Substation microclimate prediction model training method and apparatus, and computer device

By combining a dynamic weighting module, a time-series feature extraction module, and a fully connected layer, along with a cross-optimization model, the problem of low accuracy in traditional weather forecasting methods in substations is solved, improving the accuracy and generalization ability of weather forecasts and ensuring the stability of the power system.

WO2026237338A1PCT designated stage Publication Date: 2026-11-19MAINTENANCE & TEST CENTRE CSG EHV POWER TRANSMISSION CO
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2025/140629
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-16
Filing Date
2025-12-08
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Traditional weather forecasting methods in substations suffer from limitations in data sources and shallow machine learning models. This leads to excessively large training parameters when dealing with high-dimensional inputs, making overfitting easy and failing to effectively uncover temporal patterns, resulting in low forecast accuracy.

Method used

A combination of a dynamic weighting module, a temporal feature extraction module, and a fully connected layer is adopted, along with a cross-optimization model. The dynamic weighting module assigns weights to different meteorological data, the temporal feature extraction module captures time dependencies, and the fully connected layer maps them to meteorological prediction results. The model is then iteratively trained using actual meteorological data to optimize it.

Benefits of technology

It improves the accuracy and generalization ability of meteorological forecasting models, enabling them to better capture long-term and short-term weather dependencies, generate accurate meteorological forecast results, and ensure the safe and stable operation of substations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2025140629_19112026_PF_FP_ABST
    Figure CN2025140629_19112026_PF_FP_ABST
Patent Text Reader

Abstract

The present application relates to a substation microclimate prediction model training method and apparatus, and a computer device. The method comprises: inputting historical meteorological sample data into an initial dynamic weighting module to obtain weighted meteorological sample data; inputting the weighted meteorological sample data into an initial time-series feature extraction module to obtain meteorological time-series feature data; inputting the meteorological time-series feature data into an initial fully connected layer to obtain meteorological prediction sample data; performing iterative training according to the meteorological prediction sample data and actual meteorological data, so as to obtain a trained target dynamic weighting module, a trained target time-series feature extraction module, and a trained candidate fully connected layer; iteratively training the candidate fully connected layer by using a horizontal-vertical crossover optimization model to obtain a trained target fully connected layer; and obtaining a trained substation microclimate prediction model according to the target dynamic weighting module, the target time-series feature extraction module, and the target fully connected layer. The method can improve meteorological prediction model accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Training methods, devices, and computer equipment for substation micro-meteorological prediction models Technical Field

[0001] This application relates to the field of meteorological forecasting technology, and in particular to a training method, apparatus, computer equipment, computer-readable storage medium, and computer program product for a substation micro-meteorological forecasting model. Background Technology

[0002] In power systems, the operation of substations is closely related to environmental meteorological conditions. Especially under extreme weather conditions, accurate weather forecasts are of great significance for ensuring the stable operation of power equipment.

[0003] Currently, traditional weather forecasting methods are limited by data sources and shallow machine learning forecasting models. When the input sample dimension is large and the number of neurons is increased, the number of training parameters becomes too large, and overfitting is prone to occur. This makes it difficult to discover time series patterns, ultimately resulting in low accuracy of the forecasting model. Summary of the Invention

[0004] Therefore, it is necessary to provide a training method, device, computer equipment, computer-readable storage medium, and computer program product for a substation micro-meteorological forecasting model that can improve the accuracy of meteorological forecasting models, in response to the above-mentioned technical problems.

[0005] Firstly, this application provides a training method for a substation micro-meteorological prediction model, including:

[0006] Historical meteorological sample data of the substation to be predicted is obtained, and the historical meteorological sample data is input into the initial dynamic weighting module to obtain the weighted meteorological sample data;

[0007] The weighted meteorological sample data is input into the initial time series feature extraction module to obtain meteorological time series feature data;

[0008] The meteorological time-series feature data is input into the initial fully connected layer to obtain meteorological prediction sample data;

[0009] Acquire actual meteorological data, and iteratively train the initial dynamic weighting module, the initial temporal feature extraction module, and the initial fully connected layer based on the meteorological prediction sample data and the actual meteorological data to obtain the trained target dynamic weighting module, target temporal feature extraction module, and candidate fully connected layer;

[0010] A cross-optimization model is adopted, based on the historical meteorological sample data, the actual meteorological data, the target dynamic weighting module and the target temporal feature extraction module, to iteratively train the candidate fully connected layer and obtain the trained target fully connected layer.

[0011] Based on the target dynamic weighting module, the target temporal feature extraction module, and the target fully connected layer, the trained substation micro-meteorological prediction model is obtained.

[0012] In one embodiment, the step of inputting the historical meteorological sample data into the initial dynamic weighting module to obtain weighted meteorological sample data includes:

[0013] The historical meteorological sample data is processed by channel feature weighting in the initial dynamic weighting module to obtain channel-weighted sample data.

[0014] The channel-weighted sample data is processed by spatial feature weighting in the initial dynamic weighting module to obtain spatially weighted sample data, which is used as the weighted meteorological sample data.

[0015] In one embodiment, the step of processing the historical meteorological sample data through channel feature weighting in the initial dynamic weighting module to obtain channel-weighted sample data includes:

[0016] The historical meteorological sample data are respectively input into the first max pooling layer and the first average pooling layer to obtain two first local feature information.

[0017] The two first local feature information are respectively input into the multilayer perceptron to obtain two corresponding intermediate feature information;

[0018] The two intermediate feature information are added together to obtain the added feature information, and an activation function is used to process the added feature information to obtain the channel weight information.

[0019] Based on the channel weight information and the historical meteorological sample data, the channel-weighted sample data is obtained.

[0020] In one embodiment, the step of processing the channel-weighted sample data through spatial feature weighting in the initial dynamic weighting module to obtain spatially weighted sample data includes:

[0021] The channel-weighted sample data are respectively input into the second max pooling layer and the second average pooling layer to obtain two corresponding second local feature information.

[0022] The two second local feature information are merged to obtain the merged feature information;

[0023] The merged feature information is then subjected to convolution and activation function processing in sequence to obtain spatial weight information;

[0024] The spatially weighted sample data is obtained based on the spatial weight information and the channel-weighted sample data.

