A power grid dispatch optimization method based on deep learning

Abstract

The present invention provides a power grid dispatch optimization method based on a deep learning network, comprising the following steps: collecting wind turbine power output data and spatiotemporal data according to time and preprocessing the data; designing a multimodal network structure suitable for wind power prediction; designing a loss function to train the multimodal network to obtain a trained network model; and using the trained model to perform real-time wind power prediction. By employing a deep learning approach, the power grid dispatch optimization method based on a deep learning network utilizes the attention mechanism in deep learning to optimize the focus on key information, improve the accuracy of wind power prediction, and predict wind power, thereby achieving better power grid energy management and dispatch optimization, improving the utilization efficiency of renewable energy, reducing reliance on traditional energy, and ensuring the stability and efficiency of the power grid.

Description

Technical Field

[0001] The present invention relates to the intersection of artificial intelligence and power grid dispatch optimization, and specifically to a power grid dispatch optimization method based on deep learning. Background Art

[0002] In modern power systems, grid dispatch optimization is a key element in ensuring grid stability and improving energy efficiency. Traditional grid dispatch methods rely on pre-set rules and manual decision-making, which often struggle to cope with the volatility and unpredictability of renewable energy, particularly wind power. As the proportion of renewable energy in the power system continues to increase, traditional methods, lacking real-time and flexibility, often struggle to maximize the potential of wind power and effectively integrate it into the grid.

[0003] With the rapid development of machine learning and artificial intelligence technologies, deep learning techniques have been widely studied and applied to power system prediction and optimization tasks. However, existing deep learning models typically only consider time series predictions when processing wind power generation data, ignoring the spatial relationships between generators. These spatial relationships can significantly impact prediction accuracy due to factors such as the uneven spatial distribution of wind resources, wake effects, and interference effects. Therefore, studying how to comprehensively utilize time series and spatial location information to improve the accuracy of wind power generation forecasts and the efficiency of grid dispatch has become a research direction in current power system management.

[0004] Chinese patent publication number CN 117709502 A discloses "A Wind Power Prediction Method Based on the CEEMDAN-IHBA-XGBoost Model," proposing a method that can effectively process and predict the power output of wind farms through multimodal decomposition and optimized prediction of wind power. This method can use the maximum information coefficient (MIC) for feature selection and enhance the signal-to-noise ratio of the data through the CEEMDAN algorithm, ultimately using XGBoost for accurate prediction. However, this method relies primarily on traditional time series prediction technology and does not fully utilize the spatial location data of wind power stations, which may limit the comprehensiveness and accuracy of the prediction model in highly dynamic and complex wind scenarios. In addition, although the IHBA algorithm is introduced to optimize the hyperparameters of XGBoost, the flexibility and real-time performance of this method when processing high-dimensional data and complex network structures may not be as good as methods using deep learning and attention mechanisms. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a deep learning-based method for optimizing power grid dispatching, addressing the issues raised in the aforementioned background technology. This invention provides a novel multimodal network that comprehensively utilizes the spatiotemporal data of wind turbines. By converting spatial position data into 2D images and combining them with one-dimensional signal processing of time series information, this network comprehensively captures the spatial relationships and dynamic changes of individual turbines within a wind farm. This method leverages the attention mechanism within deep learning to optimize the focus on key information and improve the accuracy of wind power prediction.

[0006] In order to achieve the above object, the present invention is implemented through the following technical solution: a power grid dispatch optimization method based on deep learning, comprising the following steps:

[0007] S1. Collect historical power output data and spatiotemporal data of each wind turbine in the wind power plant and preprocess the data.

[0008] S2. Design a multimodal network structure suitable for wind power prediction, including a one-dimensional information feature extraction module, a multi-channel image spatial feature extraction module, and a feature fusion module;

[0009] S3. Design a loss function to train the multimodal network and obtain a trained network model;

[0010] S4. Use the trained model to make real-time predictions of wind power generation to optimize grid scheduling.

[0011] Furthermore, step S1 specifically includes the following contents:

[0012] S1-1. Obtain the power generation data, wind speed, wind direction, air pressure and latitude and longitude information of each wind turbine in the wind power plant every 4 hours within one year;

[0013] S1-2. Clean the acquired data. First, use the weighted linear interpolation method to process the missing values ​​of wind speed and wind direction. The formula is:

[0014] x missing =w1·x before +w2·x after

[0015]

[0016] Among them, x missing is the predicted value for the missing time point; x before and x after are the known values ​​before and after the missing point respectively; w1 and w2 are weights, calculated based on time distance; t before , t after , tmissing is the corresponding time point.

[0017] Then the IQR rule is used to deal with outliers in wind speed. The formula is:

[0018] IQR=Q3-Q1

[0019] LowerBound=Q1-k×IQR

[0020] UpperBound=Q3+k×IQR

[0021] Where Q1 is the lower quartile (25%) of all values ​​in the data set; Q3 is the upper quartile (75%) of all values ​​in the data set; IQR is the interquartile range; k is an adjustable parameter that is set according to the distribution sensitivity of the data; Upper Bound and Lower Bound are the upper and lower bounds respectively, and any value below the lower bound or above the upper bound needs to be eliminated.

[0022] S1-3. Convert the obtained data to convert the wind direction from an angle (0-360 degrees) to sine and cosine values. The formula is as follows:

[0023]

[0024] Where "Wind Direction" is the wind direction angle in degrees.

[0025] S1-4. Remove the proportional effect in the data by normalization. The formula is:

[0026]

[0027] Among them, x[n] is the original data, x normalized [n] is the normalized data, min(x) is the minimum value in the data, max(x) is the maximum value in the data, and α and β are adjustable scaling factors and offsets.

[0028] S1-5. Perform sine and cosine transformations on the longitude and latitude coordinates of the wind turbine. First, convert the degrees to radians. The formula is as follows:

[0029]

[0030] Among them, Radians Lat、Long Degrees are the radians corresponding to latitude and longitude respectively; Lat、Long are latitude and longitude respectively; π is pi.

