A power grid load time series prediction method and system fusing meteorological dynamic factors

Abstract

The present application relates to the technical field of power processing, in particular to a kind of power grid load time series prediction method and system fusing meteorological dynamics factor.Load numerical weather forecast, satellite cloud picture, radar echo chart and other multi-modal meteorological data in space-time alignment first, then extract meteorological dynamic characteristics by meteorological image time series coding network, obtain meteorological joint feature vector by cross-modal adaptive fusion gate network.Fuse it with receiving end power grid operation data, input the space-time position coding after superposition Transformer encoder to extract space-time coupling characteristics, finally, by multi-task decoder Parallel output photovoltaic output, load, cross-zone power reception prediction value, improve the prediction accuracy and power grid operation safety under extreme weather.

Description

Technical Field

[0001] This invention relates to the field of power processing technology, and in particular to a method and system for predicting power grid load time series by incorporating meteorological dynamic factors. Background Technology

[0002] In the receiving-end power grid, distributed photovoltaic (PV) output and air conditioning load are significantly affected by meteorological conditions, with source-load fluctuations tightly coupled, placing extremely high demands on short-term forecasting. Existing technologies (such as CN121235367A) often improve forecast accuracy by extracting spatial correlations between cross-regional loads and PV power, but rely solely on numerical weather prediction and do not integrate meteorological image modalities. Another type of method (such as CN121417183A) uses dynamic time warping to align multi-source time-series data, but is still limited to numerical or static features and lacks modeling of visual dynamic processes such as cloud movement. Furthermore, while forecasting methods based on dynamic graph neural networks and atmospheric physical constraints (such as CN121031916A) enhance spatial correlation and physical consistency, their inputs are mainly numerical meteorological and station operation data, without incorporating image sequences such as satellite cloud images and radar echo maps, making it difficult to perceive the impact of instantaneous changes in cloud cover on PV output.

[0003] In summary, existing technologies have gaps in multimodal meteorological data fusion, particularly in the deep interaction between image and numerical features. This leads to delayed predictions of sudden drops in photovoltaic output caused by cloud cover, making it difficult to capture net load spikes during sudden photovoltaic drops and load surges. Furthermore, early warnings are prone to failure under extreme weather conditions, threatening the safe operation of the receiving-end power grid. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies, such as the failure of early warning systems under extreme weather conditions, which threatens the safe operation of the receiving-end power grid. It provides a method and system for predicting power grid load time series by incorporating meteorological dynamic factors.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a method for time-series forecasting of power grid load by incorporating meteorological dynamic factors, comprising: loading multimodal meteorological data and performing spatiotemporal alignment processing on the multimodal meteorological data to obtain aligned multimodal meteorological data; The image data from the aligned multimodal meteorological data is input into a meteorological image temporal coding network to extract meteorological dynamic feature vectors; The numerical data in the aligned multimodal meteorological data and the meteorological dynamic feature vector are input into a cross-modal adaptive fusion gating network to obtain a joint meteorological feature vector; Load the receiving-end power grid operation data and fuse the meteorological joint feature vector with the receiving-end power grid operation data to obtain a multimodal input matrix; The spatiotemporal location codes are superimposed on the multimodal input matrix to obtain the initial latent variable matrix; The initial latent variable matrix is ​​input into the Transformer encoder to extract the spatiotemporal coupling features and obtain the encoded latent state matrix; The encoded hidden state matrix is ​​input into the multi-task decoder, which outputs photovoltaic power output prediction, load prediction, and cross-regional power reception prediction values ​​for multiple future time steps in parallel.

[0006] Optionally, load multimodal meteorological data, including: Numerical weather forecast data, meteorological satellite cloud image sequences, and weather radar echo image sequences are loaded as the multimodal meteorological data; The multimodal meteorological data is spatiotemporally aligned to obtain aligned numerical feature vectors, aligned satellite cloud image tensor sequences, and aligned radar echo image tensor sequences in the same spatiotemporal coordinate system, which are used as the aligned multimodal meteorological data.

[0007] Specifically, the image data from the aligned multimodal meteorological data is input into a meteorological image temporal coding network to extract meteorological dynamic feature vectors, including: The aligned satellite cloud image tensor sequence is input into the meteorological satellite cloud image temporal coding network to extract the dynamic feature vector of the cloud image. The aligned radar echo map tensor sequence is input into the weather radar echo map time-series coding network to extract the radar dynamic feature vector; The dynamic feature vector of the cloud image is concatenated with the dynamic feature vector of the radar to obtain the meteorological dynamic feature vector.

[0008] Specifically, the numerical data from the aligned multimodal meteorological data and the meteorological dynamic feature vector are input into a cross-modal adaptive fusion gating network to obtain a joint meteorological feature vector, including: The cross-modal adaptive fusion gating network is constructed, which includes a first linear transformation layer, a second linear transformation layer, and a weight generation network. The aligned numerical feature vector is mapped to a preset fusion dimension through the first linear transformation layer to obtain a numerical mapping vector; The meteorological dynamic feature vector is mapped to the preset fusion dimension through the second linear transformation layer to obtain the image mapping vector; The numerical mapping vector and the image mapping vector are input into the weight generation network, and the fused weight coefficients are output; wherein, the weight generation network consists of two fully connected layers and a Sigmoid activation function; Calculate the meteorological joint feature vector, where the meteorological joint feature vector = α⋅image mapping vector + (1-α)⋅numerical mapping vector; and α is the fusion weight coefficient.

[0009] The process includes, after outputting the fusion weight coefficients, applying a temporal smoothing constraint to the fusion weight coefficients: Obtain the historical fusion weight coefficients from the previous time step; Calculate the absolute value of the difference between the fusion weight coefficient at the current time step and the historical fusion weight coefficient; When the absolute value of the difference exceeds the preset smoothing threshold, the fusion weight coefficient of the current time step is updated to the historical fusion weight coefficient plus the smoothing threshold multiplied by the sign value of the absolute value of the difference; The smoothed fusion weight coefficients are used in the calculation of the meteorological joint feature vector.

[0010] Optionally, load the receiving-end power grid operation data, including: Historical load sequences, historical photovoltaic output sequences, and static physical parameter matrices for each transformer area are loaded as the operating data of the receiving-end power grid. The meteorological joint feature vector is fused with the receiving-end power grid operation data to obtain a multimodal input matrix, including: The meteorological joint feature vector is tensor-concatenated with the historical load sequence, the historical photovoltaic power output sequence, and the static physical parameter matrix of each transformer area to obtain the multimodal input matrix.

[0011] Optionally, spatiotemporal location encoding is superimposed on the multimodal input matrix to obtain an initial latent variable matrix, including: The multimodal input matrix is ​​linearly mapped to unify the feature dimensions to a preset hidden layer dimension, thereby obtaining a linear mapping matrix; Based on the absolute position of each time step in the multimodal input matrix, calculate the corresponding sine and cosine values, and concatenate them to generate a time position encoding matrix. Calculate the electrical distance between nodes in each transformer substation in the receiving-end power grid, construct the Laplace matrix, and generate the spatial topology location coding matrix after eigenvalue decomposition. The initial latent variable matrix is ​​obtained by adding the linear mapping matrix, the temporal location encoding matrix, and the spatial topological location encoding matrix.

[0012] The method further includes inputting the initial latent variable matrix into the Transformer encoder to extract spatiotemporal coupling features and obtain the encoded latent state matrix, and also includes: Threshold segmentation is performed on the last frame of the aligned radar echo tensor sequence to identify pixel regions whose combined reflectivity exceeds the strong convection threshold. The pixel region is mapped to a mask for a station area affected by strong convection through a station area-pixel space correspondence table. The initial latent variable matrix is ​​input into the Transformer encoder. In the multi-head attention calculation, the mask of the station area affected by strong convection is expanded into a mask matrix, which is then multiplied element-wise with the original attention weight matrix. The resulting weighted attention weight matrix is ​​used to extract the spatiotemporal coupling features and obtain the encoded latent state matrix.

[0013] Optionally, it also includes constructing a joint loss function and updating the network parameters with the objective of minimizing the joint loss function: The joint loss function includes a basic mean square error term, a net load asymmetric Huber penalty term, and a node power balance hard constraint term. The basic mean square error term is calculated based on the photovoltaic output forecast, the load forecast, and the cross-regional power receiving forecast. The net load asymmetric Huber penalty term is calculated based on the difference between the predicted load value and the predicted photovoltaic output value, and the actual net load value. The node power balance hard constraint term is calculated based on the photovoltaic output prediction value, the cross-regional power reception prediction value, the local backbone power base and the line loss tolerance threshold. With the goal of minimizing the joint loss function, the network parameters of the meteorological image temporal coding network, the cross-modal adaptive fusion gating network, and the Transformer encoder are updated through backpropagation.