[0025] In one embodiment, the step of inputting the weighted meteorological sample data into the initial time-series feature extraction module to obtain meteorological time-series feature data includes:

[0026] The weighted meteorological sample data is processed by feature extraction to obtain candidate feature data;

[0027] The weighted meteorological sample data is processed by one-dimensional convolution to obtain residual data;

[0028] The candidate feature data and the residual data are added together to obtain the meteorological time series feature data.

[0029] In one embodiment, the step of extracting candidate feature data from the weighted meteorological sample data includes:

[0030] The weighted meteorological sample data is then processed by dilated causal convolution to obtain dilated causal convolution data.

[0031] The dilated causal convolutional data is subjected to weight normalization, activation function processing, and random deactivation processing in sequence to obtain the candidate feature data.

[0032] Secondly, this application also provides a training device for a substation micro-meteorological prediction model, comprising:

[0033] The sample weighting module is used to acquire historical meteorological sample data of the substation to be predicted, and input the historical meteorological sample data into the initial dynamic weighting module to obtain the weighted meteorological sample data;

[0034] The feature extraction module is used to input the weighted meteorological sample data into the initial time series feature extraction module to obtain meteorological time series feature data;

[0035] The fully connected layer module is used to input the meteorological time-series feature data into the initial fully connected layer to obtain meteorological prediction sample data;

[0036] The initial training module is used to acquire actual meteorological data, and iteratively train the initial dynamic weighting module, the initial temporal feature extraction module, and the initial fully connected layer based on the meteorological prediction sample data and the actual meteorological data, to obtain the trained target dynamic weighting module, target temporal feature extraction module, and candidate fully connected layer.

[0037] The secondary training module is used to employ a cross-optimization model to iteratively train the candidate fully connected layer based on the historical meteorological sample data, the actual meteorological data, the target dynamic weighting module, and the target temporal feature extraction module, thereby obtaining the trained target fully connected layer.

[0038] The model determination module is used to obtain the trained substation micro-meteorological prediction model based on the target dynamic weighting module, the target temporal feature extraction module, and the target fully connected layer.

[0039] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0040] Historical meteorological sample data of the substation to be predicted is obtained, and the historical meteorological sample data is input into the initial dynamic weighting module to obtain the weighted meteorological sample data;

[0041] The weighted meteorological sample data is input into the initial time series feature extraction module to obtain meteorological time series feature data;

[0042] The meteorological time-series feature data is input into the initial fully connected layer to obtain meteorological prediction sample data;

[0043] Acquire actual meteorological data, and iteratively train the initial dynamic weighting module, the initial temporal feature extraction module, and the initial fully connected layer based on the meteorological prediction sample data and the actual meteorological data to obtain the trained target dynamic weighting module, target temporal feature extraction module, and candidate fully connected layer;

[0044] A cross-optimization model is adopted, based on the historical meteorological sample data, the actual meteorological data, the target dynamic weighting module and the target temporal feature extraction module, to iteratively train the candidate fully connected layer and obtain the trained target fully connected layer.

[0045] Based on the target dynamic weighting module, the target temporal feature extraction module, and the target fully connected layer, the trained substation micro-meteorological prediction model is obtained.

[0046] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0047] Historical meteorological sample data of the substation to be predicted is obtained, and the historical meteorological sample data is input into the initial dynamic weighting module to obtain the weighted meteorological sample data;

[0048] The weighted meteorological sample data is input into the initial time series feature extraction module to obtain meteorological time series feature data;

[0049] The meteorological time-series feature data is input into the initial fully connected layer to obtain meteorological prediction sample data;

[0050] Acquire actual meteorological data, and iteratively train the initial dynamic weighting module, the initial temporal feature extraction module, and the initial fully connected layer based on the meteorological prediction sample data and the actual meteorological data to obtain the trained target dynamic weighting module, target temporal feature extraction module, and candidate fully connected layer;

[0051] A cross-optimization model is adopted, based on the historical meteorological sample data, the actual meteorological data, the target dynamic weighting module and the target temporal feature extraction module, to iteratively train the candidate fully connected layer and obtain the trained target fully connected layer.

[0052] Based on the target dynamic weighting module, the target temporal feature extraction module, and the target fully connected layer, the trained substation micro-meteorological prediction model is obtained.

[0053] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0054] Historical meteorological sample data of the substation to be predicted is obtained, and the historical meteorological sample data is input into the initial dynamic weighting module to obtain the weighted meteorological sample data;

[0055] The weighted meteorological sample data is input into the initial time series feature extraction module to obtain meteorological time series feature data;

[0056] The meteorological time-series feature data is input into the initial fully connected layer to obtain meteorological prediction sample data;

[0057] Acquire actual meteorological data, and iteratively train the initial dynamic weighting module, the initial temporal feature extraction module, and the initial fully connected layer based on the meteorological prediction sample data and the actual meteorological data to obtain the trained target dynamic weighting module, target temporal feature extraction module, and candidate fully connected layer;

[0058] A cross-optimization model is adopted, based on the historical meteorological sample data, the actual meteorological data, the target dynamic weighting module and the target temporal feature extraction module, to iteratively train the candidate fully connected layer and obtain the trained target fully connected layer.

[0059] Based on the target dynamic weighting module, the target temporal feature extraction module, and the target fully connected layer, the trained substation micro-meteorological prediction model is obtained.

