Wind speed and temperature prediction method based on deep learning technology
Through the wind speed and temperature prediction method based on deep learning, the adaptive attribute interactive graph neural network and conditional variational autoencoder are used to solve the problem of unconsidered data interaction in traditional meteorological prediction, and achieve higher prediction accuracy and Interpretability.
Patent Information
- Application Number
- CN202410982897.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-08
AI Technical Summary
Traditional meteorological prediction methods are difficult to achieve high accuracy and real-time performance, especially because they fail to effectively consider the mutual influence of different meteorological data, resulting in low accuracy of prediction results.
Using the wind speed and temperature prediction method based on deep learning technology, an adaptive attribute interactive graph neural network is constructed by acquiring and filling the wind speed and temperature data of the electrical meteorological platform, and the deep characteristics of wind speed and temperature are captured using the conditional variational autoencoder and Complex relationships and conduct end-to-end learning processes to generate prediction results.
It improves data quality, enhances the training basis and prediction accuracy of the model, improves the distinction and generalization ability of wind speed and temperature prediction, and provides higher interpretability.
Smart Images

Figure CN118916635A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of deep learning and big data technology, and in particular to a wind speed and temperature prediction method based on deep learning technology. Background Art
[0002] In today's era of rapid development of information technology, deep learning and big data technology have become key drivers of innovation in all walks of life. In particular, in the field of meteorology, accurate prediction of wind speed and temperature is of great significance to agriculture, aviation, energy management, urban planning, disaster prevention and many other aspects. With the development of deep learning technology, especially the emergence of models such as convolutional neural network (CNN), recurrent neural network (RNN) and long short-term memory network (LSTM), new perspectives have been provided for the processing and analysis of meteorological data. These models can learn complex nonlinear relationships from a large amount of historical meteorological data, thereby improving the accuracy and efficiency of predictions. In addition, the application of big data technology provides a rich data resource for meteorological forecasting. By collecting and integrating meteorological data from multiple sources such as ground stations, satellites, and drones, a more comprehensive and diverse data set can be constructed to provide support for the training and verification of deep learning models.
[0003] However, traditional meteorological forecasting methods, such as numerical weather forecasting based on physical models, are often limited by computing power and data collection limitations, making it difficult to achieve high accuracy and real-time performance. For example, the transformer temperature prediction model and method based on meteorological environment data with patent application number CN202311556068.3 discloses obtaining meteorological environment monitoring data, wherein the meteorological environment monitoring data includes ambient temperature, humidity, wind speed, and altitude; analyzing the transformer temperature based on changes in ambient temperature in the meteorological environment monitoring data, and predicting the transformer temperature at different ambient temperatures; analyzing the transformer temperature based on changes in humidity in the meteorological environment monitoring data, and predicting the transformer temperature at different humidity; analyzing the transformer temperature based on changes in wind speed in the meteorological environment monitoring data, and predicting the transformer temperature at different wind speeds; analyzing the transformer temperature based on changes in altitude in the meteorological environment monitoring data, and predicting the transformer temperature at different altitudes. The transformer temperature under different ambient temperatures, humidity, wind speeds, and altitudes is analyzed based on the meteorological environment monitoring data to obtain the transformer temperature prediction values under different ambient temperatures, humidity, wind speeds, and altitudes, and a transformer temperature prediction linear regression equation is established; the transformer temperature prediction compliance coefficient is evaluated using real data, and based on the calculated transformer temperature prediction compliance coefficient, the transformer temperature prediction linear regression equation is evaluated to see whether it fits the real data. If the transformer temperature prediction linear regression equation cannot fit the real data, the transformer temperature prediction linear regression equation is optimized; the prior art can predict the change trend of the transformer temperature based on the meteorological environment data, and compared with the large number of sensors deployed to monitor the temperature, the number of sensors and maintenance work can be reduced, and the cost and workload can be reduced. However, the above-mentioned comparative documents do not consider the mutual influence of different data in the collected data, resulting in low accuracy of the prediction results. Summary of the invention
[0004] In order to solve the problems existing in the above-mentioned prior art, the present invention proposes a wind speed and temperature prediction method based on deep learning technology, which includes: obtaining wind speed and temperature data of an electric meteorological platform, and filling missing values in the data; constructing an attribute set according to the filled data, and mapping the data in the attribute set; constructing an adaptive attribute interaction graph neural network; initializing the wind speed index and temperature index in the data, and marking them as Y1 and Y2 respectively; inputting the mapped attribute features, Y1 and Y2 into the adaptive attribute interaction graph neural network for interaction, and splicing the features after interaction to obtain a total feature vector; mapping the total feature vector to the latent semantic space corresponding to the label through a CVAE network to obtain a mapped feature space vector; inputting the mapped feature space vector into the trained DNN network to obtain the wind speed and temperature prediction results of the meteorological platform.