[0014] Secondly, the present invention provides a power grid load time-series forecasting system that integrates meteorological dynamic factors, comprising: The data alignment module is used to load multimodal meteorological data and perform spatiotemporal alignment processing on the multimodal meteorological data to obtain aligned multimodal meteorological data. The feature extraction module is used to input the image data from the aligned multimodal meteorological data into the meteorological image temporal coding network to extract meteorological dynamic feature vectors; The feature fusion module is used to input the numerical data in the aligned multimodal meteorological data and the meteorological dynamic feature vector into a cross-modal adaptive fusion gating network to obtain a joint meteorological feature vector; The meteorological power grid fusion module is used to load the receiving-end power grid operation data and fuse the meteorological joint feature vector with the receiving-end power grid operation data to obtain a multimodal input matrix; The spatiotemporal location encoding module is used to superimpose spatiotemporal location encodings onto the multimodal input matrix to obtain an initial latent variable matrix; The spatiotemporal feature encoding module is used to input the initial latent variable matrix into the Transformer encoder, extract spatiotemporal coupling features, and obtain the encoded latent state matrix. The multi-task decoding output module is used to input the encoded hidden state matrix into the multi-task decoder and output the photovoltaic power output prediction, load prediction and cross-regional power reception prediction values ​​for multiple future time steps in parallel.

[0015] By implementing this invention, it is possible to load multimodal meteorological data and perform spatiotemporal alignment processing on the multimodal meteorological data to obtain aligned multimodal meteorological data; solve the problem of spatiotemporal misalignment of meteorological data from different sources, ensure the consistency of subsequent feature extraction and fusion, provide a reliable data foundation for accurately capturing dynamic meteorological processes such as cloud movement and strong convection, and avoid prediction deviations caused by data misalignment.

[0016] By implementing this invention, image data from the aligned multimodal meteorological data can be input into a meteorological image temporal coding network to extract meteorological dynamic feature vectors. This breaks through the limitations of traditional methods that only use numerical meteorological data, effectively perceives visual dynamic information such as cloud cover and radar echo changes, accurately captures the instantaneous causes of sudden drops in photovoltaic power output, and makes up for the shortcomings of existing technologies in the lack of modal modeling of meteorological images.

[0017] By implementing this invention, it is possible to input the numerical data from the aligned multimodal meteorological data and the meteorological dynamic feature vector into a cross-modal adaptive fusion gating network to obtain a joint meteorological feature vector; adaptively allocate the weights of numerical and image features, taking into account both meteorological physical laws and dynamic visual changes, to achieve deep interaction of multimodal meteorological information, improve the ability to characterize the impact of meteorological factors on power grid load, and enhance the integrity and effectiveness of meteorological features.

[0018] By implementing this invention, it is possible to load the receiving-end power grid operation data and fuse the meteorological joint feature vector with the receiving-end power grid operation data to obtain a multimodal input matrix; break down the data barriers between meteorology and power grid, establish a direct correlation between meteorological factors and power grid source load, comprehensively cover the key factors affecting load fluctuations, and provide rich and practical input information for subsequent spatiotemporal coupling feature mining.

[0019] By implementing this invention, it is possible to superimpose spatiotemporal location encoding on the multimodal input matrix to obtain an initial latent variable matrix; enable the model to perceive the temporal sequence of data and the spatial electrical relationship of the power grid distribution area, strengthen the learning of temporal dependence and spatial topological features, improve the model's understanding of the spatiotemporal distribution law of power grid load, and adapt to the complex spatial structure of the receiving-end power grid.

[0020] By implementing this invention, the initial latent variable matrix can be input into a Transformer encoder to extract spatiotemporal coupling features and obtain an encoded latent state matrix; the spatiotemporal correlation features between meteorology and power can be efficiently mined, the areas affected by extreme weather such as severe convection can be accurately identified, the pertinence and robustness of feature extraction can be enhanced, the problem of insufficient spatiotemporal coupling modeling in traditional models can be solved, and the feature perception capability under extreme weather conditions can be improved.

[0021] By implementing this invention, the encoded hidden state matrix can be input into a multi-task decoder to output photovoltaic power output prediction, load prediction, and cross-regional power receiving prediction values ​​for multiple future time steps in parallel; multi-task collaborative prediction can be achieved, and the core indicators required for grid dispatch can be output in one inference, thereby improving prediction efficiency; the mutual constraints and optimization of each task can improve the overall prediction accuracy and meet the real-time and accuracy requirements of source-load coordinated regulation of the receiving-end grid.

[0022] In summary, by implementing this invention, we can achieve deep coupling between meteorological dynamic factors and grid load, accurately capture fluctuation characteristics such as sudden drops in photovoltaic power and surges in load under extreme weather conditions, significantly improve the prediction accuracy of grid load, photovoltaic output and cross-regional power receiving under extreme weather conditions, solve the problem of failure of traditional prediction and early warning, and ensure the safe and stable operation of the receiving-end grid. Attached Figure Description

[0023] Figure 1 A flowchart illustrating a power grid load time-series prediction method incorporating meteorological dynamics factors provided by this invention; Figure 2 This is a schematic diagram of the structure of a power grid load time-series prediction system that integrates meteorological dynamic factors, provided by the present invention.

[0024] In the attached diagram, the components represented by each number are as follows: Data alignment module 11, feature extraction module 12, feature fusion module 13, meteorological power grid fusion module 14, spatiotemporal location coding module 15, spatiotemporal feature coding module 16, multi-task decoding output module 17. Detailed Implementation

[0025] Example 1, as Figure 1 As shown, this embodiment of the invention provides a method for time-series forecasting of power grid load that integrates meteorological dynamic factors, including: S100: Load multimodal meteorological data and perform spatiotemporal alignment processing on the multimodal meteorological data to obtain aligned multimodal meteorological data; S200: Input the image data from the aligned multimodal meteorological data into the meteorological image temporal coding network to extract the meteorological dynamic feature vector; S300: Input the numerical data in the aligned multimodal meteorological data and the meteorological dynamic feature vector into a cross-modal adaptive fusion gating network to obtain a joint meteorological feature vector; S400: Load the receiving-end power grid operation data and fuse the meteorological joint feature vector with the receiving-end power grid operation data to obtain a multimodal input matrix; S500: Superimpose spatiotemporal location codes onto the multimodal input matrix to obtain an initial latent variable matrix; S600: Input the initial latent variable matrix into the Transformer encoder to extract the spatiotemporal coupling features and obtain the encoded latent state matrix; S700; Input the encoded hidden state matrix into the multi-task decoder and output the photovoltaic power output prediction, load prediction, and cross-regional power reception prediction values ​​for multiple future time steps in parallel.

[0026] In step S100 of this application embodiment, loading multimodal meteorological data includes: Numerical weather forecast data, meteorological satellite cloud image sequences, and weather radar echo image sequences are loaded as the multimodal meteorological data; The multimodal meteorological data is spatiotemporally aligned to obtain aligned numerical feature vectors, aligned satellite cloud image tensor sequences, and aligned radar echo image tensor sequences in the same spatiotemporal coordinate system, which are used as the aligned multimodal meteorological data.

[0027] In step S100 of this application embodiment, the purpose of the above steps is to provide a complete multimodal meteorological input under a unified spatiotemporal benchmark for subsequent meteorological dynamic feature extraction, cross-modal feature fusion and power grid load time series prediction, to solve the prediction deviation problem caused by the spatiotemporal inconsistency between numerical meteorological data and meteorological image data, and to ensure the accuracy and real-time performance of power grid load prediction under extreme weather conditions.

[0028] To achieve the above objectives, it is first necessary to load numerical weather prediction data, meteorological satellite cloud image sequences, and weather radar echo image sequences as the multimodal meteorological data. This involves retrieving three types of data from meteorological data source systems: numerical weather prediction data, which includes gridded numerical information such as temperature, humidity, wind speed, air pressure, and precipitation probability; meteorological satellite cloud image sequences, which are sequences of satellite-captured cloud distribution images continuously acquired at fixed time intervals; and weather radar echo image sequences, which are sequences of radar-detected echo intensity images of meteorological elements such as precipitation and severe convection. These three types of data together constitute a complete meteorological input covering both numerical and image modalities. For example, loading numerical weather prediction data for the next 72 hours with a spatial resolution of 5 kilometers, loading a sequence of 48 consecutive meteorological satellite cloud images acquired every 15 minutes, and loading a sequence of 48 consecutive weather radar echo images acquired every 6 minutes—these three types of data together serve as multimodal meteorological data.