[0060] The training method, apparatus, computer equipment, computer-readable storage medium, and computer program product for the aforementioned substation micro-meteorological prediction model first acquire historical meteorological sample data of the substation to be predicted. This historical meteorological sample data is then input into an initial dynamic weighting module to obtain weighted meteorological sample data. The dynamic weighting module assigns different weights to different data elements, highlighting key features, reducing the impact of redundant features, improving the model's focus on important features, and enhancing the model's input feature representation capability. Next, the weighted meteorological sample data is input into an initial time-series feature extraction module to obtain meteorological time-series feature data. This module captures the time dependence and temporal patterns in historical meteorological data, thereby more accurately extracting time-series features related to future meteorological conditions and solving the problem of limited receptive field in traditional models when processing long-term series. Then, the meteorological time-series feature data is input into an initial fully connected layer to obtain meteorological prediction sample data. Through further processing by the fully connected layer, the time-series feature data is mapped to the output results of future meteorological predictions. Finally, actual meteorological data is acquired, and the model is iteratively trained based on the meteorological prediction sample data and the actual meteorological data. The initial dynamic weighting module, initial temporal feature extraction module, and initial fully connected layer are trained to obtain the target dynamic weighting module, target temporal feature extraction module, and candidate fully connected layer. By comparing feedback from actual meteorological data with meteorological prediction sample data, the initial dynamic weighting module, temporal feature extraction module, and fully connected layer are gradually optimized to better adapt to actual conditions, reduce errors, and improve prediction accuracy. Then, a cross-optimization model is used to iteratively train the candidate fully connected layer based on historical meteorological sample data, actual meteorological data, the target dynamic weighting module, and the target temporal feature extraction module, resulting in the trained target fully connected layer. Through the cross-optimization model, the fully connected layer is optimized a second time, which can effectively avoid the model parameters getting stuck in local optima, further improving the prediction ability and generalization ability of the fully connected layer, and enhancing the overall performance of the model. Finally, based on the target dynamic weighting module, target temporal feature extraction module, and target fully connected layer, a trained substation micro-meteorological prediction model is obtained. Integrating the target dynamic weighting module, target temporal feature extraction module, and target fully connected layer generates a highly accurate micro-meteorological prediction model, providing accurate meteorological forecasts for substations. The above method, by introducing a dynamic weighting module, a time-series feature extraction module, and a fully connected layer, combined with iterative training and cross-optimization of the model using actual meteorological data, forms a highly efficient micro-meteorological forecasting model. This model overcomes the limitations of traditional forecasting methods, enhances the focus on key meteorological elements, accurately captures the long-term and short-term dependencies of time series, and ultimately generates more accurate meteorological forecast results. The fully connected layer after secondary optimization improves the model's generalization ability, enabling it to have stronger predictive capabilities under different meteorological conditions, thus ensuring the safe and stable operation of the power system. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 is a flowchart illustrating the training method of a substation micro-meteorological prediction model in one embodiment;

[0063] Figure 2 is a flowchart illustrating the steps for obtaining weighted meteorological sample data in one embodiment;

[0064] Figure 3 is a schematic diagram of the structure of the dynamic weighting module in one embodiment;

[0065] Figure 4 is a schematic diagram of the channel feature weighting process in the dynamic weighting module of one embodiment;

[0066] Figure 5 is a schematic diagram of the spatial feature weighting process in the dynamic weighting module of one embodiment;

[0067] Figure 6 is a schematic diagram of the structure of causal convolution in one embodiment;

[0068] Figure 7 is a schematic diagram of the structure of dilated convolution in one embodiment;

[0069] Figure 8 is a schematic diagram of the temporal feature extraction module in one embodiment;

[0070] Figure 9 is a flowchart illustrating the steps of training a fully connected layer twice in one embodiment;

[0071] Figure 10 is a structural block diagram of a training device for a substation micro-meteorological prediction model in one embodiment;

[0072] Figure 11 is an internal structure diagram of a computer device in one embodiment. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0074] In one embodiment, as shown in Figure 1, a training method for a substation micro-meteorological prediction model is provided. This embodiment illustrates the method by applying it to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. The terminal can be, but is not limited to, various personal computers, laptops, smartphones, and tablets. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. In this embodiment, the method includes the following steps:

[0075] Step S101: Obtain historical meteorological sample data of the substation to be predicted, and input the historical meteorological sample data into the initial dynamic weighting module to obtain weighted meteorological sample data.

[0076] Historical meteorological sample data refers to meteorological element data collected by the meteorological monitoring system in the substation over a period of time, including but not limited to temperature, humidity, rainfall, wind speed, wind direction, air pressure, and sunshine.

[0077] One sample of data includes multiple meteorological element data at multiple consecutive time points, that is, meteorological forecasting is carried out using multiple meteorological element data at multiple consecutive time points before the forecast time point.

[0078] The initial dynamic weighting module is used to automatically adjust the weights between meteorological elements. This module analyzes the mutual influence between different meteorological elements and assigns appropriate weights to different features to ensure that important features receive higher priority in subsequent processing stages.

[0079] For example, firstly, the terminal obtains historical meteorological sample data from the substation to be predicted, and preprocesses the data, including missing value imputation, outlier detection, and data normalization. Next, the preprocessed meteorological data is input into the initial dynamic weighting module. In this module, the system calculates the relative importance of different meteorological elements over a specific time period using an algorithm, and assigns a dynamic weight to each meteorological element based on its importance. This weighting method ensures that the influence of important meteorological features is enhanced in subsequent analysis, while the influence of secondary features is appropriately weakened.

[0080] Step S102: Input the weighted meteorological sample data into the initial time series feature extraction module to obtain meteorological time series feature data.

[0081] Meteorological time-series characteristic data refers to the changing trends and characteristic patterns of meteorological sample data over time after processing by the time-series feature extraction module. This data contains the time-series characteristics of meteorological elements and can reflect the patterns of change of meteorological elements over time.

[0082] The initial time-series feature extraction module is used to extract time-series information from the weighted meteorological sample data, capturing the correlation and short-term / long-term dependence of meteorological data over time. Through this module, the temporal correlations and trend changes between meteorological elements can be identified, providing a basis for subsequent forecasting.

[0083] For example, the terminal first inputs the weighted meteorological sample data into the initial time-series feature extraction module. In this module, the terminal analyzes the time-series characteristics of the meteorological data using algorithms, extracting the features of each meteorological element in the time dimension. This module can capture the changing patterns of meteorological elements at different times and identify the potential time dependencies between meteorological elements. Through this time-series feature extraction, the terminal can effectively construct the time-series features of the meteorological samples, thereby providing more accurate input data for subsequent model training and prediction.

[0084] Step S103: Input the meteorological time series feature data into the initial fully connected layer to obtain meteorological prediction sample data.