[0031] Then calculate the sine and cosine values, the formula is as follows:

[0032] Latitude cos =cos(Radians Lat )

[0033] Latitude sin =sin(Radians Lat )

[0034] Longitude cos =cos(Radians Long )

[0035] Longitude sin =sin(Radians Long )

[0036] Among them, Latitude cos is the cosine value of the dimension; Latitude sin is the sine value of the dimension; Longitude cos is the cosine value of longitude; sin is the sine of longitude.

[0037] S1-6. Arrange all data in a time series format so that it can be input into the one-dimensional feature extraction module.

[0038] Furthermore, the S1 further includes:

[0039] S1-7. Define an M×N grid and map the geographic coordinates of each wind turbine to the grid coordinates using the following formula:

[0040]

[0041] Where x and y are the grid coordinates of the generator; N and M are the number of grid columns and rows, respectively, which are determined by the specific distribution and number of generators; longitude and latitude are the actual longitude and latitude of the power station; max_longitude and min_latitude are the maximum longitude and latitude of all generators, respectively; min_longitude and min_latitude are the minimum longitude and latitude of all generators, respectively.

[0042] S1-8. For each generator, use the x and y calculated above as grid coordinates to map the wind speed, wind direction, atmospheric pressure, and power generation data to the corresponding grid cell.

[0043] S1-9. Generate an image based on the data values ​​on the grid. The image has 5 channels, which are respectively allocated by the sine and cosine of wind speed, wind direction, atmospheric pressure, and power generation.

[0044] Furthermore, the multimodal network structure in S2 specifically includes:

[0045] S2-1. Design a multi-channel image spatial feature extraction module as follows:

[0046] S2-1-1. Input a multi-channel 2D image, use a 3×3 convolution kernel for initial convolution, followed by batch normalization and LeakyReLU activation function, then use a 1×1 convolution kernel for deep feature extraction, and then go through batch normalization and LeakyReLU activation function again;

[0047] S2-1-2. Introduce the channel and spatial dual attention mechanism. The first is the channel attention mechanism. Specifically, through the squeezing operation, the two-dimensional feature map of each channel is compressed into a single value. The formula is as follows:

[0048]

[0049] where z c represents the global feature of the cth channel, which is the global average of all features of the channel; u c is the feature map of the cth channel, u c (i, j) is the value of the feature map at position (i, j); H and W are the height and width of the feature map; ∑ is the summation symbol, which is used to accumulate the value of the entire feature map; is the normalization factor for the averaging operation.

[0050] Then comes the excitation operation, which uses a small fully connected neural network to learn the weight of each channel to strengthen the features of important channels. The formula is as follows:

[0051] s c =σ(W2δ(W1z c ))

[0052] where s c represents the weight vector of the cth channel; z is the global feature vector from the squeeze operation, which is input to the excitation network; W1 and W2 represent the weight matrices of the two fully connected layers, the first fully connected layer is used for dimensionality reduction, and the second fully connected layer is used to restore the dimension; δ is the ReLU activation function, which is used to add nonlinearity; σ is the Sigmoid activation function, which is used to limit the weight between 0 and 1.

[0053] The last step is the recalibration operation, which is used to multiply the calculated channel weights by the original feature map channel by channel to emphasize the features of important channels. The formula is:

[0054]

[0055] where s c is the weight of channel c, is the weighted feature map.

[0056] Then comes the spatial attention mechanism, the formula is as follows:

[0057]

[0058] Among them A s is the generated spatial attention map; Conv 3×3 is a 3×3 convolution kernel used to aggregate features across channels; σ is the Sigmoid function; It is the feature map processed by the channel attention mechanism.

[0059] Finally, the spatial attention mechanism is applied, and the formula is as follows:

[0060]

[0061] Represents the feature map after spatial attention adjustment; ⊙ is the element multiplication operation, representing A s Each spatial weight in Multiply the values ​​at corresponding positions in .

[0062] S2-1-3. Add a global average pooling layer, a dropout layer, and a fully connected layer after the attention mechanism, apply an adaptive activation function, and output the final feature vector; the adaptive parameterized ReLU (APReLU) activation function is expressed as follows:

[0063]

[0064] Where x is the input of the activation function, and α, β, γ, and δ are learnable parameters that can be dynamically adjusted according to the input features during training.

[0065] Furthermore, the S2 also includes the following contents:

[0066] S2-2. Design a one-dimensional information feature extraction module, as follows:

[0067] S2-2-1. Input the preprocessed time series data, first use a 3×1 convolution kernel for convolution to capture short-term dependencies, then pass it through the ReLU activation function, and then use a 5×1 convolution kernel for convolution to capture longer-term dependencies.

[0068] S2-2-2. Apply 2×2 average pooling after convolution, and then apply a batch normalization layer.

[0069] S2-2-3. Introduce the temporal attention mechanism and use a small convolutional network to generate temporal attention weights. The formula is as follows:

[0070] A t =σ(Conv 1×1 (F conv2 ))

[0071] Among them A t is the temporal attention map; F conv2 It is the feature map after the second convolution; Conv 1×1 is a small convolutional layer that generates temporal attention weights; σ is the Sigmoid activation function.

[0072] Then the attention weight is applied, the formula is as follows:

[0073]

[0074] represents the feature map after temporal attention adjustment; ⊙ is the element multiplication operation.

[0075] S2-2-4. After applying the temporal attention mechanism, the results Input to the LSTM layer to capture discontinuous patterns in the time series.

[0076] S2-2-5. After LSTM, perform global average pooling on the output, then apply the ELU activation function, and finally apply the Dropout layer and the fully connected layer.

[0077] S2-3. Design a multi-channel image spatial feature extraction module and perform feature fusion on the outputs of the one-dimensional information feature extraction module and the multi-channel image spatial feature extraction module. The details are as follows:

[0078] S2-3-1. Perform global average pooling on the output features of the two modules to ensure the same dimensions, and then use Z-score to normalize the features.

[0079] S2-3-2. Concatenate the processed one-dimensional and two-dimensional feature vectors. The formula is as follows:

[0080] F combined =[F 1D ; F 2D ]

[0081] Among them F combined represents the joint eigenvector after concatenation; F 1D and F 2D They represent the features output by the one-dimensional feature extraction module and the multi-channel image space feature extraction module respectively.