[0029] Next, the multimodal meteorological data is subjected to spatiotemporal alignment processing to obtain aligned numerical feature vectors, aligned satellite cloud image tensor sequences, and aligned radar echo image tensor sequences in the same spatiotemporal coordinate system, which are used as the aligned multimodal meteorological data.

[0030] First, a unified spatiotemporal coordinate system is established, with the receiving-end power grid coverage area as the spatial range and the prediction time step as the time unit. Then, numerical weather prediction data is interpolated and mapped to the standard coordinate system to generate aligned numerical feature vectors. Meteorological satellite cloud image sequences are converted into aligned satellite cloud image tensor sequences through geometric correction and time resampling. Weather radar echo image sequences are converted into aligned radar echo image tensor sequences through coordinate registration and time interpolation, so that the three types of data are completely matched in spatial location and timestamp.

[0031] For example, the original numerical weather forecast data with a spatial resolution of 10 km is interpolated to a standard grid of 5 km to obtain an aligned numerical feature vector with dimension N×T, where N is the number of spatial grids and T is the number of time steps. The original meteorological satellite cloud image sequence, which is one frame every 10 minutes, is resampled to one frame every 15 minutes to obtain an aligned satellite cloud image tensor sequence with size C×H×W×T, where C is the number of channels, H is the altitude, W is the width, and T is the number of time steps. The original weather radar echo image sequence in the latitude and longitude coordinate system is registered to the power grid area coordinate system to obtain an aligned radar echo image tensor sequence. All three types of data are in the same spatiotemporal coordinate system and serve as aligned multimodal meteorological data.

[0032] In step S200 of this embodiment, the image data from the aligned multimodal meteorological data is input into a meteorological image temporal coding network to extract meteorological dynamic feature vectors, including: The aligned satellite cloud image tensor sequence is input into the meteorological satellite cloud image temporal coding network to extract the dynamic feature vector of the cloud image. The aligned radar echo map tensor sequence is input into the weather radar echo map time-series coding network to extract the radar dynamic feature vector; The dynamic feature vector of the cloud image is concatenated with the dynamic feature vector of the radar to obtain the meteorological dynamic feature vector.

[0033] In step S200 of this application embodiment, the purpose of the above steps is to mine the dynamic changes of meteorological conditions such as cloud movement and strong convection development from the aligned multimodal meteorological image data, and to transform the high-dimensional image time series information into low-dimensional feature vectors, so as to provide calculable meteorological dynamic feature inputs for subsequent cross-modal feature fusion.

[0034] To achieve the above objectives, the aligned satellite cloud image tensor sequence must first be input into the meteorological satellite cloud image temporal coding network to extract the dynamic feature vector of the cloud image. The satellite cloud image tensor sequence, which is about to be spatiotemporally aligned, is fed into a dedicated meteorological satellite cloud image temporal coding network. This network captures the motion trends and morphological changes between consecutive cloud image frames through temporal convolution and attention mechanisms, compressing the spatial and temporal cloud image information into a fixed-dimensional dynamic feature vector of the cloud image.

[0035] For example, to input an aligned satellite cloud image tensor sequence with dimensions of 3 channels, 512 height, 512 width, and 48 time steps into a meteorological satellite cloud image temporal coding network, after 16 layers of temporal convolution and 8-head temporal attention calculation, the output cloud image dynamic feature vector with a dimension of 1×256 is generated.

[0036] For example, the meteorological satellite cloud image temporal coding network can be built using a temporal visual encoder based on an improved Transformer, which is a dedicated model for extracting dynamic feature vectors from cloud images.

[0037] In terms of model structure, the input is an aligned satellite cloud image tensor sequence, which passes through a 3D convolutional layer, a temporal location encoding layer, a multi-head temporal attention layer, and a fully connected output layer. The convolutional layer has 64 channels, a kernel size of 3×3×3, and a stride of 1×1×1. The attention heads are 8, and the feature hidden layer dimension is 256.

[0038] The parameters were set as follows: learning rate 0.0001, batch size 8, dropout rate 0.1, and activation function GELU.

[0039] In terms of training methodology, the training samples were derived from historical meteorological satellite cloud image sequences and corresponding meteorological dynamic annotations. Each training sample group contained a sequence of 48 consecutive aligned satellite cloud image tensors. The training run consisted of 50 rounds. The convergence criterion was that the loss on the validation set decreased by less than 0.001 for 10 consecutive rounds, and the accuracy of the cloud image dynamic feature vector extraction remained stable above 95%.

[0040] Then, the aligned radar echo map tensor sequence is input into the weather radar echo map time-series coding network to extract the radar dynamic feature vector; The radar echo tensor sequence, which is about to undergo spatiotemporal alignment, is fed into a dedicated weather radar echo temporal coding network. This network performs temporal modeling of echo intensity changes and the evolution of strong convective regions, extracts key dynamic information such as precipitation intensity and echo movement speed, and outputs a fixed-dimensional radar dynamic feature vector.

[0041] The weather radar echo image temporal coding network can be built using temporal three-dimensional convolution combined with an improved Transformer encoder, which is a dedicated model for extracting radar dynamic feature vectors.

[0042] In terms of model structure, the input is an aligned radar echo image tensor sequence, which sequentially passes through a 3D convolutional layer, a temporal location encoding layer, a multi-head temporal attention layer, a global pooling layer, and a fully connected mapping layer. The 3D convolutional layer has 48 channels, a kernel size of 3×3×2, and a stride of 1×1×1. It has 6 attention heads, a hidden layer feature dimension of 256, and uses adaptive average temporal pooling.

[0043] The parameters were set as follows: learning rate 0.00008, batch size 6, dropout rate 0.15, activation function GELU, and weight decay coefficient 0.0001.

[0044] In terms of training methodology, the training samples were derived from historical weather radar echo sequence and corresponding annotations of strong convection and precipitation intensity. Each training sample group contained a sequence of 48 consecutive aligned radar echo tensors. The training run consisted of 60 epochs. The convergence criterion was that the loss on the validation set no longer decreased and the loss value was less than 0.001 for 12 consecutive epochs, and the stability of radar dynamic feature vector extraction was higher than 94%.

[0045] Furthermore, the cloud image dynamic feature vector is concatenated with the radar dynamic feature vector to obtain the meteorological dynamic feature vector.

[0046] That is, the dynamic feature vectors of cloud images and radar are directly concatenated in the feature dimension direction, and the dynamic features extracted from the two types of image modes are integrated into a single vector to form a complete meteorological dynamic feature vector that characterizes the changes in meteorological dynamics.

[0047] For example, to concatenate the dynamic feature vector of cloud imagery with dimension 1×256 and the dynamic feature vector of radar with dimension 1×256 in the feature dimension, a dynamic feature vector of meteorological imagery with dimension 1×512 can be obtained.

[0048] In step S300 of this application embodiment, the numerical data in the aligned multimodal meteorological data and the meteorological dynamic feature vector are input into a cross-modal adaptive fusion gating network to obtain a joint meteorological feature vector, including: The cross-modal adaptive fusion gating network is constructed, which includes a first linear transformation layer, a second linear transformation layer, and a weight generation network. The aligned numerical feature vector is mapped to a preset fusion dimension through the first linear transformation layer to obtain a numerical mapping vector; The meteorological dynamic feature vector is mapped to the preset fusion dimension through the second linear transformation layer to obtain the image mapping vector; The numerical mapping vector and the image mapping vector are input into the weight generation network, and the fused weight coefficients are output; wherein, the weight generation network consists of two fully connected layers and a Sigmoid activation function; Calculate the meteorological joint feature vector, where the meteorological joint feature vector = α⋅image mapping vector + (1-α)⋅numerical mapping vector; and α is the fusion weight coefficient.

[0049] In step S300 of this application embodiment, the purpose of the above step is to adaptively fuse the aligned numerical feature vector and the meteorological dynamic feature vector across modes, automatically allocate the contribution weights of numerical features and image dynamic features, generate a unified meteorological joint feature vector that can comprehensively characterize meteorological dynamic changes, and provide standardized feature input for subsequent fusion of meteorological and power grid data.