[0085] Among them, meteorological time-series feature data is data representing the time-varying characteristics of meteorological elements after being processed by the time-series feature extraction module. It includes the time dependence of meteorological variables and reflects the dynamic characteristics of meteorological elements.

[0086] The initial fully connected layer is a neural network layer used to nonlinearly combine and map the input time-series feature data to generate the final weather forecast output.

[0087] For example, the terminal inputs meteorological time-series feature data into the initial fully connected layer. The fully connected layer performs a linear transformation on the input data through its weight matrix to obtain the meteorological prediction results for the input meteorological time-series feature data, i.e., meteorological prediction sample data.

[0088] Step S104: Obtain actual meteorological data, and iteratively train the initial dynamic weighting module, the initial temporal feature extraction module, and the initial fully connected layer based on the meteorological forecast sample data and the actual meteorological data, to obtain the trained target dynamic weighting module, the target temporal feature extraction module, and the candidate fully connected layer.

[0089] Actual meteorological data refers to the real-time weather conditions collected from the substation's meteorological monitoring system, encompassing meteorological elements related to forecasting such as temperature, humidity, wind speed, and wind direction. This actual meteorological data is used to compare with sample meteorological forecast data to evaluate the accuracy of the model's predictions.

[0090] For example, the terminal first obtains actual meteorological data from the substation and compares this data with meteorological forecast sample data to calculate the prediction error. Using this error, the terminal iteratively trains each module of the model using the backpropagation algorithm, gradually optimizing the parameters in the initial dynamic weighting module, the initial time-series feature extraction module, and the initial fully connected layer. During iterative training, the model continuously updates the weights, gradually improving its ability to extract meteorological time-series features and enhancing the accurate identification of weighted features.

[0091] Step S105: Using a cross-optimization model, based on historical meteorological sample data, actual meteorological data, target dynamic weighting module, and target temporal feature extraction module, candidate fully connected layers are iteratively trained to obtain the trained target fully connected layer.

[0092] The cross-optimization model refers to an optimization algorithm that combines horizontal and vertical cross-mechanisms to effectively avoid getting stuck in local optima during the training of fully connected layer parameters, thereby improving the global optimization capability of the fully connected layer. This optimization model iteratively adjusts the weights and biases of the fully connected layer to better adapt to the relationship between historical and actual meteorological data, thus enhancing the model's generalization ability.

[0093] For example, the terminal first acquires historical meteorological sample data and actual meteorological data, and then inputs this data into the candidate fully connected layer by combining the outputs of the optimized target dynamic weighting module and the target time-series feature extraction module. Next, the terminal optimizes the candidate fully connected layer using a cross-optimization model, gradually adjusting its weight parameters. The cross-optimization model ensures that the fully connected layer avoids getting trapped in local optima while fully mining the global features of the meteorological data during training by alternately performing horizontal cross-optimization (exploring the global optimum) and vertical cross-optimization (deepening local detail optimization).

[0094] Step S106: Based on the target dynamic weighting module, the target temporal feature extraction module, and the target fully connected layer, the trained substation micro-meteorological prediction model is obtained.

[0095] For example, the terminal integrates the trained target dynamic weighting module, target temporal feature extraction module, and target fully connected layer to form the final substation micro-meteorological prediction model. Specifically, in practical applications, the terminal first inputs real-time collected substation meteorological data into the target dynamic weighting module. This module assigns weights according to the importance of different features and then passes the weighted data to the target temporal feature extraction module. The latter learns the temporal patterns in historical data to extract temporal features useful for current prediction. Then, the extracted temporal features are input into the target fully connected layer. Through the deep learning capabilities of this layer, a prediction of future weather conditions is finally generated.

[0096] In the training method of the substation micro-meteorological prediction model mentioned above, firstly, historical meteorological sample data of the substation to be predicted is acquired and input into the initial dynamic weighting module to obtain weighted meteorological sample data. Through the dynamic weighting module, different weights can be assigned to different data elements, highlighting key features, reducing the influence of redundant features, improving the model's attention to important features, and enhancing the model's input feature representation capability. Next, the weighted meteorological sample data is input into the initial time-series feature extraction module to obtain meteorological time-series feature data. The time-series feature extraction module can capture the time dependence and time-series patterns in historical meteorological data, thereby more accurately extracting time-series features related to future meteorological conditions and solving the problem of limited receptive field in traditional models when processing long-term series. Then, the meteorological time-series feature data is input into the initial fully connected layer to obtain meteorological prediction sample data. Through further processing by the fully connected layer, the time-series feature data is mapped to the output result of future meteorological prediction. Finally, actual meteorological data is acquired, and the initial dynamic weighting module and the initial time-series feature extraction module are iteratively trained based on the meteorological prediction sample data and the actual meteorological data. The system first establishes a target dynamic weighting module, a target temporal feature extraction module, and a candidate fully connected layer. Through feedback from actual meteorological data and comparison with meteorological prediction sample data, the initial dynamic weighting module, temporal feature extraction module, and fully connected layer are gradually optimized to better adapt to actual conditions, reduce errors, and improve prediction accuracy. Then, a cross-optimization model is employed. Based on historical meteorological sample data, actual meteorological data, the target dynamic weighting module, and the target temporal feature extraction module, the candidate fully connected layer is iteratively trained to obtain the trained target fully connected layer. This cross-optimization model performs secondary optimization on the fully connected layer, effectively preventing model parameters from getting trapped in local optima, further improving the predictive and generalization abilities of the fully connected layer, and enhancing the overall model performance. Finally, based on the target dynamic weighting module, the target temporal feature extraction module, and the target fully connected layer, a trained substation micro-meteorological prediction model is obtained. Integrating these modules generates a highly accurate micro-meteorological prediction model, providing precise meteorological forecasts for substations. The above method, by introducing a dynamic weighting module, a time-series feature extraction module, and a fully connected layer, combined with iterative training and cross-optimization of the model using actual meteorological data, forms a highly efficient micro-meteorological forecasting model. This model overcomes the limitations of traditional forecasting methods, enhances the focus on key meteorological elements, accurately captures the long-term and short-term dependencies of time series, and ultimately generates more accurate meteorological forecast results. The fully connected layer after secondary optimization improves the model's generalization ability, enabling it to have stronger predictive capabilities under different meteorological conditions, thus ensuring the safe and stable operation of the power system.