[0082] Then use a fully connected layer from F combined The feature weights are learned in the following formula:

[0083] w 1D ,w 2D =softmax(W·F combined +b)

[0084] where w 1D ,w 2D is the obtained weight; W is the weight matrix of the fully connected layer; b is the bias term; F combined represents the joint feature vector; the softmax activation function ensures that the sum of the output weights is 1.

[0085] Then perform weighted feature fusion, the formula is as follows:

[0086] F fused =w 1D ·F 1D +w 2D ·F 2D

[0087] Among them F fused represents the fused features; w 1D and w 2D is the weight obtained through training; F 1D and F 2D They represent the features output by the one-dimensional feature extraction module and the multi-channel image space feature extraction module respectively.

[0088] S2-3-3. After feature fusion, apply the first fully connected layer with the ReLU activation function, then add the Dropout layer, and then apply the second fully connected layer with the ReLU activation function. Finally, output the predicted value.

[0089] Furthermore, the loss function designed in S3 is as follows:

[0090] S3-1. The loss function consists of three parts: prediction error loss, modal coordination loss, and regularization term.

[0091] S3-1-1. The prediction error loss part measures the difference between the predicted value and the actual value through the mean square error. The formula is as follows:

[0092]

[0093] where y i is the actual power value of sample i, is the power value predicted by the model, and N is the total number of samples.

[0094] S3-1-2, modality coordination loss aims to optimize the coordination between different modality features (image and time series data), and the formula is as follows:

[0095]

[0096] Among them, F 1D,j and F 2D,j They represent the extraction of one-dimensional and two-dimensional features on the j-th sample respectively; corr represents the correlation coefficient; M is the number of modes.

[0097] S3-1-3, the regularization term is used to control the complexity of the model and prevent overfitting. The formula is as follows:

[0098]

[0099] where w k represents the weight of each layer of the model, and λ is the regularization coefficient.

[0100] Furthermore, the final loss function of the three parts is expressed as follows:

[0101] L=L pred +αL modal +L reg

[0102] Here, α is an adjustable hyperparameter used to balance the importance of prediction error loss and modality coordination loss.

[0103] Furthermore, the detailed steps of performing real-time prediction and power grid dispatch optimization in S4 are as follows:

[0104] S4-1. Before conducting real-time forecasts, you need to ensure that all relevant real-time data sources are ready and can be accessed in real time. This data includes: wind speed and direction data for each generator, air pressure and temperature data, power generation data, and latitude and longitude data.

[0105] S4-2. As in the training phase, all real-time data must undergo the same preprocessing steps to ensure data consistency and model effectiveness. This includes data cleaning, data transformation, and feature extraction.

[0106] S4-3. Use the preprocessed data to perform real-time predictions using the trained multimodal network model. The model will output the predicted power generation for the next 7 days.

[0107] S4-4. The predicted wind power data will be directly input into the grid dispatching system. The system will dynamically adjust the allocation of power generation resources in the grid based on the predicted results and the current load demand of the grid.

[0108] Furthermore, the following aspects are optimized in S4-4:

[0109] S4-4-1. Power generation efficiency: Adjust the output of the wind farm based on the predicted results of wind power generation to improve power generation efficiency.

[0110] S4-4-2. Power supply reliability: Reduce the risk of power outages due to unstable wind power through forecast data.

[0111] S4-4-3. Cost control: Optimize the power supply combination, reduce the use of expensive peak-shaving power stations, and reduce power generation costs.

[0112] Beneficial effects of the present invention:

[0113] 1. This deep learning-based grid dispatch optimization method uses advanced data preprocessing techniques, such as sine-cosine transformation of wind direction and IQR outlier processing, to effectively improve the quality of input data. This refined processing helps the model better understand and utilize data, thereby reducing prediction errors.

[0114] 2. This deep learning-based grid dispatch optimization method simultaneously processes time series and spatial image data. The multimodal network of the present invention comprehensively analyzes the dynamics and spatial distribution of wind turbines. This integrated approach provides the model with richer information, facilitating more accurate forecasts of wind power generation over the next several hours to a week.

[0115] 3. By converting the spatial positions of wind turbines into 2D images, this invention captures the relative positions and potential interactions between turbines. This consideration of spatial relationships significantly improves the model's accuracy in predicting overall wind farm dynamics. The model's integrated channel and spatial attention mechanisms enable it to automatically identify and focus on the most important data features for prediction, such as areas of sudden wind speed changes. This dynamic attention capability enables more accurate predictions and faster responses.

[0116] 4. This invention designs a composite loss function that incorporates a modal coordination loss, which not only optimizes prediction error but also strengthens consistency across different data modalities. This approach ensures optimal coordination between different data sources, thereby improving overall prediction stability and accuracy. By applying techniques such as algorithm optimization and hardware acceleration, the model can be rapidly deployed and maintain efficient operation in resource-constrained environments. This is particularly critical for power grid operations requiring real-time scheduling, ensuring that the model is both accurate and practical.

[0117] 5. This invention rapidly responds to environmental changes and grid demands by processing and analyzing wind turbine spatiotemporal data in real time. This rapid data processing capability enables grid operators to instantly adjust grid loads and power generation plans to address sudden changes in wind speed and direction, thereby reducing energy waste caused by insufficient or excessive power generation.

[0118] 6. The prediction results of this invention can help operators more efficiently allocate and utilize renewable energy, particularly wind energy. By accurately predicting wind power generation over the next few days or even hours, power grids can more finely adjust the ratios between different energy sources, optimize the overall energy mix, and reduce reliance on traditional fossil fuels. BRIEF DESCRIPTION OF THE DRAWINGS

[0119] Figure 1 A flowchart of a power grid dispatch optimization method based on a deep learning network provided by an embodiment of the present invention;

[0120] Figure 2 A network structure diagram of a power grid dispatch optimization method based on a deep learning network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0121] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0122] See also Figures 1 to 2 , the present invention provides the following technical solution: a power grid dispatch optimization method comprising the following steps:

[0123] S1. Collect historical power output data and spatiotemporal data of wind turbines and preprocess the data.