[0005] Beneficial effects of the present invention:
[0006] 1. The present invention aims at the problem of missing data values and uses BayImpute to fill in missing values, thereby improving data quality and strengthening the training basis of the model.
[0007] 2. The present invention aims at the interaction problem between different attributes and different attributes and labels. By introducing attribute interaction graph neural network and conditional variational autoencoder, the deep-level characteristics and complex relationships of wind speed and temperature prediction data can be captured.
[0008] 4. Aiming at the problem of feature expression under different indicators, the present invention takes temperature and wind speed as conditions and uses conditional variational autoencoder for prediction, which enhances the model's ability to distinguish and generalize these two related indicators.
[0009] 5. The method of the present invention is an end-to-end learning process, from data preprocessing to feature interaction extraction and then to prediction, which helps the model to learn and optimize better while providing higher interpretability. BRIEF DESCRIPTION OF THE DRAWINGS
[0010] Figure 1 This is a flow chart of a method for predicting wind speed and temperature based on deep learning technology of the present invention;
[0011] Figure 2 This is a model structure diagram of a wind speed and temperature prediction method based on deep learning technology of the present invention. DETAILED DESCRIPTION
[0012] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0013] A wind speed and temperature prediction method based on deep learning technology, such as Figure 1 and Figure 2As shown, the method includes: obtaining wind speed and temperature data of the meteorological platform, and filling the missing values of the data; constructing an attribute set according to the filled data, and mapping the data in the attribute set; constructing an adaptive attribute interaction graph neural network; initializing the wind speed index and temperature index in the data, and marking them as Y1 and Y2 respectively; inputting the mapped attribute features, Y1 and Y2 into the adaptive attribute interaction graph neural network for interaction, splicing the features after interaction to obtain a total feature vector; mapping the total feature vector to the latent semantic space corresponding to the label through the CVAE network to obtain a mapped feature space vector; inputting the mapped feature space vector into the trained DNN network to obtain wind speed and temperature prediction results.
[0014] In this embodiment, mapping the data in the attribute set includes using the BayImpute method to fill missing values in the data; the calculation formula of the BayImpute method is as follows:
[0015]
[0016] Among them, α is the prior parameter of the mean, μ0 is the mean of m values near the missing value, and w i is the weight parameter of these m values.
[0017] The continuous value attributes in the data are mapped through the linear layer, and the category attributes are transformed into dense features through Embedding. Each attribute is regarded as a node to build an attribute interaction graph neural network, which includes:
[0018] Step 1. Map the continuous value attributes through the linear layer, map the categorical attributes through Embedding, and then combine them together. The calculation formula for the above process is as follows:
[0019] h j =W j ·x j , for j∈R
[0020] h i =Embedding(x i ), for i∈C
[0021] The index set of the categorical feature attributes is C, and the index set of the continuous feature attributes is R. i is the value of the i-th category feature, W j is the weight vector of the jth continuous feature, x j is the value of the jth continuous feature attribute. Then all the continuous value features and categorical value features are concatenated together. Assuming there are N attributes and the mapping dimension is H, a vector representation of the shape [N,H] can be obtained.