[0097] In an exemplary embodiment, as shown in Figure 2, the above step S101, which inputs historical meteorological sample data into the initial dynamic weighting module to obtain weighted meteorological sample data, can also be achieved through the following steps:

[0098] Step S201: The historical meteorological sample data is processed by channel feature weighting in the initial dynamic weighting module to obtain channel-weighted sample data.

[0099] Step S202: The channel-weighted sample data is processed by spatial feature weighting in the initial dynamic weighting module to obtain spatially weighted sample data, which is used as the weighted meteorological sample data.

[0100] For example, the terminal inputs historical meteorological sample data into the initial dynamic weighting module. First, channel weighting is performed, specifically through max pooling and average pooling operations to generate feature maps. Then, a nonlinear mapping is applied to the feature maps, and finally, the channel attention weights are calculated and multiplied by the sample data to generate weighted sample data. Next, the terminal inputs the channel-weighted sample data into the spatial weighting processing part of the initial dynamic weighting module. Max pooling and average pooling operations are performed on the feature maps to generate two types of spatial feature maps. Then, a 1×1 convolution operation is used to merge these two feature maps and perform weight calculation to obtain the spatial attention matrix. Finally, the sample data obtained after combining channel and spatial weighting can more accurately reflect the spatiotemporal characteristics of meteorological data. Figure 3 shows a schematic diagram of the dynamic weighting module in one embodiment.

[0101] In this embodiment, the combination of channel weighting and spatial weighting effectively improves the efficiency and accuracy of extracting spatiotemporal features from meteorological sample data. Channel weighting highlights key features, while spatial weighting enhances the model's sensitivity to location information. The combined effect of these two methods greatly improves the accuracy and robustness of the micro-meteorological prediction model, making the prediction results more reliable, especially in complex meteorological environments.

[0102] In an exemplary embodiment, step S201 above, which processes historical meteorological sample data through channel feature weighting in the initial dynamic weighting module to obtain channel-weighted sample data, further includes: inputting the historical meteorological sample data into a first max pooling layer and a first average pooling layer respectively to obtain two first local feature information; inputting the two first local feature information into a multilayer perceptron respectively to obtain two intermediate feature information; adding the two intermediate feature information to obtain added feature information, and processing the added feature information using an activation function to obtain channel weight information; and obtaining channel-weighted sample data based on the channel weight information and the historical meteorological sample data.

[0103] Among them, the first max pooling layer, the first average pooling layer, the multilayer perceptron, and the activation function are all units required for channel feature weighting processing in the initial dynamic weighting module.

[0104] Among them, the max pooling layer highlights important features in the data by retaining the maximum value in each feature channel, while the average pooling layer smooths out noise in the data by calculating the average value of each feature channel, thereby obtaining two different first local feature information.

[0105] Among them, the multilayer perceptron is a non-linear model structure used to enhance the representation capability of features.

[0106] For example, the terminal inputs historical meteorological sample data into a first max pooling layer and a first average pooling layer, compressing them to obtain two first local feature information types that retain the number of feature channels (in this embodiment, one channel represents a meteorological factor). These two feature information types have different feature focus methods: one emphasizes highlighting local extreme value features, while the other emphasizes the overall smoothness of the features. Next, the terminal inputs these two first local feature information types into a multilayer perceptron, generating two intermediate feature information types through the perceptron's nonlinear mapping capability. Then, the terminal processes the summed feature information using activation functions. By using activation functions such as Sigmoid and ReLU, nonlinear processing further enhances the model's expressive power, generating the final channel weight information. The channel weight information represents the importance of each channel. Based on this weight information, the terminal weights the historical meteorological sample data to obtain the final channel-weighted sample data. Figure 4 shows a flowchart of the channel feature weighting process in the dynamic weighting module of one embodiment.

[0107] In this embodiment, channel-weighted processing, through the synergistic effect of max pooling, average pooling, multilayer perceptron, and activation functions, effectively improves the discriminative and representational capabilities of features in historical meteorological sample data. The channel-weighted sample data more accurately reflects the contribution of meteorological elements in different channels, contributing to improved accuracy and robustness of subsequent prediction models.

[0108] In an exemplary embodiment, step S202 above, which processes the channel-weighted sample data through spatial feature weighting in the initial dynamic weighting module to obtain spatially weighted sample data, further includes: inputting the channel-weighted sample data into a second max pooling layer and a second average pooling layer respectively to obtain two second local feature information; merging the two second local feature information to obtain merged feature information; sequentially performing convolution and activation function processing on the merged feature information to obtain spatial weight information; and obtaining spatially weighted sample data based on the spatial weight information and the channel-weighted sample data.

[0109] Among them, the second max pooling layer, the second average pooling layer, and the activation function are all units required for spatial feature weighting processing in the initial dynamic weighting module.

[0110] For example, firstly, the terminal inputs the channel-weighted sample data into the second max-pooling layer and the second average-pooling layer, extracting two different second local feature information from the channels of the input feature data. Next, the terminal merges these two local feature information channels, fusing the feature information from the max-pooling layer and the average-pooling layer to form a more comprehensive merged feature information. This merging operation ensures that key spatial features (in this embodiment, the time dimension is analogous to "space") are preserved while eliminating redundant information. The merged feature information is further processed through convolution. Convolution aims to capture the local dependencies of features in the spatial dimension (in this embodiment, the temporal dimension), especially the mutual influence between different locations. The convolution kernel traverses the merged feature information through a sliding window, extracting deep-level features of spatial distribution and enhancing the model's understanding of spatial structure. Then, the terminal applies activation functions to the convolutional feature information, typically using non-linear activation functions such as Sigmoid and ReLU, to further enhance the model's non-linear expressive power, ultimately generating spatial weight information. Finally, the terminal weights the channel-weighted sample data according to the calculated spatial weight information, generating the final spatially weighted sample data. This ensures that the model can not only adjust the weights according to the importance of channel features, but also optimize the data based on the spatial dimension variations of the samples, thus obtaining more accurate weighted sample data. Figure 5 shows a flowchart of the spatial feature weighting process in the dynamic weighting module of one embodiment.

