0038] The output gate determines which part of the information should be output, i.e.
[0039] ,
[0040] in and These are the weight matrix and bias vector of the forget gate, respectively.
[0041] Furthermore, ( , , ,…, The data is input into an LSTM network, which then provides a prediction for the next data point. ,from The ori_diff variable is extracted to correct the next forecast data.
[0042] Furthermore, the training process of the LSTM-based wind speed forecast correction model is as follows:
[0043] The dataset is divided into a training set and a validation set;
[0044] On the training set, the optimal hyperparameter configuration, including the number of LSTM units, initial learning rate, and batch size, is obtained by using the 5-fold cross-validation method.
[0045] The model weights are trained using the stochastic gradient descent method, with a maximum number of training generations set.
[0046] Wind speed data is a continuous variable, and the mean squared error (MSE) is used as the loss function:
[0047]
[0048] Train the model until the loss function is minimized, obtain the optimal model parameters, and the training is complete.
[0049] The beneficial effects of this invention are as follows: Compared with traditional wind speed correction models, numerical results show that deep learning algorithms have better feature representation capabilities and can improve prediction accuracy. This invention can also significantly simplify the workflow while maintaining prediction accuracy. Furthermore, this method is based on objective calculations and is free from any human subjectivity. Exploring this post-processing system will help the field of wind speed forecasting shift from a forecasting model combining subjective and objective methods to a purely objective forecasting model.
[0050] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0051] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:
[0052] Figure 1 This is a clustering graph of GPS locations;
[0053] Figure 2 A heatmap of NWP prediction errors for a certain location;
[0054] Figure 3 Time series prediction plot for correction;
[0055] Figure 4 Here is a diagram of the RNN model structure;
[0056] Figure 5 Diagram of the internal structure of an LSTM neuron;
[0057] Figure 6 This is a schematic diagram of LSTM time series processing;
[0058] Figure 7 Training curves for artificial neural network models;
[0059] Figure 8 A plot showing the overall RMSE results based on location;
[0060] Figure 9 A graph showing the overall RMSE results based on the forecast period;
[0061] Figure 10 This is a schematic diagram of the wind speed forecast correction method based on artificial neural networks described in this invention.
[0062] Figure 11 The above are RMSE heatmaps of an embodiment of the present invention, wherein (a) is the heatmap before correction and (b) is the heatmap after correction. Detailed Implementation
[0063] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0064] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0065] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0066] A wind speed forecast correction method based on artificial neural networks, comprising the following steps:
[0067] S1: Organize the historical forecast grid data of the 10-meter wind field in the target area and the corresponding real data, and perform data preprocessing;
[0068] S2: Use clustering algorithms such as K-means to perform cluster analysis on the data and generate category features, which will be used as one of the features in the next step;
[0069] S3: The Time Series Prediction (TSP) method is used to analyze the characteristics of the 10-meter wind field data, generating eight feature vectors including ori_diff and Max. The categorical features generated in S2, along with the forecast lead time of the sample, are then added to the feature vectors. A sliding window method is used to process historical data and generate sequence samples.
[0070] S4: Artificial Neural Network Model Design and Training;
[0071] S5: Validate and evaluate the trained artificial neural network model.
[0072] In step S1, historical forecast grid data containing 10-meter wind field data within the target area are organized, and corresponding real-time grid data are collected. The historical data spans more than one year, including hourly forecasts starting at 00:00 and 12:00 (UTC 00:00 and UTC 12:00) daily, with a forecast lead time of 48 hours. The organized data is parsed; the wind field data format is HDF5. The HDF5 data is parsed and processed using the pandas module in Python, and then stored in a MongoDB database.
[0073] Historical wind field forecast grid data of 10 meters within the target area and the corresponding real-time data are selected, as shown in Table 1.
[0074] Table 1
[0075]
[0076] In step S2, cluster analysis is performed on the preprocessed data from S1 using clustering algorithms such as K-means to generate category features, which will be used as one of the features in the next step.
[0077] The features, created through feature engineering, can be divided into two groups: statistical features and location features. Statistical features are general features representing historical NWP wind speed errors and additional relevant information. Location features are groups clustered using a K-means unsupervised learning algorithm that uses geographic location data. Figure 1 The data shows the clustering of geographic location data, where locations are classified into 3 groups.
[0078] In step S3, the Time Series Prediction (TSP) method is used to analyze the characteristics of the 10-meter wind field data, generating eight feature vectors including ori_diff, Max, Min, mean, median, Q_25, Q_75, and std (see Table 2). The categorical features generated in S2, along with the forecast lead time of the sample, are then added to the feature vectors. A sliding window method is used to process historical data and generate sequence samples. Each sequence is 30 bytes long, contains eight features, is spaced 12 hours apart, and spans 15 days; that is, each sample has a shape of 30 bytes. 8. The label corresponding to each sample is the next ori_diff for that sequence.