[0050] To achieve the above objectives, it is first necessary to construct the cross-modal adaptive fusion gating network, which includes a first linear transformation layer, a second linear transformation layer, and a weight generation network. By constructing a network structure with linear mapping and adaptive weight calculation capabilities, a model carrier is provided for the unified fusion of numerical data and dynamic image features. For example, a cross-modal adaptive fusion gating network is constructed, which includes a first linear transformation layer, a second linear transformation layer, and a weight generation network consisting of two fully connected layers and a Sigmoid activation function. The preset fusion dimension is set to 512 dimensions.

[0051] The first and second linear transformation layers are the core mapping components of the cross-modal adaptive fusion gating network, used to unify feature dimensions to achieve cross-modal weighted fusion. The first linear transformation layer receives aligned numerical feature vectors, maps them to a preset fusion dimension, and outputs a numerical mapping vector.

[0052] The second linear transformation layer receives the meteorological dynamic feature vector, maps it to the same preset fusion dimension, and outputs an image mapping vector. Both layers use linear weighted transformation to ensure consistent output dimensions.

[0053] For example, with the preset fusion dimension set to 512, the first linear transformation layer converts the 1×128 aligned numerical feature vector into a 1×512 numerical mapping vector, and the second linear transformation layer converts the 1×512 meteorological dynamic feature vector into a 1×512 image mapping vector.

[0054] Then, the aligned numerical feature vector is mapped to a preset fusion dimension through the first linear transformation layer to obtain a numerical mapping vector; The aligned numerical feature vectors are mapped to a preset fusion dimension through the first linear transformation layer to obtain a numerical mapping vector. The numerical features of different original dimensions are unified to the set dimension through linear transformation to eliminate the dimensional differences between modes. For example, the aligned numerical feature vectors with an original dimension of 1×128 are mapped to a preset fusion dimension of 512 through the first linear transformation layer to obtain a 1×512 numerical mapping vector.

[0055] Next, the meteorological dynamic feature vector is mapped to the preset fusion dimension through the second linear transformation layer to obtain the image mapping vector; The meteorological dynamic feature vector is mapped to a preset fusion dimension through the second linear transformation layer to obtain an image mapping vector, so that the meteorological dynamic features and numerical features maintain the same dimension, ensuring that subsequent weighted calculations can be performed. For example, the original meteorological dynamic feature vector with a dimension of 1×512 is mapped to a preset fusion dimension of 512 through the second linear transformation layer to obtain a 1×512 image mapping vector.

[0056] Furthermore, the numerical mapping vector and the image mapping vector are input into the weight generation network to output fused weight coefficients; wherein, the weight generation network consists of two fully connected layers and a Sigmoid activation function; The numerical mapping vector and the image mapping vector are input into the weight generation network, which outputs fused weight coefficients. The weight generation network consists of two fully connected layers and a sigmoid activation function. The two fully connected layers extract the intermodal correlation features, and the sigmoid activation function constrains the output to the interval between 0 and 1 to obtain adaptive fused weight coefficients.

[0057] For example, to input a 1×512 numerical mapping vector and a 1×512 image mapping vector into a weight generation network, after two fully connected layers and sigmoid activation, the output is a fusion weight coefficient α of 0.45.

[0058] Finally, the meteorological joint feature vector is calculated, which is: meteorological joint feature vector = α⋅image mapping vector + (1-α)⋅numerical mapping vector; where α is the fusion weight coefficient.

[0059] That is, the meteorological joint feature vector is calculated. The meteorological joint feature vector is equal to α multiplied by the image mapping vector plus 1 minus α multiplied by the numerical mapping vector, where α is the fusion weight coefficient. The two types of mapping vectors are weighted and combined according to the adaptive weight to obtain the final fusion feature.

[0060] For example, with α = 0.45, multiplying the image mapping vector by 0.45 and adding 0.55 multiplied by the numerical mapping vector yields a 1×512 meteorological joint feature vector.

[0061] In step S300 of this embodiment, after outputting the fusion weight coefficients, the method further includes applying a time-series smoothing constraint to the fusion weight coefficients: Obtain the historical fusion weight coefficients from the previous time step; Calculate the absolute value of the difference between the fusion weight coefficient at the current time step and the historical fusion weight coefficient; When the absolute value of the difference exceeds the preset smoothing threshold, the fusion weight coefficient of the current time step is updated to the historical fusion weight coefficient plus the smoothing threshold multiplied by the sign value of the absolute value of the difference; The smoothed fusion weight coefficients are used in the calculation of the meteorological joint feature vector.

[0062] In step S300 of this application embodiment, the purpose of the above steps is to avoid drastic changes in the fusion weight coefficients in continuous time steps, ensure the temporal stability of the meteorological joint feature vector, and improve the temporal smoothness and physical rationality of power grid load forecasting.

[0063] To achieve the above objectives, it is first necessary to obtain the historical fusion weight coefficients from the previous time step; This involves obtaining the historical fusion weight coefficients from the previous time step, reading the fusion weight coefficients calculated in the previous time step from the model cache or time series records, and using them as the benchmark value for smoothing constraints. For example, if the current time step is step 20, the historical fusion weight coefficient of 0.42 calculated in step 19 is read.

[0064] Then, calculate the absolute value of the difference between the fusion weight coefficient at the current time step and the historical fusion weight coefficient; This involves calculating the absolute value of the difference between the fusion weight coefficient at the current time step and the historical fusion weight coefficient. The absolute value of the difference between the fusion weight coefficient output at the current time step and the historical fusion weight coefficient is then used to obtain the weight fluctuation range. For example, if the fusion weight coefficient at the current time step is 0.68 and the historical fusion weight coefficient is 0.42, the absolute value of the difference between the two is 0.26.

[0065] Next, when the absolute value of the difference exceeds a preset smoothing threshold, the fusion weight coefficient of the current time step is updated to the historical fusion weight coefficient plus the smoothing threshold multiplied by the sign value of the absolute value of the difference. When the absolute value of the difference exceeds a preset smoothing threshold, the fusion weight coefficient at the current time step is updated to the historical fusion weight coefficient plus the smoothing threshold multiplied by the sign value of the absolute value of the difference. This limits the magnitude of a single weight change according to the preset smoothing threshold, suppressing abnormal jumps. For example, if the preset smoothing threshold is 0.10, and the absolute value of the difference is 0.26, which is greater than 0.10 and has a positive sign value, the updated fusion weight coefficient is 0.42 plus 0.10, which equals 0.52.

[0066] Finally, the smoothed fusion weight coefficients are used to calculate the meteorological joint feature vector.

[0067] The smoothed fusion weight coefficients are used to calculate the meteorological joint feature vector. The weight coefficients that have undergone time-series smoothing constraints are used to replace the original weight coefficients to complete the weighted fusion calculation. For example, the smoothed 0.52 is used as the fusion weight coefficient α and substituted into the formula to calculate a stable meteorological joint feature vector.

[0068] In step S400 of this application embodiment, loading the receiving-end power grid operation data includes: Historical load sequences, historical photovoltaic output sequences, and static physical parameter matrices for each transformer area are loaded as the operating data of the receiving-end power grid. The meteorological joint feature vector is fused with the receiving-end power grid operation data to obtain a multimodal input matrix, including: The meteorological joint feature vector is tensor-concatenated with the historical load sequence, the historical photovoltaic power output sequence, and the static physical parameter matrix of each transformer area to obtain the multimodal input matrix.

[0069] In step S400 of this application embodiment, the purpose of the above steps is to construct a unified input carrier that integrates meteorological dynamic characteristics and the actual operating status of the receiving-end power grid, forming a multimodal input matrix that can be directly sent into the subsequent spatiotemporal coding module, so as to provide a complete and matching feature basis for power grid load time series prediction.

[0070] To achieve the above objectives, it is first necessary to load historical load sequences, historical photovoltaic power output sequences, and static physical parameter matrices for each distribution area as the receiving-end grid operation data. This involves reading user power consumption data, distributed photovoltaic power generation data, and a fixed parameter matrix reflecting the lines, capacity, and topology of the distribution areas from the power grid dispatching system and the data acquisition terminals in the distribution areas within a specified time range. These data together constitute complete receiving-end power grid operation data. For example, historical load sequences with a time resolution of 15 minutes and a 7-day time range for future forecasts, historical photovoltaic output sequences within the same spatiotemporal range, and static physical parameter matrices for each distribution area, consisting of line impedance, rated capacity, and node numbers for 32 distribution areas, are loaded together as receiving-end power grid operation data.