[0124] S2. Design a multimodal network structure suitable for wind power prediction, including a one-dimensional information feature extraction module and a multi-channel image spatial feature extraction module;

[0125] S3. Design a loss function to train the multimodal network and obtain a trained network model;

[0126] S4. Use the trained model to make real-time predictions of wind power generation to optimize grid scheduling.

[0127] The step S1 is specifically as follows:

[0128] S1-1. Obtain 4-hourly historical power generation data, wind speed, wind direction, air pressure, and the longitude and latitude of each wind turbine for a year. This data, which includes at least one year of seasonal variations, helps the model learn seasonal variations in wind power generation patterns. 4-hourly power generation data helps capture more detailed variations in wind speed and direction, which is particularly important for real-time scheduling.

[0129] S1-2. Clean the acquired data. First, use weighted linear interpolation to address missing values ​​for wind speed and direction. In traditional linear interpolation, the value of a predicted point is based on the average of two adjacent points. However, weighted linear interpolation uses time or spatial distance as a weight, weighting adjacent points so that observation points closer to the predicted point have greater influence. The improved formula is:

[0130] x missing =w1·x before +w2·x after

[0131]

[0132] Among them, x missing is the predicted value for the missing time point; x before and x after are the known values ​​before and after the missing point respectively; w1 and w2 are weights, calculated based on time distance; t before , t after , t missing is the corresponding time point.

[0133] The IQR rule is then used to process outliers in wind speed. IQR (interquartile range) is a commonly used statistical method for identifying outliers in data. IQR is based on the quartiles of the data and is a very robust outlier detection method that is not affected by extreme values ​​in the data. Its formula is:

[0134] IQR=Q3-Q1

[0135] LowerBound=Q1-k×IQR

[0136] UpperBound=Q3+k×IQR

[0137] Where Q1 is the lower quartile (25%) of all values ​​in the data set; Q3 is the upper quartile (75%) of all values ​​in the data set; IQR is the interquartile range; k is an adjustable parameter that is set according to the distribution sensitivity of the data; Upper Bound and Lower Bound are the upper and lower bounds respectively. Any value below the lower bound or above the upper bound is considered an outlier and needs to be removed.

[0138] S1-3. Perform data conversion on the obtained data. Wind direction is angular data, usually expressed in degrees (0° to 360°). In order to make the wind direction data more meaningful in the mathematical model, the wind direction is converted from an angle to sine and cosine values, which can maintain its cyclic nature and avoid mutations. For example, 359 degrees and 1 degree are very different in value, but in fact they point to almost the same direction. Using only a single angle value may cause the model to misunderstand this close relationship. By using sine and cosine conversion, the wind direction data can correctly express its periodicity mathematically, ensuring that, for example, the north wind (0° or 360°) and the direction close to the north wind (such as 1° or 359°) are appropriately represented as similar in the model. Therefore, conversion is required, and the formula is as follows:

[0139]

[0140]

[0141] Where "Wind Direction" is the wind direction angle in degrees. The sine and cosine functions provide a way to decompose an angle into two orthogonal components (on the unit circle). Wind_Cos represents the projection of the angle onto the horizontal axis, reflecting the east-west component. Wind_sin represents the projection of the angle onto the vertical axis, reflecting the north-south component.

[0142] S1-4. Normalize the data. Normalization is the process of adjusting the scale of the data so that it is within a small, specific range (between 0 and 1). This helps improve algorithm performance and computational efficiency. The normalization method used here introduces a scaling factor to further control the scaling range and center point of the data to adjust the final distribution of the data. The formula is:

[0143]

[0144] Among them, x[n] is the original data, x normalized [n] is the normalized data, min(x) is the minimum value in the data, max(x) is the maximum value in the data, and α and β are adjustable scaling factors and offsets.

[0145] S1-5. Longitude and latitude are strong geographic indicators in themselves, but directly inputting them as features into the model is not the most effective approach. Spatial data (latitude and longitude) needs to be converted into a format that is more useful for machine learning models. This method performs sine and cosine transformations on the longitude and latitude coordinates of wind turbines to help the model capture the continuity and periodicity of geographic locations and avoid discontinuities caused by sudden changes, such as from 180 degrees to -180 degrees. First, convert degrees to radians. Longitude and latitude are usually in degrees and need to be converted to radians for trigonometric calculations. The formula is as follows:

[0146]

[0147] Among them, Radians Lat、Long Degrees are the radians corresponding to latitude and longitude respectively; Lat、Long are latitude and longitude respectively; π is pi.

[0148] Then calculate the sine and cosine values ​​of the latitude to represent the north-south position information. The formula is as follows:

[0149] Latitude cos =cos(Radians Lat )

[0150] Latitude sin =sin(Radians Lat )

[0151] Then calculate the sine and cosine values ​​of the longitude to represent the position information in the east-west direction. The formula is as follows:

[0152] Longitude cos =cos(Radians Long )

[0153] Longitude sin =sin(Radians Long )

[0154] Among them, Latitude cos is the cosine value of the dimension; Latitude sin is the sine value of the dimension; Longitude cos is the cosine value of longitude; sin is the sine of longitude.

[0155] S1-6. Arrange all data into a time series format so that it can be input into the one-dimensional feature extraction module.

[0156] S1-7. The multi-channel image spatial feature extraction module requires a preprocessing process that maps the wind turbine's wind speed, wind direction, air pressure, and power generation data to a two-dimensional grid and converts it into a multi-channel image. First, an M×N grid is defined based on the generator's spatial coverage and actual distribution accuracy. The geographic coordinates of each wind turbine are mapped to the grid coordinates using the following formula:

[0157]

[0158] Where x and y are the grid coordinates of the generator; N and M are the number of grid columns and rows, respectively, which are determined by the specific distribution and number of generators; longitude and latitude are the actual longitude and latitude of the power station; max_longitude and min_latitude are the maximum longitude and latitude of all generators, respectively; min_longitude and min_latitude are the minimum longitude and latitude of all generators, respectively.

[0159] S1-8. For each generator, use the x and y calculated above as grid coordinates and map the wind speed, wind direction, atmospheric pressure, and power generation data to the corresponding grid cell.