[0079] Table 2
[0080]
[0081] In step S4, the wind speed prediction process is defined in the design of the artificial neural network model. = (.),in It is the predicted wind speed of 10 meters per second. This is a numerical weather prediction model. Numerical weather prediction models contain errors, partly due to the quality of the data driving the model and computational limitations. Some errors are systematic and need to be reduced by applying statistical post-processing methods.
[0082] The current numerical weather prediction (NWP) model forecasts 10-meter wind speed data every 12 hours over a period of 13–48 hours. However, actual measurements are collected hourly, therefore, prediction errors exist. It can be calculated using the following formula. When collecting actual measured values...
[0083] = - (1)
[0084] In the formula, Let be the NWP wind speed at time t. The wind speed at 10 meters per second was measured at time t. Figure 2 This displays a heatmap showing the prediction error for a location.
[0085] In view of this, in order to capture the variation pattern of wind speed prediction error, this invention uses the time series prediction (TSP) method to analyze wind field data.
[0086] Time series forecasting (TSP) methods develop models based on previous data and apply them to predict future values over a period of time. According to the principles of TSP, historical errors are learned and modeled to predict future errors, such as... Figure 3 As shown. After accounting for the prediction error, the NWP wind speed can be easily corrected simply by adding the prediction error. This process is represented by the following equation:
[0087] = (2)
[0088] = + (3)
[0089] in It is the prediction error for time t. This is the corrected NWP wind speed.
[0090] However, TSP-based algorithms only utilize historical error information, while other features such as temperature and air pressure are often ignored. Furthermore, TSP-based algorithms require separate modeling for different locations and prediction periods, which is extremely difficult and time-consuming. To broaden the application scenarios of prediction models, this invention chooses to use deep learning algorithms—artificial neural network models—because they can automatically learn arbitrarily complex mappings from input to output, which has been widely used in recent time series prediction tasks.
[0091] Deep learning-based time series prediction algorithms can take into account both location patterns and prediction periods. Therefore, a generalized model can be trained using location patterns and prediction periods as input features.
[0092] Recurrent Neural Networks (RNNs) are suitable for modeling time series data with a cyclic structure, which uses neural networks to model the functional relationship between input features from the most recent past to the future target variable. Figure 4 The structure of an RNN is shown.
[0093] Learn the hidden state from consecutive time slots from historical data; specifically, output... The prediction depends not only on the input And it also depends on the hidden state. As shown in Equation 4. By analyzing previous output Weighting is applied to remember previous information. However, RNNs suffer from the vanishing gradient problem due to the multiplication of the weight matrix at each time step. Therefore, RNNs are not good at capturing non-stationary dependencies that occur over long periods of time.
[0094] = (4)
[0095] In view of this, consider that Long Short-Term Memory Artificial Neural Network (LSTM) is a variant of RNN, which can overcome the gradient vanishing problem in RNN. Figure 5 The structure of an LSTM unit is shown, which incorporates a unit state c and three gates to enable long-term storage and access to information. Each gate consists of a sigmoid neural network layer and a pointwise multiplication operation, selectively allowing information to pass through. First, the input gate processes the current input... And the previous output Weighting is performed, as shown in Equation 5.
[0096] =σ (5)
[0097] in and These are the weight matrix and bias vector of the input gate, respectively, and σ() is the sigmoid function with an output range of (0, 1). When it is close to 1, it tends to select all information from the cell state, while when it is close to 0, it tends to discard input information. Then, the selected information is multiplied by the input vector.
[0098] := (6)
[0099] in and These are the weight matrix and bias vector of the input layer, respectively. The selected information is represented as...
[0100] := ⊙ (7)
[0101] Where ⊙ represents element-wise multiplication. It is the input information composed of the current input and the previous output. It is the selection information of the input layer stored in the cell state.
[0102] The forget gate determines which part of the information stored in the cell state should be forgotten, as shown in the following equation:
[0103] = , (8)
[0104] in and These are the weight matrix and bias vector of the forget gate, respectively. Then, the new cell state is...
[0105] (9)
[0106] The output gate determines which part of the information should be output, i.e.
[0107] , (10)
[0108] in and These are the weight matrix and bias vector of the forget gate, respectively.
[0109] In this invention, ( , , ,…, The data is input into an LSTM network, which then provides a prediction for the next data point. We started from... Extracting the `ori_diff` variable allows for corrections to the next forecast data, such as... Figure 6 As shown.