[0071] Next, the meteorological joint feature vector is fused with the receiving-end power grid operation data to obtain a multimodal input matrix, including: tensor splicing the meteorological joint feature vector with the historical load sequence, the historical photovoltaic power output sequence, and the static physical parameter matrix of each transformer area to obtain the multimodal input matrix.

[0072] This involves aligning and concatenating the combined meteorological features with the three types of power grid data mentioned above in both the transformer substation and time dimensions. First, meteorological features are extracted independently for each transformer substation: using a pre-established substation-pixel spatial correspondence table, a local meteorological feature vector is sampled from the meteorological feature map output by the meteorological image time-series coding network for each substation. This local meteorological feature vector has a dimension of 1×128, and the feature vectors differ between different substations, reflecting spatial differences in meteorological conditions such as cloud cover and radar echo intensity. Then, for each substation at each time step, the meteorological feature vector (1×128), the historical load features (1×64), the historical photovoltaic output features (1×64), and the static physical parameters (1×16, shared across all time steps) of that substation are concatenated along the feature dimension to form a comprehensive feature vector (1×272) for that substation at that time step. Finally, the multimodal input matrix is ​​organized into a three-dimensional tensor according to the station area dimension (32) and the time step dimension (48), resulting in a multimodal input matrix of dimension 32×48×272. In this matrix, the meteorological characteristics of different stations at the same time step are independent of each other, which can effectively characterize the spatial heterogeneity of meteorological processes such as strong convection. In step S500 of the embodiment of this application, the spatiotemporal location encoding is superimposed on the multimodal input matrix to obtain an initial latent variable matrix, including: The multimodal input matrix is ​​linearly mapped to unify the feature dimensions to a preset hidden layer dimension, thereby obtaining a linear mapping matrix; Based on the absolute position of each time step in the multimodal input matrix, calculate the corresponding sine and cosine values, and concatenate them to generate a time position encoding matrix. Calculate the electrical distance between nodes in each transformer substation in the receiving-end power grid, construct the Laplace matrix, and generate the spatial topology location coding matrix after eigenvalue decomposition. The initial latent variable matrix is ​​obtained by adding the linear mapping matrix, the temporal location encoding matrix, and the spatial topological location encoding matrix.

[0073] In step S500 of this application embodiment, the purpose of the above steps is to inject temporal location information and power grid spatial topology information into the multimodal input matrix, unify the feature dimensions and form an initial latent variable matrix with spatiotemporal location awareness, so as to ensure that the Transformer encoder can effectively capture the spatiotemporal coupling characteristics of meteorology and power grid.

[0074] To achieve the above objectives, it is first necessary to perform a linear mapping on the multimodal input matrix to unify the feature dimensions to the preset hidden layer dimensions and obtain a linear mapping matrix. In essence, a linear transformation maps the original feature dimensions of the multimodal input matrix to a predetermined unified hidden layer dimension, eliminating the feature dimension mismatch problem and providing standardized input for subsequent encoding calculations. For example, a 32×48×272 multimodal input matrix (32 transformer areas, 48 ​​time steps, 272 features) is linearly mapped to a preset hidden layer dimension of 256, resulting in a 32×48×256 linear mapping matrix. This matrix retains the transformer area and time step dimensions, allowing for the subsequent superposition of spatial location encoding and temporal location encoding, respectively.

[0075] Then, based on the absolute position of each time step in the multimodal input matrix, the corresponding sine and cosine values ​​are calculated and concatenated to generate a time position encoding matrix. Based on the absolute position of each time step in the multimodal input matrix, the corresponding sine and cosine values ​​are calculated and concatenated to generate a time position encoding matrix. The position code is calculated according to the sine and cosine function formulas based on the time step number, and the temporal sequence information is embedded in the matrix so that the model can perceive the temporal sequence relationship. For example, if the multimodal input matrix contains 64 time steps, the sine and cosine values ​​are calculated according to the position number and concatenated to generate a time position encoding matrix with a dimension of 64×256.

[0076] Next, the electrical distance between each transformer substation node in the receiving-end power grid is calculated, a Laplace matrix is ​​constructed, and a spatial topology location coding matrix is ​​generated after eigenvalue decomposition. First, the electrical distance between each transformer substation node in the receiving-end power grid is calculated based on the line impedance and topological connection relationship of the substation. Then, a Laplace matrix is ​​constructed and eigenvalue decomposition is performed to transform the spatial topology of the power grid into a learnable location code. For example, the electrical distance is calculated for 32 transformer substation nodes and a Laplace matrix is ​​constructed. After eigenvalue decomposition, a spatial topology location code matrix with a dimension of 32×256 is generated.

[0077] The electrical distance is a physical indicator used to quantify the degree of electrical connection between nodes in the receiving-end power grid. It is calculated based on the line impedance, topological connection relationship and power transmission path between the substations, and can truly reflect the spatial coupling relationship between power grid nodes.

[0078] The eigenvalue decomposition is a linear algebraic operation performed on the Laplace matrix to decompose it into a combination of eigenvalues ​​and eigenvectors, thereby extracting the essential spatial topological features of the receiving-end power grid nodes. First, an orthogonal decomposition is performed on the Laplace matrix constructed from electrical distances to obtain the corresponding eigenvalues ​​and eigenvectors. Then, the first few eigenvectors are selected to form a spatial topological location encoding matrix.

[0079] Furthermore, the linear mapping matrix, the temporal location encoding matrix, and the spatial topological location encoding matrix are added together via a broadcast mechanism to obtain the initial latent variable matrix. Specifically: The time location coding matrix has a shape of [48, 256], representing the location coding for 48 time steps. It is broadcast as [32, 48, 256], meaning that each station area shares the same time coding.

[0080] The spatial topology location coding matrix has a shape of [32, 256], representing the spatial topology coding of 32 stations. It is broadcast as [32, 48, 256], meaning that each station shares the same spatial coding at all time steps.

[0081] The linear mapping matrix [32,48,256] is added element by element to the two broadcast-encoded matrices to obtain an initial latent variable matrix of shape [32,48,256].

[0082] For example, the 32×48×256 linear mapping matrix is ​​added to the broadcast temporal location coding matrix (48×256→32×48×256) and the broadcast spatial topological location coding matrix (32×256→32×48×256) to obtain an initial latent variable matrix of dimension 32×48×256. This matrix can then be flattened into a [32×48,256] input to the Transformer encoder.

[0083] In step S600 of this embodiment, the initial latent variable matrix is ​​input into the Transformer encoder to extract spatiotemporal coupling features and obtain the encoded latent state matrix. The method further includes: Threshold segmentation is performed on the last frame of the aligned radar echo tensor sequence to identify pixel regions whose combined reflectivity exceeds the strong convection threshold. The pixel region is mapped to a mask for a station area affected by strong convection through a station area-pixel space correspondence table. The initial latent variable matrix is ​​input into the Transformer encoder. In the multi-head attention calculation, the mask of the station area affected by strong convection is expanded into a mask matrix, which is then multiplied element-wise with the original attention weight matrix. The resulting weighted attention weight matrix is ​​used to extract the spatiotemporal coupling features and obtain the encoded latent state matrix.

[0084] In step S600 of this application embodiment, the purpose of the above step is to introduce spatial mask constraints for strong convective weather during the process of extracting spatiotemporal coupling features by the Transformer encoder, strengthen the focus on the features of the transformer area affected by strong convection, improve the accuracy and pertinence of spatiotemporal coupling feature extraction under extreme weather conditions, and thus obtain an encoded hidden state matrix that is more in line with the actual operating state of the power grid.

[0085] To achieve the above objectives, it is first necessary to perform threshold segmentation on the last frame of the aligned radar echo tensor sequence to identify pixel regions with combined reflectivity exceeding the strong convection threshold. Using a set strong convection threshold as the segmentation standard, the algorithm iterates through every pixel in the last frame of the aligned radar echo tensor sequence and marks pixels with reflectivity values ​​greater than the strong convection threshold as strong convection pixel regions. For example, if the strong convection threshold is set to 45 dBZ, threshold segmentation is performed on the last frame of the aligned radar echo tensor sequence, identifying 128 pixel regions with combined reflectivity exceeding 45 dBZ.

[0086] The severe convection threshold is obtained by combining historical meteorological observation data with power grid fault event statistics. First, multi-year combined reflectivity data from weather radar echoes within the receiving-end power grid coverage area is collected. This data is then matched against power grid anomalies such as photovoltaic power drops and load fluctuations caused by severe convective weather during the same period, and corresponding reflectivity thresholds are selected. Next, the final fixed threshold is determined by combining the severe convective weather judgment standards issued by the local meteorological department with numerical weather forecast verification. For example, after statistical analysis and verification, the local severe convection threshold is set to 45 dB Z, used to align the threshold segmentation of the radar echo tensor sequence and identify the area affected by severe convection.