[0160] S1-9. Generate an image based on the data values ​​on the grid. The image has 5 channels. Each parameter (wind speed, sine and cosine of wind direction, air pressure, power generation) can be assigned to a channel of the image, similar to the RGB channels of a color image.

[0161] As an embodiment of the present application, the multi-channel image spatial feature extraction module in step S2 specifically includes:

[0162] S2-1-1. Input a multi-channel 2D image and perform an initial convolution using a 3×3 convolution kernel, followed by batch normalization and the LeakyReLU activation function to enhance the model's learning stability and nonlinear expression capabilities. Then, use a 1×1 convolution kernel for deep feature extraction, followed by batch normalization and the LeakyReLU activation function again.

[0163] S2-1-2. Introducing a dual attention mechanism of channel and space. First, channel attention is applied to weight certain channels, and then spatial attention is used to further refine the important areas of the feature map. The first is the channel attention mechanism, which compresses the two-dimensional feature map of each channel into a single value through a squeezing operation. The formula is as follows:

[0164]

[0165] where z c represents the global feature of the cth channel, which is the global average of all features of the channel; u c is the feature map of the cth channel, u c (i, j) is the value of the feature map at position (i, j); H and W are the height and width of the feature map; ∑ is the summation symbol, which is used to accumulate the value of the entire feature map; is the normalization factor for the averaging operation.

[0166] Then comes the excitation operation, which uses a small fully connected neural network to learn the weight of each channel to strengthen the features of important channels. The formula is as follows:

[0167] s c =σ(W2δ(W1z c ))

[0168] where s c represents the weight vector of the cth channel; z is the global feature vector from the squeeze operation, which is input to the excitation network; W1 and W2 represent the weight matrices of the two fully connected layers, the first fully connected layer is used for dimensionality reduction, and the second fully connected layer is used to restore the dimension; δ is the ReLU activation function, which is used to add nonlinearity; σ is the Sigmoid activation function, which is used to limit the weight between 0 and 1.

[0169] The last step is the recalibration operation, which is used to multiply the calculated channel weights by the original feature map channel by channel to emphasize the features of important channels. The formula is:

[0170]

[0171] where s c is the weight of channel c, is the weighted feature map.

[0172] Then comes the spatial attention mechanism, which requires generating a spatial attention map. Specifically, a 3×3 convolution kernel is used to aggregate features across channels, and then the Sigmoid function is used to obtain the importance of each position. The formula is as follows:

[0173]

[0174] Among them A s The generated spatial attention map is a two-dimensional matrix of the same size as the input feature map, where the value of each element is between 0 and 1, representing the importance weight of the corresponding position; Conv 3×3 is a 3×3 convolution kernel used to aggregate input feature maps across channels The convolution operation does not change the spatial dimension of the feature map because padding is used to keep the dimension unchanged. σ is the Sigmoid function, which is used to normalize the convolution output to the range of [0, 1] to ensure that the result can be used as a multiplication weight. It is the feature map processed by the channel attention mechanism.

[0175] Finally, the spatial attention mechanism is applied to the feature map after channel adjustment to enhance the features of important spatial regions. The formula is as follows:

[0176]

[0177] Represents the feature map after spatial attention adjustment, which is the data to be passed to the next processing stage; ⊙ is the element multiplication operation, representing A s Each spatial weight in This step is performed element by element, which means that each pixel is weighted by its corresponding spatial attention weight to emphasize the more important parts of the feature map; The adjusted feature map output from the channel attention mechanism has been preliminarily adjusted by the channel weights, and this step further refines it by the spatial weights.

[0178] S2-1-3. A global average pooling layer is added after the attention mechanism to reduce parameters and prepare for classification or regression tasks. A dropout layer is then introduced to prevent overfitting. All features are then integrated through a fully connected layer. Finally, an adaptive parameterized ReLU activation function (APReLU) is applied for final activation. The parameters (slope and intercept) of this function are dynamically adjusted for each feature (treated as an independent channel) to adapt to the characteristics of different features, and the final feature vector is output. The adaptive parameterized ReLU (APReLU) activation function is expressed as follows:

[0179]

[0180] Where x is the input of the activation function, and α, β, γ, and δ are learnable parameters that can be dynamically adjusted according to the input features during training.

[0181] As an embodiment of the present application, the one-dimensional information feature extraction module in step S2 specifically includes:

[0182] S2-2-1. Input the preprocessed time series data. First, convolve it with a 3×1 convolution kernel to capture short-term dependencies. Then, pass it through the ReLU activation function to increase nonlinearity to help the model learn more complex data patterns. Then, use a 5×1 convolution kernel for a second convolution to capture longer-term time dependencies.

[0183] S2-2-2. Apply a 2×2 average pooling operation after convolution to reduce the temporal dimension of the data while retaining important information; then apply a batch normalization layer to stabilize network training, speed up convergence, and reduce internal covariate shift.

[0184] S2-2-3. Introduce a temporal attention mechanism to enhance the model's focus on key time points. Specifically, a small convolutional network is used to generate temporal attention weights, which are implemented through a simple 1x1 convolutional layer followed by a Sigmoid activation function. The size of the attention weight is the same as the feature map, and the formula is as follows:

[0185] A t =σ(Conv 1×1 (F conv2 ))

[0186] Among them A t is the temporal attention map; F conv2 It is the feature map that has been convolved for the second time; Conv 1×1 is a small convolutional layer that generates temporal attention weights; σ is the Sigmoid activation function.

[0187] Then the attention weight is applied, the formula is as follows:

[0188]

[0189] Represents the feature map after temporal attention adjustment; ⊙ is the element-wise multiplication of the temporal attention weight and the feature map to strengthen the model's attention to important time segments.

[0190] S2-2-4. After applying the temporal attention mechanism, the results Input to the LSTM layer. LSTM can effectively retain information in long time series, which is very important for prediction. However, in time series analysis, many important events or changes are often short-lived and discontinuous. LSTM can not only capture these discontinuous patterns, but also pre-adjust the feature maps through the attention mechanism. The data provided has already emphasized these key time points, allowing LSTM to focus more on these parts that may contain important events or state changes, thereby improving the accuracy and timeliness of predictions.