[0111] In this embodiment, spatial feature weighting fully considers the distribution characteristics of meteorological features at different spatial locations. By combining max pooling, average pooling, convolution, and activation functions, the model's feature extraction capability in the spatial dimension is enhanced, effectively increasing the influence of spatial features on the prediction results and further improving the overall prediction performance of the model.

[0112] In an exemplary embodiment, step S102 above, which inputs the weighted meteorological sample data into the initial time-series feature extraction module to obtain meteorological time-series feature data, further includes: performing feature extraction processing on the weighted meteorological sample data to obtain candidate feature data; performing one-dimensional convolution processing on the weighted meteorological sample data to obtain residual data; and adding the candidate feature data and the residual data to obtain meteorological time-series feature data.

[0113] Feature extraction refers to extracting candidate features from weighted meteorological sample data using specific algorithms (such as convolutional neural networks or other time-series feature extraction techniques). Candidate features represent significant temporal patterns and trends in the time-series data. The purpose of this process is to enhance relevant information in the time-series data and reduce unnecessary redundancy, making subsequent convolutional processing more efficient.

[0114] One problem is that excessively large network layers can lead to gradient explosion or vanishing gradients, causing network collapse and a decline in learning ability. The introduction of residual data effectively addresses this problem of network degradation caused by increasing network depth.

[0115] For example, the terminal inputs weighted meteorological sample data into the time-series feature extraction module and obtains candidate feature data through a specific feature extraction algorithm. This candidate feature data contains representative time-series features extracted from the input samples, aiming to capture key changes and trends in the meteorological data. Next, the terminal inputs the weighted meteorological sample data into a one-dimensional convolution processing module to calculate residual data. This residual data captures feature information not fully expressed during the convolution process, preserving key information from the input samples. This process effectively mitigates the information loss problem that may occur as network depth increases, ensuring the integrity of feature information. Then, the terminal adds the candidate feature data and the residual data to fuse the residuals and candidate features obtained from the convolution process. This preserves the time-series features captured by the convolutional network while enhancing the network's stability and generalization ability. The added feature data contains refined meteorological time-series feature information, providing a more reliable input for subsequent prediction layers.

[0116] In this embodiment, feature extraction and residual addition are employed to effectively improve the accuracy of meteorological time-series feature extraction, ensuring the model's sensitivity to meteorological changes in the time-series dimension. Through this method, the model can better understand complex time-series patterns in meteorological data, improving prediction accuracy and robustness.

[0117] In an exemplary embodiment, the above-mentioned process of extracting features from the weighted meteorological sample data to obtain candidate feature data further includes: processing the weighted meteorological sample data through dilated causal convolution to obtain dilated causal convolution data; and sequentially performing weight normalization, activation function processing, and random deactivation processing on the dilated causal convolution data to obtain candidate feature data.

[0118] Dilated causal convolution refers to performing a convolution operation on weighted sample data using dilated causal convolution. The purpose of dilated causal convolution is to preserve the causal relationships within the time series data while expanding the receptive field of the convolution kernel by introducing holes (dilation) during the convolution operation, thereby better capturing long-term dependent feature information. Through this convolution process, important temporal patterns over longer time ranges can be extracted from time-series data, avoiding the loss of crucial historical information.

[0119] The causal convolution structure is composed of a series of stacked one-dimensional convolutions, as shown in Figure 6. It is clear that the output at the current time step is closely related to the historical inputs. For example, the output at time t is only convolved with the two inputs from the current time step and the previous time step. t and historical information {x t-3 x t-2 x t-1 x t However, when processing long sequences, causal convolutions have to increase the number of hidden layers or the size of the convolution kernel, which makes the network structure more complex. Causal convolutions also suffer from the drawback of having a small receptive field.

[0120] To overcome the challenge of small receptive field in causal convolutions, temporal convolutional neural networks incorporate dilated convolutions into their network structure. The advantage of this structure is that, with the same number of hidden layers and kernel size, dilated convolutions can achieve a larger receptive field, which is beneficial for solving long-term series prediction problems of micro-meteorological elements. The structure of dilated convolution is shown in Figure 7.

[0121] For example, the terminal inputs weighted meteorological sample data into the dilated causal convolution module. The dilated causal convolution operation processes time-series data by increasing the span of the convolution kernel, ultimately obtaining dilated causal convolution data. This more effectively captures important information in long-term series while avoiding the problem of short-series convolution easily ignoring long-range dependencies. Then, the terminal performs a series of processing steps on the dilated causal convolution data to enhance feature expressiveness and model robustness: Weight normalization: Weight normalization aims to standardize the convolution results, limiting the range of network weights, ensuring consistent feature distribution, and preventing model instability due to excessive weight differences. This step helps accelerate model convergence and prevents gradient explosion or vanishing problems. Activation function processing: After weight normalization, the terminal applies activation functions to the data, typically using non-linear activation functions such as ReLU (Modified Linear Unit). The role of activation functions is to introduce non-linear characteristics, enabling the model to learn and express complex temporal relationships. The ReLU function effectively filters out irrelevant or negative feature values, retaining positively activated feature information. Random Dropout Processing: Finally, the terminal performs random dropout processing on the activated data. Random dropout is a regularization technique used to prevent overfitting during model training. By randomly dropping connections of some neurons, the network can avoid over-reliance on certain specific features, improving the model's generalization ability. After these processing steps, the terminal finally obtains candidate feature data. This candidate feature data integrates temporal features extracted by dilated causal convolution, standardized features after weight normalization, nonlinear features processed by activation functions, and generalization ability enhanced by random dropout. Through these steps, the model can better capture and express complex temporal features, providing high-quality input for subsequent prediction stages. Figure 8 shows a schematic diagram of the temporal feature extraction module in one embodiment.