[0110] Training process of wind speed forecast correction model based on Long Short-Term Memory (LSTM) artificial neural network
[0111] First, the dataset is split into training and validation sets in a 7:3 ratio. On the training set, a 5-fold cross-validation method is used to obtain the optimal hyperparameter configuration, including the number of LSTM units, the initial learning rate, and the batch size. The initial number of units is determined to be 5, the learning rate to be 0.01, and the batch size to be 32. Stochastic gradient descent is used to train the model weights, with a maximum training iteration count of 20. This module is implemented in Python, building and training a neural network for each forecast timeframe.
[0112] Wind speed data is a continuous variable, and mean squared error (MSE) is used.
[0113] (11)
[0114] This loss function is suitable for training regression models, enabling the model to converge relatively quickly. The loss function reaches its minimum point around the 16th generation, which is then saved as the optimal model parameters. The training curve (taking 13 hours as an example) is shown below. Figure 7 As shown.
[0115] In step S5, the trained Long Short-Term Memory (LSTM) artificial neural network model is validated and evaluated. In this invention, actual weather forecast data is used to verify the superiority of the artificial neural network correction model, comparing the original numerical forecast wind speed error with the corrected wind speed error. (The following experiments were implemented on a Linx PC equipped with an AMD Ryzen 5 3550H, 2.1GHz CPU, 16GB RAM, and Python 3.8 with Tensorflow 2.8.0.) For the TSP-based algorithm, the training data is the past raw NWP wind speed error for a specific location and a specific forecast period. Conversely, for the deep learning-based algorithm, the past raw NWP wind speed error for all locations and all forecast periods is used to train a general model.
[0116] Location-based assessment
[0117] To evaluate the correction and prediction performance of the proposed wind speed forecast correction model across all locations, the root mean square error (RMSE) was used as the performance metric. A smaller RMSE value indicates better performance of the proposed model.
[0118] (12)
[0119] where t∈ To predict time slots, n∈ This is for the prediction period. Therefore... Evaluate the overall prediction performance of location loc.
[0120] Figure 8 The overall prediction performance was evaluated. Box plots were used to compare location-based overall performance. The first box on the left shows the uncorrected predictions. Rolling mean, ETS, and ARIMA represent the predictions of the NWP model corrected using regular data, while LSTM and CNN represent the predictions corrected using deep learning algorithms. Figure 8 As can be seen, without prediction correction, the lowest RMSE is 1.1, while the average RMSE is almost 2. Both the TSP-based algorithm and the prediction correction algorithm based on the Long Short-Term Memory network can reduce the prediction error to 1.0. Specifically, Rolling mean and LSTM show the best performance.
[0121] In summary, correcting the prediction results is necessary because it can significantly improve prediction accuracy. However, as mentioned earlier, the Rolling mean, ETS, and ARIMA methods all require building local numerical weather prediction models. This means that if the location changes, the model will no longer be applicable. Artificial neural network models do not require repeated training with location patterns and prediction cycle data, making them more versatile. Furthermore, Figure 8 The results show that the CNN algorithm lowers the upper and lower bounds of the RMSE compared to the LSTM algorithm, but the LSTM algorithm has a lower average RMSE. Reliability has the highest priority in weather forecasting. Therefore, LSTM remains the most recommended numerical weather prediction correction algorithm.
[0122] Evaluation based on prediction cycle
[0123] Similarly, the RSME of the wind speed forecast correction model was calculated for different forecast periods to evaluate its correction and forecast performance.
[0124] (13)
[0125] Where l∈ It is a set of locations.
[0126] To verify the application of the above algorithm in short-term wind speed forecasting, the forecast results for 48 hours were plotted, as shown below. Figure 9 As shown in the figure, the RMSE is 2.05 without prediction correction. After prediction correction, the prediction error is reduced to below 1.7. Note that the RMSE of the CNN algorithm is higher in the first 20 hours, then drops to 1.7, and remains stable in subsequent periods. The numerical forecast of the wind speed forecast correction model based on Long Short-Term Memory (LSTM) networks shows superior overall performance compared to other algorithms.
[0127] In summary, the wind speed forecast correction model based on Long Short-Term Memory (LSTM) artificial neural network can respond promptly throughout the wind speed forecast correction process and achieve excellent wind speed forecast correction results.
[0128] The flowchart of this embodiment is as follows: Figure 10 As shown, after training the artificial neural network model, it was run on the computing platform. Ten grid points of data were randomly selected for correction. A heatmap of the root mean square error (RMSE) of 100 original and post-processed model data relative to the measured data was also extracted, with the initial report time being 12:00 and the forecast time being 13:00. Figure 11 As shown, Figure 11 In the figure, (a) represents the data before correction and (b) represents the data after correction. The darker the color, the greater the error. The average RMSE of the model forecast data is about 2.0, and the RMSE after correction is about 1.6.