[0191] S2-2-5. After the LSTM, the output is globally average pooled to compress each feature dimension of the time series into a single value, simplifying subsequent processing. The ELU activation function is then applied. The ELU is particularly useful when processing time series data. Due to its exponential form, the ELU maintains a certain gradient when handling negative values. This makes it more effective than ReLU and other variants when training deep networks, helping to accelerate learning and increase network robustness. A Dropout layer is then applied to further prevent overfitting and enhance the model's generalization capabilities. Finally, a fully connected layer is applied to integrate the learned features through one or more fully connected layers and output the final prediction result.

[0192] As an embodiment of the present application, the feature fusion module in step S2 specifically includes:

[0193] S2-3-1. Global average pooling is performed on the output features of both modules. Global average pooling maps multidimensional features to a one-dimensional vector, unifying the dimensions of the outputs of different modules and simplifying subsequent fusion steps. Whether it is two-dimensional image data or one-dimensional time series data, global average pooling provides a compact global feature description. Z-score is then used to normalize the features to ensure that the features from different sources are on the same order of magnitude, facilitating fusion.

[0194] S2-3-2. After flattening and normalizing the features, concatenate the outputs of the two modules to form a longer feature vector, as shown below:

[0195] F combined =[F 1D ; F 2D ]

[0196] Among them F combined represents the joint eigenvector after concatenation; F 1D and F 2D They represent the features output by the one-dimensional feature extraction module and the multi-channel image space feature extraction module respectively.

[0197] Because different features contribute differently to the final task, their importance should be reflected by different weights. To this end, this module introduces a weighted feature fusion method. Weighted fusion allows the model to consider the relative importance of each feature when making decisions, thereby improving prediction accuracy and model generalization ability. The steps are as follows:

[0198] First, a fully connected layer is used from F combined The role of the fully connected layer here is to combine the features F extracted from different modalities combined This layer performs a linear transformation on the combined input features using the weight matrix W and the bias term b, and then uses the softmax function to ensure that the sum of the output weights is 1. This output can be directly used as the coefficient of each feature in the subsequent weighted sum. The formula is as follows:

[0199] w 1D ,w 2D =softmax(W·F combined +b)

[0200] where w 1D ,w 2D is the obtained weight; W is the weight matrix of the fully connected layer; b is the bias term; F combined represents the joint feature vector; the softmax activation function ensures that the sum of the output weights is 1.

[0201] Then perform weighted feature fusion, the formula is as follows:

[0202] F fused =w 1D ·F 1D +w 2D ·F 2D

[0203] Among them F fused represents the fused features; w 1D and w 2D is the weight obtained through training; F 1D and F 2D They represent the features output by the one-dimensional feature extraction module and the multi-channel image space feature extraction module respectively.

[0204] This weighted fusion approach enables the model to automatically adjust the contribution of different features to the final decision based on their importance. This is particularly important when processing multimodal data, as different modalities may contribute differently to the final task.

[0205] S2-3-3. After feature fusion, the first fully connected layer is applied using the ReLU activation function to increase the model's expressiveness. A dropout layer is then added to reduce overfitting. A second fully connected layer is then applied using the ReLU activation function. This step uses two fully connected layers to further process the fused features, helping the model capture the complex relationships between different feature combinations. Finally, the predicted value is output.

[0206] As an embodiment of the present application, the loss function designed in step 3 is designed as follows:

[0207] S3-1. To improve the accuracy of wind power forecasting and optimize the network's performance when processing different data modalities, a loss function is proposed that combines prediction error, modal weight adjustment, and regularization. This function considers both prediction accuracy and feature coordination, improving the model's adaptability to complex data. The details are as follows:

[0208] S3-1-1. The prediction error loss part measures the difference between the predicted value and the actual value through the mean square error. This is a common loss function form in regression tasks. The formula is as follows:

[0209]

[0210] where y i is the actual power value of sample i, is the power value predicted by the model, and N is the total number of samples.

[0211] S3-1-2. In order to promote information coordination between different modalities (image spatial features and time series features), a modality weight adjustment loss function based on the internal correlation of the two features is designed. The formula is as follows:

[0212]

[0213] Among them, F 1D,j and F 2D,j They represent the extraction of one-dimensional and image features on the j-th sample; corr represents the correlation coefficient, which is used to measure F 2D,j ,F 1D,j A linear relationship exists between them; M is the number of modes.

[0214] Specifically, if F 2D,j and F 1D,j If the two features are perfectly correlated, the correlation coefficient is 1, and the modal coordination loss component of the loss function is 0, indicating that the two features are fully informationally coordinated. If the two features are uncorrelated or negatively correlated, the loss increases, prompting the model to adjust parameters to enhance the correlation between the two modal features. In this way, the modal coordination loss helps the model learn how to better integrate information from different data sources, improving the performance of prediction tasks.

[0215] S3-1-3. Introduce the L2 regularization term to control model complexity and prevent overfitting. The formula is as follows:

[0216]

[0217] where w k The weights of each layer of the model are learned during training using the Adam optimization algorithm. λ is the regularization coefficient, which controls the strength of regularization. Increasing λ increases the penalty for weights, forcing them toward smaller values, which helps prevent overfitting.

[0218] S3-1-4. The final loss function combining the three parts is expressed as follows:

[0219] L=L pred +αL modal +L reg

[0220] Here, α is an adjustable hyperparameter used to balance the importance of prediction error loss and modality coordination loss.

[0221] Specifically, the loss function is designed to promote the model to learn how to integrate information from different data sources and improve the consistency and accuracy of the overall prediction.

[0222] As an embodiment of the present application, the real-time prediction and power grid dispatch optimization in step S4 specifically include:

[0223] Before making real-time forecasts, S4-1 ensures that all key real-time data sources are instantly accessible and processed by the system. These data sources include wind speed, direction, air pressure, temperature, and real-time power generation data for each wind turbine, as well as the precise geographic location (latitude and longitude) of each turbine.