[0087] Then, the pixel region is mapped to a mask of a station affected by strong convection through a station area-pixel space correspondence table; Based on a pre-established one-to-one correspondence table between power grid substations and radar image pixel coordinates, the strong convection pixel regions in the image are converted into binary mask identifiers of the corresponding affected substations in the power grid, thus marking the substations affected by strong convection. For example, through the substation-pixel spatial correspondence table, the identified strong convection pixel regions are mapped to the masks of the 5th, 9th, and 17th substations out of 32 substations, which are marked as 1, indicating substations affected by strong convection.

[0088] The transformer area-pixel spatial correspondence table is a pre-established spatial mapping table between the receiving-end power grid transformer areas and meteorological image pixels. This table is generated through coordinate transformation and matching based on the actual latitude and longitude coordinates of each transformer area in the receiving-end power grid, as well as the pixel coordinates and geographic projection parameters aligned with satellite cloud image tensor sequences and radar echo image tensor sequences. Each row in the table records four items: transformer area number, transformer area center latitude and longitude, corresponding image pixel row number, and pixel column number, used to accurately map strong convection pixel areas in the image to the affected transformer areas.

[0089] Next, the initial latent variable matrix is ​​input into the Transformer encoder. In the multi-head attention calculation, the mask of the station area affected by strong convection is expanded into a mask matrix, which is then multiplied element-wise with the original attention weight matrix. The resulting weighted attention weight matrix is ​​used to extract the spatiotemporal coupling features and obtain the encoded latent state matrix.

[0090] First, the mask of the transformer area affected by strong convection is expanded into a matrix with the same dimension as the multi-head attention weights. Then, the attention weights of the affected transformer areas are increased by multiplying them element by element. Finally, the spatiotemporal coupled feature extraction is completed and the encoded latent state matrix is ​​output. For example, the 32-dimensional mask of the transformer area affected by strong convection is expanded into a 32×32×8 mask matrix corresponding to 8 attention heads. This matrix is ​​then multiplied element by element with the original attention weight matrix. The weighted attention weights are used to extract features, resulting in an encoded latent state matrix with a dimension of 32×48×256 (32 transformer areas, 48 ​​time steps, and 256 features). This matrix preserves the spatiotemporal structure and can be used by the subsequent multi-task decoder.

[0091] S700; Input the encoded hidden state matrix into the multi-task decoder and output the photovoltaic power output prediction, load prediction, and cross-regional power reception prediction values ​​for multiple future time steps in parallel.

[0092] In step S700 of this application embodiment, the purpose of the above step is to simultaneously output the photovoltaic power output prediction value, load prediction value and cross-regional power receiving prediction value for multiple future time steps through multi-task parallel decoding based on the extracted spatiotemporal coupling characteristics, so as to realize the integrated and accurate prediction of multi-dimensional key operating indicators of the receiving-end power grid.

[0093] To achieve the above objectives, the encoded latent state matrix needs to be input into a multi-task decoder, which outputs photovoltaic power output prediction, load prediction, and cross-regional power reception prediction values ​​for multiple future time steps in parallel. The multi-task decoder has three independent decoding branches, corresponding to the three prediction tasks of photovoltaic power output, load, and cross-regional power reception, respectively. Each independent decoding branch shares the spatiotemporal coupling characteristics of the encoded latent state matrix. Through fully connected mapping and time-series regression calculation, the three prediction results for the specified future time steps are output synchronously in the same forward inference process.

[0094] For example, the 32×48×256 encoded latent state matrix can be flattened into [32×48,256] and then input into a multi-task decoder containing three parallel and independent decoding branches; or the three-dimensional structure can be maintained, and decoding can be performed frame by frame at each time step. In this embodiment, the encoded latent state matrix is ​​split into 48 slices of [32,256] along the time step dimension, which are then input into the decoder sequentially, and the prediction results for the next 48 time steps are output.

[0095] Among them, the photovoltaic output forecast is the predicted power generation of distributed photovoltaic power within the receiving-end grid at multiple future time steps. The load forecast is the predicted overall electricity load of the receiving-end grid at multiple future time steps. The inter-regional power receiving forecast is the predicted power purchased by the receiving-end grid from external grids at multiple future time steps. These three are output in parallel by a multi-task decoder based on spatiotemporal coupling characteristics, used to characterize the source-load status of the receiving-end grid and the external power receiving situation, and are the core output results of grid load time-series forecasting. For example, the photovoltaic output forecast can be 6.2 MW, the load forecast can be 15.7 MW, and the inter-regional power receiving forecast can be 3.5 MW.

[0096] The multi-task decoder is built using a multi-branch parallel decoding architecture based on the Transformer structure.

[0097] In terms of model structure, the input is an encoded hidden state matrix, with three independent decoding branches corresponding to photovoltaic power output, load, and inter-regional power reception, respectively. Each independent decoding branch contains two decoding attention layers, one feedforward network, and one fully connected output layer. The number of attention heads is 8, the hidden layer dimension is 256, and the output time step is 48 steps.

[0098] For the hyperparameter settings, the learning rate is 0.0001, the batch size is 8, the dropout rate is 0.1, the activation function is GELU, and the weight decay coefficient is 0.0001.

[0099] In terms of training methodology, training samples are derived from historical meteorological joint feature vectors, receiving-end power grid operation data, encoded hidden state matrices, and corresponding ground truth labels. Each training sample group contains an encoded hidden state matrix and ground truth labels for the next 48 steps. The training run consists of 50 rounds. The convergence criterion is that the joint loss on the validation set no longer decreases after 10 consecutive rounds, and the average relative error of the three predicted values ​​is all below 3%.

[0100] In this embodiment of the application, the method further includes constructing a joint loss function and updating the network parameters with the goal of minimizing the joint loss function: The joint loss function includes a basic mean square error term, a net load asymmetric Huber penalty term, and a node power balance hard constraint term. The basic mean square error term is calculated based on the photovoltaic output forecast, the load forecast, and the cross-regional power receiving forecast. The net load asymmetric Huber penalty term is calculated based on the difference between the predicted load value and the predicted photovoltaic output value, and the actual net load value. The node power balance hard constraint term is calculated based on the photovoltaic output prediction value, the cross-regional power reception prediction value, the local backbone power base and the line loss tolerance threshold. With the goal of minimizing the joint loss function, the network parameters of the meteorological image temporal coding network, the cross-modal adaptive fusion gating network, and the Transformer encoder are updated through backpropagation.

[0101] In this embodiment of the application, the purpose of the above steps is to construct a joint loss function that includes error fitting, asymmetric penalty and physical constraints, update the model parameters with the goal of minimizing the joint loss, improve the prediction accuracy and ensure the physical rationality of the power grid power balance, so that the model output is more in line with the actual operation rules of the receiving-end power grid.

[0102] To achieve the above objectives, it is first necessary to construct a joint loss function, which includes a basic mean squared error term, a net load asymmetric Huber penalty term, and a node power balance hard constraint term. The three loss terms are combined with fixed weights to form an overall optimization objective that takes into account both fitting accuracy and physical constraints. For example, the basic mean squared error term, the net load asymmetric Huber penalty term, and the node power balance hard constraint term can be weighted in a ratio of 1 to 0.8 to 0.5 to form a complete joint loss function.

[0103] Specifically, the basic mean square error term needs to be calculated first based on the photovoltaic output forecast, the load forecast, and the cross-regional power receiving forecast. The basic mean square error term is calculated based on the photovoltaic power output forecast, load forecast, and cross-regional power reception forecast. The mean square error between the three forecast values ​​and their corresponding actual values ​​is calculated to measure the overall forecast deviation. For example, the basic mean square error term is obtained by averaging the squared differences between the three forecast values ​​and the actual values ​​for each of the next 48 time steps.

[0104] Then, the net load asymmetric Huber penalty term is calculated based on the difference between the load forecast value and the photovoltaic output forecast value and the actual net load value; The net load asymmetric Huber penalty term is calculated based on the difference between the predicted load and the predicted photovoltaic output, as well as the actual net load. Different penalties are applied to overestimation and underestimation of the net load, with a focus on suppressing peak prediction deviations in the net load. For example, the difference between the predicted net load and the actual net load is calculated using the asymmetric Huber function. When the net load is underestimated, the penalty coefficient is 1.5, and when it is overestimated, the penalty coefficient is 1.0, thus obtaining the net load asymmetric Huber penalty term.