[0122] In an exemplary embodiment, as shown in Figure 9, step S105 above employs a cross-optimization model. Based on historical meteorological sample data, actual meteorological data, a target dynamic weighting module, and a target temporal feature extraction module, iteratively trains candidate fully connected layers to obtain the trained target fully connected layer. This can also be achieved through the following steps:

[0123] Step S901: Initialize the parameters of multiple fully connected layers as the initial population, obtain the basic parameters of the cross-sectional optimization model, and obtain the fitness function.

[0124] Step S902: Calculate the fitness of individuals in all populations.

[0125] Step S903: Record the optimal parameters for each individual and the global optimal parameters.

[0126] Step S904: Determine whether the number of iterations has reached the preset value. If not, proceed to step S905; otherwise, proceed to step S908.

[0127] Step S905: Perform horizontal crossover and execute the competition operator.

[0128] Step S906: Perform vertical crossover and execute the competition operator.

[0129] Step S907: A new cross-sectional population is obtained. Return to step S902.

[0130] Step S908 yields the optimal fully connected layer parameters.

[0131] For example, during the optimization process, it is necessary to determine the fitness function. Mean squared error and regularization terms are selected as the fitness function. Mean squared error ensures that the predicted results are close to the actual values, and the introduction of regularization terms prevents overfitting and improves the model's generalization ability. The fitness function can then be expressed as follows:

[0132]

[0133] Where N represents the total number of samples in the training set; This represents the parameters (threshold or bias) of the fully connected layer; M is the total number of parameters used for training the fully connected layer. and This represents the actual and predicted values ​​of micrometeorological elements; This represents the coefficient of the regularization term, which is 0.005.

[0134] The most significant feature of the cross-cutting algorithm is the dual-crossing mechanism that combines horizontal cross-cutting (hc) and vertical cross-cutting (vc). During the optimization process, the two cross-cutting operators are performed alternately and iterated continuously until the number of iterations reaches the upper limit or the loss meets the requirements.

[0135] Horizontal cross-exchange refers to the exchange of information between individuals in a population. Let S represent the population, M the number of individuals, and D the dimension of each individual. Let S(i) and S(j) represent two different individuals in the population, with S(i, d) and S(j, d) representing the d-th dimension information of individuals i and j, respectively. The formula for the horizontal cross-exchange of information between these two individuals is:

[0136]

[0137] in, , For father and New individuals generated after horizontal crossover; , representing the probability that two sets of parent generations exchange information in the same dimension; This increases the emphasis on the differences between the two parent features, allowing for the exploration of a larger search space beyond the parent features, and introducing additional randomness and diversity, which is very important for the algorithm's global search capability.

[0138] Vertical crossover targets the offspring of horizontal crossover, involving the exchange of information across different dimensions of an individual within the population. This crossover method increases population diversity and allows individuals to escape local optima under certain conditions, enhancing their optimization capabilities. The formula for vertical crossover is:

[0139]

[0140] in, Offspring generated by vertical crossover; It is used to measure the degree of information exchange between individuals across different dimensions; , These represent the d1 and d2 dimensions of the parent individual S, respectively.

[0141] In this embodiment, by employing a cross-optimization model to iteratively optimize the fully connected layer, the training efficiency and global search capability of the model can be effectively improved, ultimately resulting in a target fully connected layer with superior performance. By initializing multiple fully connected layer parameters as an initial population and combining a dual cross-optimization mechanism of horizontal and vertical cross-optimization, sufficient information exchange can occur between and within individuals, enhancing the model's ability to explore the global optimum. Simultaneously, the introduction of competition operators ensures that individuals with lower fitness are gradually eliminated, thereby driving the overall population towards a better direction. The fitness function ensures the model's ability to fit meteorological data during training, and the parameters of the fully connected layer are gradually optimized during iterative updates.

[0142] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0143] Based on the same inventive concept, this application also provides a training device for a substation micro-meteorological prediction model, which is used to implement the training method for the substation micro-meteorological prediction model described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more training device embodiments for a substation micro-meteorological prediction model provided below can be found in the limitations of the training method for the substation micro-meteorological prediction model described above, and will not be repeated here.

[0144] In an exemplary embodiment, as shown in FIG10, a training device for a substation micro-meteorological prediction model is provided, comprising: a sample weighting module 1001, a feature extraction module 1002, a fully connected layer module 1003, an initial training module 1004, a secondary training module 1005, and a model determination module 1006, wherein:

[0145] The sample weighting module 1001 is used to acquire historical meteorological sample data of the substation to be predicted, and input the historical meteorological sample data into the initial dynamic weighting module to obtain the weighted meteorological sample data.

[0146] The feature extraction module 1002 is used to input the weighted meteorological sample data into the initial time series feature extraction module to obtain meteorological time series feature data;

[0147] The fully connected layer module 1003 is used to input meteorological time-series feature data into the initial fully connected layer to obtain meteorological prediction sample data;

[0148] The initial training module 1004 is used to acquire actual meteorological data, and based on the meteorological forecast sample data and actual meteorological data, iteratively trains the initial dynamic weighting module, the initial temporal feature extraction module and the initial fully connected layer to obtain the trained target dynamic weighting module, the target temporal feature extraction module and the candidate fully connected layer.

[0149] The secondary training module 1005 is used to adopt the cross-optimization model, based on historical meteorological sample data, actual meteorological data, target dynamic weighting module and target temporal feature extraction module, to iteratively train candidate fully connected layers and obtain the trained target fully connected layer.

[0150] The model determination module 1006 is used to obtain the trained substation micro-meteorological prediction model based on the target dynamic weighting module, the target time series feature extraction module, and the target fully connected layer.

[0151] In one embodiment, the sample weighting module 1001 is further configured to process historical meteorological sample data by channel feature weighting in the initial dynamic weighting module to obtain channel-weighted sample data; and to process the channel-weighted sample data by spatial feature weighting in the initial dynamic weighting module to obtain spatially weighted sample data, which is used as weighted meteorological sample data.

[0152] In one embodiment, the sample weighting module 1001 is further configured to input historical meteorological sample data into a first max pooling layer and a first average pooling layer respectively to obtain two first local feature information; input the two first local feature information into a multilayer perceptron respectively to obtain two intermediate feature information; add the two intermediate feature information to obtain the added feature information, and use an activation function to process the added feature information to obtain channel weight information; and obtain channel-weighted sample data based on the channel weight information and historical meteorological sample data.