[0224] As with the training phase, all real-time data in S4-2 must undergo a standardized preprocessing process to maintain data consistency and ensure model accuracy. This preprocessing step includes data cleaning (removing errors and outliers), data transformation (converting wind direction to sine and cosine values), and appropriate feature extraction.

[0225] The S4-3 uses pre-processed real-time data to generate predictions using a trained multimodal network model. Leveraging an integrated attention mechanism and advanced deep learning techniques, the model outputs hourly forecasted power generation for the next seven days, providing accurate data support for grid dispatch.

[0226] The wind power data predicted by S4-4 will be transmitted to the power grid dispatching system in real time. The system will dynamically adjust the power generation resource allocation based on this data and the current load demand of the power grid. Specific optimization measures include:

[0227] S4-4-1 Power generation efficiency optimization: Adjust the power generation output of each wind farm based on the forecast results of wind power generation, optimize the operating status of the generator sets and the power generation scheduling plan, so as to maximize the wind power utilization efficiency and overall power generation efficiency.

[0228] S4-4-2 Improved power supply reliability: By using accurate forecasting data, the system can identify potential power shortages or surpluses in advance, allowing preventive measures to be taken, such as starting or shutting down auxiliary power generation facilities, to reduce the risk of power outages due to unstable wind power.

[0229] S4-4-3 optimizes the power generation mix, reducing overall power generation costs by reducing reliance on expensive peak-shaving power plants. At the same time, by precisely dispatching renewable energy, it reduces carbon emissions and environmental impact, achieving both economic and environmental benefits.

[0230] Through the above steps, the present invention not only improves the prediction accuracy of wind power generation, but also significantly improves the operating efficiency and reliability of the power grid through intelligent power grid scheduling, while reducing operating costs, demonstrating its practical application value in modern power grid management.

[0231] The basic principles, main features and advantages of the present invention are shown and described above. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention.

[0232] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A power grid dispatch optimization method based on deep learning, characterized by: The steps include: S1. Collect historical power output data and spatiotemporal data of each wind turbine in the wind power plant and pre-process the data; S2. Design a multimodal network structure suitable for wind power generation prediction, including a one-dimensional information feature extraction module, a multi-channel image space feature extraction module, and a feature fusion module; the multimodal network structure in S2 specifically includes: S2-1. Design a multi-channel image spatial feature extraction module as follows: S2-1-1. Input a multi-channel 2D image, perform initial convolution using a 3×3 convolution kernel, followed by batch normalization and LeakyReLU activation function, then use a 1×1 convolution kernel for deep feature extraction, and then go through batch normalization and LeakyReLU activation function again; S2-1-2. Introduce the channel and spatial dual attention mechanism. The first is the channel attention mechanism. Specifically, through the squeezing operation, the two-dimensional feature map of each channel is compressed into a single value. The formula is as follows: ; in represents the global feature of the c-th channel, which is the global average of all features of the channel; is the feature map of the c-th channel, The feature map is at position The value of; and is the height and width of the feature map; ∑ is the summation symbol, which is used to accumulate the value of the entire feature map; is the normalization factor for the averaging operation; Then comes the excitation operation, which uses a small fully connected neural network to learn the weight of each channel to strengthen the features of important channels. The formula is as follows: ; in represents the weight vector of the c-th channel; is the global feature vector from the squeeze operation, which is input to the excitation network; and Represents the weight matrix of two fully connected layers, the first fully connected layer is used for dimensionality reduction, and the second fully connected layer is used for dimensionality recovery; is the ReLU activation function, used to increase nonlinearity; Sigmoid activation function is used to limit the weight between 0 and 1; The last step is the recalibration operation, which is used to multiply the calculated channel weights by the original feature map channel by channel to emphasize the features of important channels. The formula is: ; in is the weight of channel c, is the weighted feature map; Then comes the spatial attention mechanism, the formula is as follows: ; in is the generated spatial attention map; It is a 3×3 convolution kernel used to aggregate features across channels; is the Sigmoid function; It is the feature map processed by the channel attention mechanism; Finally, the spatial attention mechanism is applied, and the formula is as follows: ; Represents the feature map after spatial attention adjustment; is the element-wise multiplication operation, indicating Each spatial weight in Multiply the values ​​at corresponding positions in ; S2-1-3. Add a global average pooling layer, a dropout layer, and a fully connected layer after the attention mechanism, apply an adaptive activation function, and output the final feature vector; the activation function is expressed as follows: ; in is the input of the activation function, are learnable parameters that can be dynamically adjusted according to input features during training S3. Design a loss function to train the multimodal network and obtain a trained network model; S4. Use the trained model to make real-time predictions of wind power generation to optimize grid scheduling.

2. A power grid dispatch optimization method based on deep learning according to claim 1, characterized in that: Step S1 specifically includes the following contents: S1-1. Obtain the power generation data, wind speed, wind direction, air pressure and latitude and longitude information of each wind turbine in the wind power plant every 4 hours within one year; S1-2. Clean the acquired data. First, use the weighted linear interpolation method to process the missing values ​​of wind speed and wind direction. The formula is: ; ; in, is the predicted value for the missing time point; and are the known values ​​before and after the missing point respectively; and is the weight, calculated based on time distance; , , is the corresponding time point; Then the IQR rule is used to deal with outliers in wind speed. The formula is: ; ; ; Where Q1 is the lower quartile (25%) of all values ​​in the data set; Q3 is the upper quartile (75%) of all values ​​in the data set; IQR is the interquartile range; k is an adjustable parameter that is set according to the sensitivity of the data distribution; Upper Bound and Lower Bound are the upper and lower bounds respectively, and any value below the lower bound or above the upper bound needs to be eliminated; S1-3. Perform data conversion on the acquired data, converting the wind direction from an angle (0-360 degrees) into sine and cosine values. The formula is as follows: ; ; Where "Wind Direction" is the wind direction angle in degrees; S1-4. Remove the proportional effect in the data by normalization. The formula is: ; in, is the original data, is the normalized data, is the minimum value in the data, is the maximum value in the data, and are adjustable scaling factors and offsets; S1-5. Perform sine and cosine transformations on the longitude and latitude coordinates of the wind turbine. First, convert the degrees to radians. The formula is as follows: ; in, are the arc degrees corresponding to latitude and longitude respectively; are latitude and longitude respectively; π is pi; Then calculate the sine and cosine values, the formula is as follows: ; ; ; ; in, is the cosine of the dimension; is the sine of the dimension; is the cosine of longitude; is the sine of longitude; S1-6. Arrange all data in a time series format so that it can be input into the one-dimensional feature extraction module.