[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A wind speed forecast correction method based on artificial neural networks, characterized in that: Includes the following steps: S1: Organize the historical forecast grid data of the 10-meter wind field in the target area and the corresponding real data, and perform data preprocessing; S2: Use clustering algorithms to perform cluster analysis on the data and generate category features; S3: Use time series forecasting methods to analyze the characteristics of 10-meter wind field data, generate feature vectors, and then add the category features generated in step S2 and the forecast lead time of the samples to the feature vectors. Use the sliding window method to process historical data to generate sequence samples. S4: Design and training of an LSTM-based wind speed forecast correction model; S5: Correct the wind speed forecast using the trained LSTM-based wind speed forecast correction model; In step S1, the historical forecast grid data containing 10-meter wind field data in the target area is organized, and the corresponding real-time grid data is collected. The historical data spans more than one year, including hourly forecasts reported daily from 00:00 UTC and 12:00 UTC, with a forecast lead time of 48 hours. The organized wind field data is parsed, and then the parsed and processed data is stored in the MongoDB database. In step S2, the K-means clustering algorithm is used to perform cluster analysis on the data preprocessed in S1 to generate category features, which are used as one of the features in the next step. The features include statistical features and location features. The statistical features are general features that represent historical NWP wind speed errors and additional related information. The location features are groups clustered by the K-means unsupervised learning algorithm using geographic location data. In step S3, eight feature vectors are generated, including ori_diff, Max, Min, mean, median, Q_25, Q_75, and std. Each generated sequence has a length of 30, contains 8 features, is spaced 12 hours apart, and spans 15 days, meaning each sample has a shape of 30.
8. The label corresponding to each sample is the next ori_diff of that sequence.
2. The wind speed forecast correction method based on artificial neural networks according to claim 1, characterized in that: In step S3, the wind speed prediction process is defined as... = (.),in It is the predicted wind speed of 10 meters per second. It is a numerical weather prediction model; when actual measurements are collected, the prediction error... for: = - In the formula, Let be the NWP wind speed at time t. The wind speed at 10 meters per second was measured at time t. The analysis of wind field data using time series forecasting methods specifically includes: developing a model based on previous data and applying it to predict future values over a period of time; learning and modeling historical errors to predict future errors; and correcting the NWP wind speed by adding the predicted errors, as shown in the following formula: = = + in It is the prediction error for time t. This is the corrected NWP wind speed.
3. The wind speed forecast correction method based on artificial neural networks according to claim 1, characterized in that: In step S4, a generalized model is trained using location patterns and prediction periods as input features, specifically including: This paper utilizes a Long Short-Term Memory (LSTM) artificial neural network to model time series data with cyclic structures, modeling the functional relationship between input features and the target variable from the most recent past to the future. The LSTM incorporates one unit state c and three gates. Each gate consists of a sigmoid neural network layer and a pointwise multiplication operation. The input gate handles the current input... And the previous output Weighting: =s in and These are the weight matrix and bias vector of the input gate, respectively. σ() is a sigmoid function with an output range of (0, 1). When it is close to 1, it tends to select all information in the cell state, while when it is close to 0, it tends to discard input information. Then, the information selection is multiplied by the input vector: := in and These are the weight matrix and bias vector of the input layer, respectively; The selected information is represented as follows: := ⊙ Where ⊙ represents element-wise multiplication. It is the input information composed of the current input and the previous output. It is the selection information of the input layer stored in the cell state; The forget gate determines which part of the information stored in the cell state should be forgotten, as shown in the following formula: = , in and These are the weight matrix and bias vector of the forget gate, respectively; The new cell state is: The output gate determines which part of the information should be output, i.e. , in and These are the weight matrix and bias vector of the forget gate, respectively.
4. The wind speed forecast correction method based on artificial neural networks according to claim 3, characterized in that: Will( , , ,…, The data is input into an LSTM network, which then provides a prediction for the next data point. ,from The ori_diff variable is extracted to correct the next forecast data.
5. The wind speed forecast correction method based on artificial neural networks according to claim 1, characterized in that: The training process of the LSTM-based wind speed forecast correction model is as follows: The dataset is divided into a training set and a validation set; On the training set, the optimal hyperparameter configuration, including the number of LSTM units, initial learning rate, and batch size, is obtained by using the 5-fold cross-validation method. The model weights are trained using the stochastic gradient descent method, with a maximum number of training generations set. Wind speed data is a continuous variable, and the mean squared error (MSE) is used as the loss function: Train the model until the loss function is minimized, obtain the optimal model parameters, and the training is complete.