A power load prediction method based on multi-modal data fusion and related devices
By extracting temporal and spatial features through multimodal data fusion and cross-attention mechanism, and combining multi-task learning and dynamic uncertainty weighting, the problem of insufficient accuracy in power load forecasting under complex terrain is solved, and high-precision and robust power load forecasting is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING XIAOZHUANG UNIV
- Filing Date
- 2026-02-12
- Publication Date
- 2026-05-29
AI Technical Summary
Existing power load forecasting methods are insufficient in capturing features when faced with complex load data that is non-stationary and has long-term time-dependent characteristics, making it difficult to meet the refined requirements of new power systems. Furthermore, numerical weather forecasting does not make sufficient use of spatial information in complex terrain, resulting in large prediction errors.
By employing a multimodal data fusion method, the temporal features of meteorological station observation sequence data and the spatial features of numerical weather prediction grid image sequence data are extracted using a bidirectional cross-attention mechanism and then deeply fused. Combining a multi-task learning framework and a dynamic uncertainty weighting strategy, the loss function is optimized to generate a high-precision meteorological correction sequence, which is then used for power load forecasting.
It significantly improves the accuracy and robustness of power load forecasting in complex terrain areas, solves the nonlinear systematic bias of NWP in complex terrain, and enhances the forecast accuracy and consistency under extreme weather conditions.
Smart Images
Figure CN122118677A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of deep learning technology, and in particular relates to a power load forecasting method and related apparatus based on multimodal data fusion. Background Technology
[0002] Electricity load forecasting is fundamental to power system planning, dispatching, and control. With the accelerated construction of new power systems and the significant increase in the penetration rate of distributed power sources such as wind and solar power, along with the emergence of new loads like electric vehicles and virtual power plants, the power system's sensitivity to weather fluctuations has increased significantly, and the randomness and non-stationarity of load changes have intensified. Accurate electricity load forecasting has become a core foundation for grid optimization dispatching, renewable energy consumption, and electricity market trading. Related research shows that even a small reduction in forecasting error can bring significant economic benefits.
[0003] In recent years, driven by the development of data-driven technologies, machine learning and deep learning methods have been widely applied in the field of power load forecasting, and forecasting methods have gradually evolved from traditional statistical models to high-dimensional deep neural network architectures. Although these traditional machine learning methods perform well in predicting stationary loads, they heavily rely on manual feature engineering, and their feature capture capabilities remain significantly limited when dealing with complex load data that is non-stationary and has long-term time-dependent characteristics, making it difficult to meet the needs of refined forecasting for new power systems. Subsequently, the introduction of the Transformer architecture completely changed this situation. In particular, models such as Informer and Autoformer, through probabilistic sparse attention mechanisms and sequence decomposition mechanisms, have made breakthrough progress in capturing long-term time-dependent characteristics, becoming benchmark work in this field. Meanwhile, numerical weather prediction (NWP), due to its wide-area coverage, has become an indispensable auxiliary element in load forecasting. To mitigate the systematic bias in NWP (Non-Wide Load Prediction), the academic community has begun to explore the use of deep learning to establish a non-linear mapping from NWP predictions to actual station values, aiming to further improve the accuracy of downstream load forecasting through high-quality meteorological input. However, how to fully utilize the spatial information of NWP in complex terrain and solve the physical coupling problem of multiple meteorological elements remains a challenge for current technological development. Summary of the Invention
[0004] The purpose of this invention is to provide a power load forecasting method and related device based on multimodal data fusion. By extracting temporal and spatial features from multimodal meteorological data containing NWP data, and using a bidirectional cross-attention mechanism to deeply fuse the temporal and spatial features, the bottleneck of single-modal information utilization is eliminated, and the model dynamically focuses on the most valuable local area in the NWP data for forecasting, thereby improving the accuracy of power load forecasting.
[0005] To achieve the above objectives, the present invention is implemented using the following technical solution:
[0006] In a first aspect, the present invention provides a power load forecasting method based on multimodal data fusion, comprising:
[0007] Acquire multimodal meteorological data for the target area, wherein the multimodal meteorological data includes time-consistent meteorological station observation sequence data and numerical weather prediction grid image sequence data; the meteorological station observation sequence data and the numerical weather prediction grid image sequence data contain multiple corresponding meteorological elements;
[0008] The multimodal meteorological data to be tested is input into a pre-trained multimodal meteorological correction model to obtain multi-element meteorological correction residuals. The multimodal meteorological correction model extracts temporal features from meteorological station observation sequence data and spatial features from numerical weather prediction grid image sequence data. It then deeply fuses the temporal and spatial features using a bidirectional cross-attention mechanism to obtain fused features. Based on these fused features, a multi-task learning framework is used to predict the correction residuals for each meteorological element. Finally, a dynamic uncertainty weighting strategy is employed to adaptively optimize the multi-task joint loss function.