[0153] In one embodiment, the sample weighting module 1001 is further configured to input the channel-weighted sample data into the second max pooling layer and the second average pooling layer respectively to obtain two second local feature information; merge the two second local feature information to obtain merged feature information; perform convolution and activation function processing on the merged feature information in sequence to obtain spatial weight information; and obtain spatially weighted sample data based on the spatial weight information and the channel-weighted sample data.

[0154] In one embodiment, the feature extraction module 1002 is further configured to: extract candidate feature data from the weighted meteorological sample data; perform one-dimensional convolution on the weighted meteorological sample data to obtain residual data; and add the candidate feature data and the residual data together to obtain meteorological time-series feature data.

[0155] In one embodiment, the feature extraction module 1002 is further configured to process the weighted meteorological sample data through dilated causal convolution to obtain dilated causal convolution data; and to process the dilated causal convolution data by performing weight normalization, activation function processing and random deactivation processing in sequence to obtain candidate feature data.

[0156] The modules in the training device for the aforementioned substation micro-meteorological prediction model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0157] In an exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram is shown in Figure 11. The computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface of the computer device is used for exchanging information between the processor and external devices. The communication interface of the computer device is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a training method for a substation micro-meteorological prediction model. The display unit of the computer device is used to form a visually visible image and may be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0158] Those skilled in the art will understand that the structure shown in Figure 11 is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0159] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0160] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0161] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0162] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0163] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0164] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0165] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for training a substation microclimate prediction model, characterized in that, The method includes: Historical meteorological sample data of the substation to be predicted is obtained, and the historical meteorological sample data is input into the initial dynamic weighting module to obtain the weighted meteorological sample data; The weighted meteorological sample data is input into the initial time series feature extraction module to obtain meteorological time series feature data; The meteorological time-series feature data is input into the initial fully connected layer to obtain meteorological prediction sample data; Acquire actual meteorological data, and iteratively train the initial dynamic weighting module, the initial temporal feature extraction module, and the initial fully connected layer based on the meteorological prediction sample data and the actual meteorological data to obtain the trained target dynamic weighting module, target temporal feature extraction module, and candidate fully connected layer; A cross-optimization model is adopted, based on the historical meteorological sample data, the actual meteorological data, the target dynamic weighting module and the target temporal feature extraction module, to iteratively train the candidate fully connected layer and obtain the trained target fully connected layer. Based on the target dynamic weighting module, the target temporal feature extraction module, and the target fully connected layer, the trained substation micro-meteorological prediction model is obtained.

2. The method of claim 1, wherein, The step of inputting the historical meteorological sample data into the initial dynamic weighting module to obtain the weighted meteorological sample data includes: The historical meteorological sample data is processed by channel feature weighting in the initial dynamic weighting module to obtain channel-weighted sample data. The channel-weighted sample data is processed by spatial feature weighting in the initial dynamic weighting module to obtain spatially weighted sample data, which is used as the weighted meteorological sample data.

3. The method according to claim 2, characterized in that, The process of weighting the historical meteorological sample data using channel features in the initial dynamic weighting module to obtain channel-weighted sample data includes: The historical meteorological sample data are respectively input into the first max pooling layer and the first average pooling layer to obtain two first local feature information; The two first local feature information are respectively input into the multilayer perceptron to obtain two corresponding intermediate feature information; The two intermediate feature information are added together to obtain the added feature information, and an activation function is used to process the added feature information to obtain the channel weight information. Based on the channel weight information and the historical meteorological sample data, the channel-weighted sample data is obtained.

4. The method according to claim 2, characterized in that, The process of weighting the channel-weighted sample data and then performing spatial feature weighting in the initial dynamic weighting module to obtain spatially weighted sample data includes: The channel-weighted sample data are respectively input into the second max pooling layer and the second average pooling layer to obtain two corresponding second local feature information. The two second local feature information are merged to obtain the merged feature information; The merged feature information is then subjected to convolution and activation function processing in sequence to obtain spatial weight information; The spatially weighted sample data is obtained based on the spatial weight information and the channel-weighted sample data.

5. The method according to claim 1, characterized in that, The step of inputting the weighted meteorological sample data into the initial time-series feature extraction module to obtain meteorological time-series feature data includes: The weighted meteorological sample data is processed by feature extraction to obtain candidate feature data; The weighted meteorological sample data is processed by one-dimensional convolution to obtain residual data; The candidate feature data and the residual data are added together to obtain the meteorological time series feature data.

6. The method according to claim 5, characterized in that, The step of extracting candidate feature data from the weighted meteorological sample data includes: The weighted meteorological sample data is then processed by dilated causal convolution to obtain dilated causal convolution data. The dilated causal convolution data is subjected to weight normalization, activation function processing, and random deactivation processing in sequence to obtain the candidate feature data.

7. A training device for a substation micro-meteorological prediction model, characterized in that, The device includes: The sample weighting module is used to acquire historical meteorological sample data of the substation to be predicted, and input the historical meteorological sample data into the initial dynamic weighting module to obtain the weighted meteorological sample data; The feature extraction module is used to input the weighted meteorological sample data into the initial time series feature extraction module to obtain meteorological time series feature data; The fully connected layer module is used to input the meteorological time-series feature data into the initial fully connected layer to obtain meteorological prediction sample data; The initial training module is used to acquire actual meteorological data, and iteratively train the initial dynamic weighting module, the initial temporal feature extraction module, and the initial fully connected layer based on the meteorological prediction sample data and the actual meteorological data, to obtain the trained target dynamic weighting module, target temporal feature extraction module, and candidate fully connected layer. The secondary training module is used to employ a cross-optimization model to iteratively train the candidate fully connected layer based on the historical meteorological sample data, the actual meteorological data, the target dynamic weighting module, and the target temporal feature extraction module, thereby obtaining the trained target fully connected layer. The model determination module is used to obtain the trained substation micro-meteorological prediction model based on the target dynamic weighting module, the target temporal feature extraction module, and the target fully connected layer.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.