[0022] Step 2. Consider each of these N attributes as a node and construct an attribute interaction graph neural network. The formula is as follows:
[0023]
[0024] where h i ′ is the updated feature representation of node i, σ is the activation function, and d i is the degree of node i, d j is the degree of node j, A ij It is an element of the adjacency matrix of the graph, indicating whether there is an edge between node i and node j. Here is a fully connected graph neural network, so A is an adjacency matrix of all 1s, W is a weight matrix, and X j is the feature representation of node j.
[0025] In this embodiment, the two indicators of wind speed and temperature are represented by Embedding initialization and recorded as Y1 and Y2 respectively, and each attribute feature is interacted with Y1 and Y2 using an adaptive deep attribute interaction network, specifically including the following steps:
[0026] Step 1: Use the embedding vector to calculate the similarity between attributes, which is implemented here through RMS (Regularized Mahalanobis Similarity) and calculate the adaptive weight matrix. The calculation formula is as follows:
[0027] For each attribute i, its embedding is represented as x i ∈R H ; For two labels Y1, Y2, their embedding representations are y1∈R H ,y2∈R H ; Combine the embedding representations of attributes and labels to get the matrix X∈R (N+2)×H ; Here N means there are N attributes, and H means the embedding dimension of Embedding.
[0028] Calculate the similarity between attributes and labels as follows:
[0029]
[0030] S shrinked =(1-γ)S+γI
[0031]
[0032] Where μ is the mean vector, S is the covariance matrix of X, and S shrinked is the regularized covariance matrix, is its inverse matrix, and M is a (N+2)×(N+2) similarity matrix, which represents the similarity between all attributes. The stability and robustness of the calculation are improved by regularizing the covariance matrix.
[0033] Calculate the adaptive weight matrix W A :
[0034]
[0035] Step 2: Introduce an adaptive deep attribute interaction network to enable each attribute feature to interact deeply with other attribute features and indicators, so that the model can learn richer and more relevant feature representations. The calculation process is as follows:
[0036] Calculate the attention score through Historical Weight-Adjusted Attention:
[0037] Q=XW Q , K = XW k , V = XW v
[0038] f(Q,K)=LeakyReLu(QW1+KW2+b)W3+γ*History(W1,W2,W3)
[0039]
[0040] Among them, W1, W2, W3, W Q ,W K ,W V ∈R H×H is the weight matrix, W4∈R H×N is the weight matrix.
[0041] Introduce the adaptive weight matrix and update the attention weight:
[0042] A ′ =A+W A
[0043] Among them A ′ is the adjusted attention weight.
[0044] Stack multiple such adaptive attention layers:
[0045] Z l =AdaptAttention(Z l-1 )
[0046] After each self-attention layer, skip connections and layer normalization are introduced:
[0047] Zl =LayerNorm(Z l-1 +AdaptAttention(Z l-1 ))
[0048] Step 3: The weighted representations of all attributes are fused into a comprehensive feature vector for subsequent latent semantic space vector mapping. The calculation formula is as follows:
[0049] Z E =[E1; E2; ...; E N ]
[0050] Z d =Linear(Z E )
[0051] where E1, E2, …, E N Represents the vector representation of each attribute that combines attribute and label features, Z d Indicates that Z E After linear layer mapping, the final fusion feature vector representation is obtained, Z d ∈R H .
[0052] In this embodiment, Conditional Variational Autoencoder (CVAE) is used to map the feature space to the latent semantic space corresponding to the label under the condition that the two indicators Y1 and Y2 are respectively used as conditions; for the mapped feature space vectors Z1 and Z2, DNN is used to predict the results of the two corresponding indicators of inflow wind speed and temperature, including the following steps:
[0053] Step 1: Use the conditional encoder to encode the input features and labels Y1, Y2 to obtain the mean distribution μ and standard deviation distribution σ in the latent space. The calculation process is as follows:
[0054] First, define Weighted Layer Normalization for normalization, and its calculation formula is as follows:
[0055]
[0056] where W∈R d×d is the weight parameter, H∈R n×d is a matrix that needs to be normalized, β is a d-dimensional bias vector, and W and β are both learnable parameters that are optimized by back propagation during training.