3. A power grid dispatch optimization method based on deep learning according to claim 2, characterized in that: Said S1 further comprises: S1-7. Define an M×N grid and map the geographic coordinates of each wind turbine to the grid coordinates using the following formula: ; ; Where x and y are the grid coordinates of the generator; N and M are the number of columns and rows in the grid, respectively, which are determined by the specific distribution and number of generators; longitude and latitude are the actual longitude and latitude of the power station; max_longitude and max_latitude are the maximum longitude and latitude of all generators, respectively; min_longitude and min_latitude are the minimum longitude and latitude of all generators, respectively. S1-8. For each generator, use the x and y calculated above as grid coordinates and map the wind speed, wind direction, atmospheric pressure, and power generation data to the corresponding grid cell; S1-9. Generate an image based on the data values ​​on the grid. The image has 5 channels, which are respectively allocated by the sine and cosine of wind speed, wind direction, atmospheric pressure, and power generation.

4. The power grid dispatch optimization method based on deep learning according to claim 1, characterized in that: The S2 also includes the following contents: S2-2. Design a one-dimensional information feature extraction module, as follows: S2-2-1. Input the preprocessed time series data and first perform convolution with a 3×1 convolution kernel to capture short-term dependencies. Then, after passing the ReLU activation function, convolution with a 5×1 convolution kernel is performed again to capture longer-term dependencies. S2-2-2. Apply 2×2 average pooling after convolution, and then apply batch normalization layer; S2-2-3. Introduce the temporal attention mechanism and use a small convolutional network to generate temporal attention weights. The formula is as follows: ; in is the temporal attention map; It is the feature map after the second convolution; A small convolutional layer to generate temporal attention weights; is the Sigmoid activation function; Then the attention weight is applied, the formula is as follows: ; Represents the feature map after temporal attention adjustment; It is an element-wise multiplication operation; S2-2-4. After applying the temporal attention mechanism, the results Input to the LSTM layer to capture discontinuous patterns in the time series; S2-2-5. After LSTM, perform global average pooling on the output, then apply the ELU activation function, and finally apply the Dropout layer and the fully connected layer; S2-3. Design a multi-channel image spatial feature extraction module and perform feature fusion on the outputs of the one-dimensional information feature extraction module and the multi-channel image spatial feature extraction module. The details are as follows: S2-3-1. Perform global average pooling on the output features of the two modules to ensure the same dimensions, and then use Z-score to normalize the features; S2-3-2. Concatenate the processed one-dimensional and two-dimensional feature vectors. The formula is as follows: ; in represents the joint eigenvector after concatenation; and Represent the features output by the one-dimensional feature extraction module and the multi-channel image space feature extraction module respectively; Then use a fully connected layer from The feature weights are learned in the following formula: ; in is the obtained weight; is the weight matrix of the fully connected layer; is the bias term; represents the joint feature vector; the softmax activation function ensures that the sum of the output weights is 1; Then perform weighted feature fusion, the formula is as follows: ; in Represents the fused features; and is the weight obtained through training; and Represent the features output by the one-dimensional feature extraction module and the multi-channel image space feature extraction module respectively; S2-3-3. After feature fusion is completed, the first fully connected layer is applied, using the ReLU activation function; then the Dropout layer is added; then the second fully connected layer is applied, using the ReLU activation function; finally, the predicted value is output.

5. The power grid dispatch optimization method based on deep learning according to claim 1, characterized in that: The loss function designed in S3 is as follows: S3-1. The loss function consists of three parts: prediction error loss, modal coordination loss, and regularization term; S3-1-1. The prediction error loss part measures the difference between the predicted value and the actual value through the mean square error. The formula is as follows: ; in It is a sample The actual power value, is the power value predicted by the model, and N is the total number of samples; S3-1-2, modality coordination loss aims to optimize the coordination between different modality features (image and time series data), and the formula is as follows: ; in, and Respectively, they represent the extraction of one-dimensional and two-dimensional features on the j-th sample; corr represents the correlation coefficient; M is the number of modes; S3-1-3, the regularization term is used to control the complexity of the model and prevent overfitting. The formula is as follows: ; in represents the weights of each layer of the model, is the regularization coefficient.

6. The power grid dispatch optimization method based on deep learning according to claim 5, characterized in that: The final loss function of the three parts is expressed as follows: ; in, is an adjustable hyperparameter used to balance the importance of prediction error loss and modality coordination loss.

7. The power grid dispatch optimization method based on deep learning according to claim 1, characterized in that: The detailed steps for performing real-time prediction and power grid dispatch optimization in S4 are as follows: S4-1. Before conducting real-time forecasts, it is necessary to ensure that all relevant real-time data sources are ready and can be accessed in real time. This data includes: wind speed and direction data, air pressure and temperature data, power generation data, and latitude and longitude data for each generator; S4-2. As in the training phase, all real-time data must undergo the same preprocessing steps to ensure data consistency and model effectiveness; this includes data cleaning, data transformation, and feature extraction. S4-3. Use the pre-processed data to perform real-time predictions using the trained multimodal network model; the model will output the predicted power generation for the next 7 days. S4-4. The predicted wind power generation data will be directly input into the power grid dispatching system; the system will dynamically adjust the power generation resource allocation in the power grid based on the prediction results and the current load demand of the power grid.

8. The power grid dispatch optimization method based on deep learning according to claim 7, characterized in that: The following aspects are optimized in S4-4: S4-4-1. Power generation efficiency: Adjust the output of the wind farm based on the forecast results of wind power generation to improve power generation efficiency; S4-4-2. Power supply reliability: Use forecast data to reduce the risk of power outages caused by unstable wind power; S4-4-3. Cost control: Optimize the power supply combination, reduce the use of expensive peak-shaving power stations, and reduce power generation costs.

Images