[0105] The actual net load value is directly calculated from the measured operating data of the receiving-end power grid. The calculation method is to subtract the measured photovoltaic output value from the measured load value at the same time step. This data can be obtained synchronously and directly from the power grid energy management system, the distribution area acquisition terminal, and the photovoltaic grid-connected monitoring device. After spatiotemporal alignment and outlier removal, a standard time series label is obtained. For example, if the measured total load of the distribution area at a certain time step is 15 MW and the measured photovoltaic output is 6 MW, the difference between the two yields an actual net load value of 9 MW, which is used for calculating the net load asymmetric Huber penalty term.

[0106] Next, the node power balance hard constraint term is calculated based on the photovoltaic output prediction value, the cross-regional power receiving prediction value, the local backbone power base and the line loss tolerance threshold; This involves calculating the node power balance hard constraint terms based on the predicted photovoltaic output, the predicted cross-regional power reception, the local backbone power base, and the line loss tolerance threshold. This verifies the power supply and demand balance relationship at the grid nodes. If the power exceeds the line loss tolerance range, a penalty is imposed. For example, if the local backbone power base is 10 MW and the line loss tolerance threshold is set at 0.5 MW, when the predicted power deviates from the balance formula by more than 0.5 MW, the node power balance hard constraint terms are calculated based on the magnitude of the deviation.

[0107] The local backbone power base is obtained by statistically analyzing the long-term stable output average of conventional generating units such as thermal and hydropower in the receiving-end power grid. The average of the minimum technical output over a continuous period of more than one year is taken as a fixed benchmark value. For example, if the long-term stable output of the backbone power in a certain receiving-end power grid is 10 MW, it is determined as the local backbone power base. The line loss tolerance threshold is determined based on the historical distribution area line loss rate statistics of the receiving-end power grid and the power grid operation specifications. The safety margin of the maximum actual line loss of the distribution area in the past three years is taken as the threshold. For example, if the line loss of the distribution area is generally between 0.3 MW and 0.5 MW, the line loss tolerance threshold is set to 0.5 MW in combination with the operation specifications.

[0108] Finally, with the goal of minimizing the joint loss function, the network parameters of the meteorological image temporal coding network, the cross-modal adaptive fusion gating network, and the Transformer encoder are updated through backpropagation.

[0109] Based on the gradient information of the joint loss function, the weight parameters of the three core networks are updated layer by layer to gradually reduce the overall loss. For example, following the gradient descent method, the joint loss is propagated in reverse, and the parameters of the meteorological image temporal coding network, the cross-modal adaptive fusion gating network, and the Transformer encoder are adjusted sequentially to continuously reduce the joint loss until convergence.

[0110] For example, we can first calculate the gradient of the joint loss with respect to the model output, and then propagate backward along the network to calculate the gradients of the parameters of each layer of the Transformer encoder, the cross-modal adaptive fusion gating network, and the meteorological image temporal coding network. Then, we use gradient descent to synchronously update all trainable parameters according to the learning rate to reduce prediction error and constraint penalties. We repeat the process of forward inference, calculating loss, backpropagating gradients, and updating parameters until the joint loss no longer decreases significantly, at which point we can determine that the network has converged.

[0111] Example 2, as Figure 2 As shown, based on the same inventive concept as the grid load time-series prediction method integrating meteorological dynamic factors provided in Embodiment 1, this embodiment of the invention also provides a grid load time-series prediction system integrating meteorological dynamic factors, including: The data alignment module 11 is used to load multimodal meteorological data and perform spatiotemporal alignment processing on the multimodal meteorological data to obtain aligned multimodal meteorological data. Feature extraction module 12 is used to input the image data in the aligned multimodal meteorological data into the meteorological image temporal coding network to extract meteorological dynamic feature vectors; Feature fusion module 13 is used to input the numerical data in the aligned multimodal meteorological data and the meteorological dynamic feature vector into a cross-modal adaptive fusion gating network to obtain a joint meteorological feature vector; The meteorological power grid fusion module 14 is used to load the receiving-end power grid operation data and fuse the meteorological joint feature vector with the receiving-end power grid operation data to obtain a multimodal input matrix; Spatiotemporal location encoding module 15 is used to superimpose spatiotemporal location encoding on the multimodal input matrix to obtain an initial latent variable matrix; The spatiotemporal feature encoding module 16 is used to input the initial latent variable matrix into the Transformer encoder, extract spatiotemporal coupling features, and obtain the encoded latent state matrix. The multi-task decoding output module 17 is used to input the encoded hidden state matrix into the multi-task decoder and output the photovoltaic power output prediction value, load prediction value and cross-regional power receiving prediction value for multiple future time steps in parallel.

[0112] Furthermore, the data alignment module 11 includes the following execution steps: Numerical weather forecast data, meteorological satellite cloud image sequences, and weather radar echo image sequences are loaded as the multimodal meteorological data; The multimodal meteorological data is spatiotemporally aligned to obtain aligned numerical feature vectors, aligned satellite cloud image tensor sequences, and aligned radar echo image tensor sequences in the same spatiotemporal coordinate system, which are used as the aligned multimodal meteorological data.

[0113] Furthermore, the feature extraction module 12 includes the following execution steps: The aligned satellite cloud image tensor sequence is input into the meteorological satellite cloud image temporal coding network to extract the dynamic feature vector of the cloud image. The aligned radar echo map tensor sequence is input into the weather radar echo map time-series coding network to extract the radar dynamic feature vector; The dynamic feature vector of the cloud image is concatenated with the dynamic feature vector of the radar to obtain the meteorological dynamic feature vector.

[0114] Furthermore, the feature fusion module 13 includes the following execution steps: The cross-modal adaptive fusion gating network is constructed, which includes a first linear transformation layer, a second linear transformation layer, and a weight generation network. The aligned numerical feature vector is mapped to a preset fusion dimension through the first linear transformation layer to obtain a numerical mapping vector; The meteorological dynamic feature vector is mapped to the preset fusion dimension through the second linear transformation layer to obtain the image mapping vector; The numerical mapping vector and the image mapping vector are input into the weight generation network, and the fused weight coefficients are output; wherein, the weight generation network consists of two fully connected layers and a Sigmoid activation function; Calculate the meteorological joint feature vector, where the meteorological joint feature vector = α⋅image mapping vector + (1-α)⋅numerical mapping vector; and α is the fusion weight coefficient.

[0115] The process includes, after outputting the fusion weight coefficients, applying a temporal smoothing constraint to the fusion weight coefficients: Obtain the historical fusion weight coefficients from the previous time step; Calculate the absolute value of the difference between the fusion weight coefficient at the current time step and the historical fusion weight coefficient; When the absolute value of the difference exceeds the preset smoothing threshold, the fusion weight coefficient of the current time step is updated to the historical fusion weight coefficient plus the smoothing threshold multiplied by the sign value of the absolute value of the difference; The smoothed fusion weight coefficients are used in the calculation of the meteorological joint feature vector.

[0116] Furthermore, the meteorological power grid integration module 14 includes the following execution steps: Historical load sequences, historical photovoltaic output sequences, and static physical parameter matrices for each transformer area are loaded as the operating data of the receiving-end power grid. The meteorological joint feature vector is fused with the receiving-end power grid operation data to obtain a multimodal input matrix, including: The meteorological joint feature vector is tensor-concatenated with the historical load sequence, the historical photovoltaic power output sequence, and the static physical parameter matrix of each transformer area to obtain the multimodal input matrix.

[0117] Furthermore, the spatiotemporal location encoding module 15 includes the following execution steps: The multimodal input matrix is ​​linearly mapped to unify the feature dimensions to a preset hidden layer dimension, thereby obtaining a linear mapping matrix; Based on the absolute position of each time step in the multimodal input matrix, calculate the corresponding sine and cosine values, and concatenate them to generate a time position encoding matrix. Calculate the electrical distance between nodes in each transformer substation in the receiving-end power grid, construct the Laplace matrix, and generate the spatial topology location coding matrix after eigenvalue decomposition. The initial latent variable matrix is ​​obtained by adding the linear mapping matrix, the temporal location encoding matrix, and the spatial topological location encoding matrix.