[0057] The MLP network is defined as follows:
[0058] h1=Relu(XW1+b1)
[0059]
[0060] h4=WLN(β·h2+γ·h3+X)
[0061] where X∈R n×d is a matrix put in, W1∈R d×h The weight matrix, b1 is an h-dimensional bias vector, W2, W2∈R h×n The weight matrix is , b2, b3 are d-dimensional bias vectors, and β, γ are learnable parameters. The MLP network enhances the nonlinear modeling ability, generalization ability, and expression ability of the model by stacking linear layers, using LeakyReLU activation, introducing weight regularization, applying weighted summation, and possible residual connections, making it suitable for processing complex data relationships.
[0062] For Y1:
[0063] μ1=MLP([Z d ; E y1 ])
[0064] σ1=MLP([Z d ; E y1 ])
[0065] For Y2:
[0066] μ2=MLP([Z d ; E y2 ])
[0067] σ2=MLP([Z d ; E y2 ])
[0068] Step 2: Sample random noise ∈1, ∈2 from the standard normal distribution, apply the sampling process, and generate latent variables X1, X2. These variables reflect the data characteristics under the given label. The calculation process is as follows:
[0069] X1=u1+∈1·exp(σ1)
[0070] X2=u2+∈2·exp(σ2)
[0071] Where ∈1, ∈2 are sampled from the standard normal distribution, μ1, μ1 are the mean distributions in the corresponding label latent semantic space, and σ1, σ2 correspond to the variance distributions in the label latent semantic space.
[0072] Step 3: Use the conditional prediction network to decode and predict the latent variables X1 and X2 to obtain the predicted values of labels Y1 and Y2 The calculation process is as follows:
[0073] The DNN network formula is as follows:
[0074]
[0075] Step 4: Calculate the loss function and optimize the model based on the loss.
[0076] The loss function formula here is:
[0077]
[0078]
[0079] Among them, β, δ, σ are hyper parameters, Y 1,i , Y 2,i is the label, Q(Z|X) is the posterior distribution of the latent variable Z given X, and P(Z) is the prior distribution of the latent variable, which is used to measure the difference between the distribution of the latent variable Z generated by the encoder and the prior distribution P(Z).
[0080] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A wind speed and temperature prediction method based on deep learning technology, characterized in that: include: The wind speed and temperature data of the electric meteorological platform are obtained, and the missing values are filled in the data; an attribute set is constructed according to the filled data, and the data in the attribute set is mapped; an adaptive attribute interaction graph neural network is constructed; the wind speed index and temperature index in the data are initialized and marked as Y1 and Y2 respectively; the mapped attribute features, Y1 and Y2 are input into the adaptive attribute interaction graph neural network for interaction, and the features after interaction are concatenated to obtain the total feature vector; the total feature vector is mapped to the latent semantic space corresponding to the label through the CVAE network to obtain the mapped feature space vector; the mapped feature space vector is input into the trained DNN network to obtain the wind speed and temperature prediction results of the meteorological platform.
2. The method for predicting wind speed and temperature based on deep learning technology according to claim 1, characterized in that: The missing value filling process includes: Among them, α is the prior parameter of the mean, μ0 is the mean of m values near the missing value, and w i is the weight parameter of these m values.
3. The method for predicting wind speed and temperature of a meteorological platform according to claim 1, characterized in that: Mapping the data in the attribute set includes mapping the continuous value attributes in the attribute set through the linear layer and mapping the category attributes through Embedding; Its expression is: h j =W j ·x j ,forj∈R h i =Embedding(x i ),fori∈C Among them, C is the index set of category feature attributes, R is the index set of continuous feature attributes, and x i is the value of the i-th category feature, W j is the weight vector of the jth continuous feature, x j is the value of the jth continuous feature attribute.