[0118] Furthermore, the spatiotemporal feature encoding module 16 includes the following execution steps: Threshold segmentation is performed on the last frame of the aligned radar echo tensor sequence to identify pixel regions whose combined reflectivity exceeds the strong convection threshold. The pixel region is mapped to a mask for a station area affected by strong convection through a station area-pixel space correspondence table. The initial latent variable matrix is ​​input into the Transformer encoder. In the multi-head attention calculation, the mask of the station area affected by strong convection is expanded into a mask matrix, which is then multiplied element-wise with the original attention weight matrix. The resulting weighted attention weight matrix is ​​used to extract the spatiotemporal coupling features and obtain the encoded latent state matrix.

Claims

1. A method for time-series forecasting of power grid load integrating meteorological dynamic factors, characterized in that, include: Load multimodal meteorological data and perform spatiotemporal alignment processing on the multimodal meteorological data to obtain aligned multimodal meteorological data; The image data from the aligned multimodal meteorological data is input into a meteorological image temporal coding network to extract meteorological dynamic feature vectors. The numerical data from the aligned multimodal meteorological data and the meteorological dynamic feature vectors are input into a cross-modal adaptive fusion gating network to obtain a joint meteorological feature vector. The receiving-end power grid operation data is loaded, and the joint meteorological feature vectors are fused with the receiving-end power grid operation data to obtain a multimodal input matrix. The spatiotemporal position coding is superimposed on the multimodal input matrix to obtain an initial latent variable matrix. The initial latent variable matrix is ​​input into a Transformer encoder to extract spatiotemporal coupling features to obtain an encoded latent state matrix. The encoded latent state matrix is ​​input into a multi-task decoder to output photovoltaic power output predictions, load predictions, and cross-regional power receiving predictions for multiple future time steps in parallel.

2. The method as described in claim 1, characterized in that, Load multimodal meteorological data, including: Numerical weather forecast data, meteorological satellite cloud image sequences, and weather radar echo image sequences are loaded as the multimodal meteorological data; The multimodal meteorological data is spatiotemporally aligned to obtain aligned numerical feature vectors, aligned satellite cloud image tensor sequences, and aligned radar echo image tensor sequences in the same spatiotemporal coordinate system, which are used as the aligned multimodal meteorological data.

3. The method as described in claim 2, characterized in that, The image data from the aligned multimodal meteorological data is input into a meteorological image temporal coding network to extract meteorological dynamic feature vectors, including: The aligned satellite cloud image tensor sequence is input into the meteorological satellite cloud image temporal coding network to extract the dynamic feature vector of the cloud image. The aligned radar echo map tensor sequence is input into the weather radar echo map time-series coding network to extract the radar dynamic feature vector; The dynamic feature vector of the cloud image is concatenated with the dynamic feature vector of the radar to obtain the meteorological dynamic feature vector.

4. The method as described in claim 3, characterized in that, The numerical data from the aligned multimodal meteorological data and the meteorological dynamic feature vector are input into a cross-modal adaptive fusion gating network to obtain a joint meteorological feature vector, including: The cross-modal adaptive fusion gating network is constructed, which includes a first linear transformation layer, a second linear transformation layer, and a weight generation network. The aligned numerical feature vector is mapped to a preset fusion dimension through the first linear transformation layer to obtain a numerical mapping vector; The meteorological dynamic feature vector is mapped to the preset fusion dimension through the second linear transformation layer to obtain the image mapping vector; The numerical mapping vector and the image mapping vector are input into the weight generation network, and the fused weight coefficients are output; wherein, the weight generation network consists of two fully connected layers and a Sigmoid activation function; Calculate the meteorological joint feature vector, where the meteorological joint feature vector = α⋅image mapping vector + (1-α)⋅numerical mapping vector; and α is the fusion weight coefficient.

5. The method as described in claim 4, characterized in that, Following the output of the fusion weight coefficients, a time-series smoothing constraint is also applied to the fusion weight coefficients: Obtain the historical fusion weight coefficients from the previous time step; Calculate the absolute value of the difference between the fusion weight coefficient at the current time step and the historical fusion weight coefficient; When the absolute value of the difference exceeds the preset smoothing threshold, the fusion weight coefficient of the current time step is updated to the historical fusion weight coefficient plus the smoothing threshold multiplied by the sign value of the absolute value of the difference; The smoothed fusion weight coefficients are used in the calculation of the meteorological joint feature vector.

6. The method as described in claim 1, characterized in that, Load the receiving-end power grid operation data, including: Historical load sequences, historical photovoltaic output sequences, and static physical parameter matrices for each transformer area are loaded as the operating data of the receiving-end power grid. The meteorological joint feature vector is fused with the receiving-end power grid operation data to obtain a multimodal input matrix, including: The meteorological joint feature vector is tensor-concatenated with the historical load sequence, the historical photovoltaic power output sequence, and the static physical parameter matrix of each transformer area to obtain the multimodal input matrix.

7. The method as described in claim 1, characterized in that, The initial latent variable matrix is ​​obtained by superimposing spatiotemporal location codes onto the multimodal input matrix, including: The multimodal input matrix is ​​linearly mapped to unify the feature dimensions to a preset hidden layer dimension, thereby obtaining a linear mapping matrix; Based on the absolute position of each time step in the multimodal input matrix, calculate the corresponding sine and cosine values, and concatenate them to generate a time position encoding matrix. Calculate the electrical distance between nodes in each transformer substation in the receiving-end power grid, construct the Laplace matrix, and generate the spatial topology location coding matrix after eigenvalue decomposition. The initial latent variable matrix is ​​obtained by adding the linear mapping matrix, the temporal location encoding matrix, and the spatial topological location encoding matrix.

8. The method as described in claim 3, characterized in that, The initial latent variable matrix is ​​input into the Transformer encoder to extract spatiotemporal coupling features and obtain the encoded latent state matrix, which also includes: Threshold segmentation is performed on the last frame of the aligned radar echo tensor sequence to identify pixel regions whose combined reflectivity exceeds the strong convection threshold. The pixel region is mapped to a mask for a station area affected by strong convection through a station area-pixel space correspondence table. The initial latent variable matrix is ​​input into the Transformer encoder. In the multi-head attention calculation, the mask of the station area affected by strong convection is expanded into a mask matrix, which is then multiplied element-wise with the original attention weight matrix. The resulting weighted attention weight matrix is ​​used to extract the spatiotemporal coupling features and obtain the encoded latent state matrix.

9. The method as described in claim 1, characterized in that, It also includes constructing a joint loss function and updating network parameters with the goal of minimizing the joint loss function: The joint loss function includes a basic mean square error term, a net load asymmetric Huber penalty term, and a node power balance hard constraint term. The basic mean square error term is calculated based on the photovoltaic output forecast, the load forecast, and the cross-regional power receiving forecast. The net load asymmetric Huber penalty term is calculated based on the difference between the predicted load value and the predicted photovoltaic output value, and the actual net load value. The node power balance hard constraint term is calculated based on the photovoltaic output prediction value, the cross-regional power reception prediction value, the local backbone power base and the line loss tolerance threshold. With the goal of minimizing the joint loss function, the network parameters of the meteorological image temporal coding network, the cross-modal adaptive fusion gating network, and the Transformer encoder are updated through backpropagation.

10. A power grid load time-series forecasting system integrating meteorological dynamic factors, characterized in that, For implementing the power grid load time-series forecasting method incorporating meteorological dynamic factors as described in any one of claims 1-9, the system comprises: The data alignment module is used to load multimodal meteorological data and perform spatiotemporal alignment processing on the multimodal meteorological data to obtain aligned multimodal meteorological data. The feature extraction module is used to input the image data from the aligned multimodal meteorological data into the meteorological image temporal coding network to extract meteorological dynamic feature vectors; The feature fusion module is used to input the numerical data in the aligned multimodal meteorological data and the meteorological dynamic feature vector into a cross-modal adaptive fusion gating network to obtain a joint meteorological feature vector; The meteorological power grid fusion module is used to load the receiving-end power grid operation data and fuse the meteorological joint feature vector with the receiving-end power grid operation data to obtain a multimodal input matrix; The spatiotemporal location encoding module is used to superimpose spatiotemporal location encodings onto the multimodal input matrix to obtain an initial latent variable matrix; The spatiotemporal feature encoding module is used to input the initial latent variable matrix into the Transformer encoder, extract spatiotemporal coupling features, and obtain the encoded latent state matrix. The multi-task decoding output module is used to input the encoded hidden state matrix into the multi-task decoder and output the photovoltaic power output prediction, load prediction and cross-regional power reception prediction values ​​for multiple future time steps in parallel.

Images