4. The method for predicting wind speed and temperature based on deep learning technology according to claim 1, characterized in that: The adaptive attribute interaction graph neural network is: Among them, h i ′ is the updated feature representation of node i, σ is the activation function, and d i is the degree of node i, d j is the degree of node j, A ij is the element of the graph adjacency matrix, W is the weight matrix, X j is the feature representation of node j.
5. The method for predicting wind speed and temperature based on deep learning technology according to claim 1, characterized in that: Adopting adaptive attribute interaction graph neural network to interact with input data includes: Step 1: Use the embedding vector to calculate the similarity between attributes, and calculate the adaptive weight matrix through the similarity between attributes; Step 2: Use an adaptive deep attribute interaction network to perform deep information interaction between each attribute feature and other attribute features and indicators to obtain the interacted features; Step 3: Use the attention mechanism to perform weighted fusion on the interacted features according to the adaptive weight matrix to obtain the total feature vector.
6. A wind speed and temperature prediction method based on deep learning technology according to claim 5, characterized in that: The similarity between the attributes and labels is calculated as: S shrinked =(1-γ)S+γI Where μ is the mean vector, S is the covariance matrix of X, and S shrinked is the regularized covariance matrix, is its inverse matrix, and M is a (N+2)×(N+2) similarity matrix.
7. The method for predicting wind speed and temperature based on deep learning technology according to claim 5, characterized in that: The attention mechanism is used to perform weighted fusion of the interacted features, including: Step 1: Calculate the attention score through Historical Weight-Adjusted Attention; its expression is: Q=XW Q ,K=XW k ,V=XW v f(Q,K)=LeakyReLu(QW1+KW2+b)W3+γ*History(W1,W2,W3) Among them, W1, W2, W3, W Q ,W K ,W V ∈R H×H is the weight matrix, W4∈R H×N is the weight matrix; Step 2: Update the attention weight according to the adaptive weight; Step 3: Use the updated attention weights to calculate the attention features of the input attributes, and use skip connections and layer normalization to connect the attention features; Step 4: Fusion the weighted representations of all attributes into a comprehensive feature vector, which is expressed as: Z E =[E1;E2;…;E N ] WITH d =Linear(Z E ) Among them, E1, E2, …, E N Represents the vector representation of each attribute that combines attribute and label features, Z d Indicates that Z E After linear layer mapping, the final fused feature vector representation is obtained.
8. The method for predicting wind speed and temperature based on deep learning technology according to claim 1, characterized in that: The CVAE network processes input data by: using a conditional encoder to encode the input features and labels Y1, Y2 to obtain the mean distribution μ and standard deviation distribution σ in the latent space; sampling random noise ∈1, ∈2 from a standard normal distribution; sampling the random noise to generate latent variables X1, X2; where the latent variables are the mapped feature space vectors.
9. The method for predicting wind speed and temperature of a meteorological platform according to claim 8, characterized in that: The formula for generating latent variables X1 and X2 is: X1=u1+∈1·exp(σ1) X2=u2+∈2·exp(σ2) Among them, ∈1, ∈2 are parameters sampled from the standard normal distribution, μ1, μ2 are the mean distributions in the corresponding label latent semantic space, and σ1, σ2 correspond to the variance distributions in the label latent semantic space.
10. The method for predicting wind speed and temperature based on deep learning technology according to claim 1, characterized in that: The loss function for training the DNN network is: Among them, β, δ, σ are hyper parameters, Y 1,i , Y 1,i is the label, Q(Z|X) is the posterior distribution of the latent variable Z given X, and P(Z) is the prior distribution of the latent variable.
Citation Information
Patent Citations
High-temperature disaster forecasting method based on directed graph neural network
CN114626512A
Space time sequence prediction method based on loop graph operator neural network
CN116610919A
Multi-fan wind speed space-time prediction method and system
CN116894384A
New energy meteorological large model construction method and device and power prediction method
CN118332937A
High-temperature disaster forecast method based on directed graph neural network
US20230375745A1