[0009] Meteorological correction sequences are obtained based on gridded image sequence data from numerical weather prediction and multi-element meteorological correction residuals.
[0010] The meteorological correction sequence and historical load data are input into the power load forecasting model, which outputs the power load forecast results for future times.
[0011] Optionally, the structure of the multimodal meteorological correction model includes a multimodal feature extraction module, a bidirectional cross-modal attention fusion module, a shared feature fusion module, and a multi-task residual prediction module, which are sequentially connected.
[0012] The multimodal feature extraction module includes a parallel temporal feature extraction unit and a spatial feature extraction unit;
[0013] The temporal feature extraction unit includes a serial Autoformer encoder and an autocorrelation mechanism, and the spatial feature extraction unit includes a serial CNN encoder and a global average pooling layer.
[0014] The bidirectional cross-modal attention fusion module includes a serial bidirectional attention extraction unit and a cross-modal attention fusion unit;
[0015] The attention extraction unit includes a parallel time-guided spatial attention stream and a spatially guided time attention stream;
[0016] The cross-modal attention fusion unit includes a serial multi-head attention mechanism, a splicing layer, a feedforward neural network, and layer normalization.
[0017] The shared feature fusion module includes a global average pooling layer;
[0018] The multi-task residual prediction module includes independent multilayer sensor regression heads for each meteorological element.
[0019] Optionally, the step of inputting the multimodal meteorological data to be measured into a pre-trained multimodal meteorological correction model to obtain multi-factor meteorological correction residuals includes:
[0020] The time-series feature extraction unit decomposes the meteorological station observation sequence data through the Autoformer encoder to extract the seasonal term, and then aggregates the time delay information of the seasonal term through the autocorrelation mechanism to obtain the time-series feature embedding.
[0021] The spatial feature extraction unit performs multi-scale convolutional encoding on the numerical weather forecast grid image sequence data through a CNN encoder to extract multi-scale features. The multi-scale features are then aligned with the temporal feature embedding by a global average pooling layer to obtain the spatial feature embedding.
[0022] Through a bidirectional attention extraction unit, a time-guided spatial attention stream is constructed using temporal feature embedding as the query and spatial feature embedding as the key and value; and a space-guided temporal attention stream is constructed using spatial feature embedding as the query and temporal feature embedding as the key and value.
[0023] A multi-head attention mechanism is used to perform bidirectional feature interaction between spatial attention stream and temporal attention stream to obtain the interactive features. The interactive features are then concatenated through a concatenation layer and nonlinearly fused through a feedforward neural network and layer normalization to generate a fused feature sequence.
[0024] The fused feature sequence is subjected to global average pooling by the shared feature fusion module to obtain a global shared feature representation;
[0025] The multi-task residual prediction module inputs the globally shared feature representation into the independent multilayer perceptron regression head corresponding to each meteorological element, and maps it to obtain the corrected residuals of each meteorological element.
[0026] Optionally, the method of using a time-series feature extraction unit to perform sequence decomposition on meteorological station observation sequence data through an Autoformer encoder, and then aggregating the time-delay information of seasonal terms through an autocorrelation mechanism basis includes:
[0027] Perform sequence decomposition on the input meteorological station observation sequence data and extract the trend term:
[0028] ,
[0029] in, For trend items, The input is the meteorological station observation sequence. For fill operation, For average pooling;
[0030] The seasonal term is obtained by comparing the input meteorological station observation sequence data with the trend term:
[0031] ,
[0032] in, For seasonal items;
[0033] The seasonal terms are mapped to the query matrix using a linear projection matrix. Key matrix Sum matrix :
[0034] ,
[0035] ,
[0036] ,
[0037] in, , , It is a linear projection matrix;
[0038] The query matrix is transformed using the Fast Fourier Transform. Key matrix Transform to the frequency domain and calculate the autocorrelation of the seasonal term:
[0039] ,
[0040] in, For periodic correlation strength, For inverse fast Fourier transform, For Fast Fourier Transform, Indicates time delay. This represents element-wise multiplication. Indicates complex conjugation;
[0041] Multiple time delays are selected based on the autocorrelation of the seasonal term, and attention weights are generated based on the autocorrelation to adjust the value matrix. Perform the corresponding time-delay rolling aggregation to obtain the autocorrelation output:
[0042] ,
[0043] in, For the first The autocorrelation output of the layer, For time-delay index, For the amount of time delay, For scrolling, For normalized probability, Indicates the first A time delay;
[0044] The autocorrelation output of the current layer and the seasonal term input of the previous layer are subjected to sequence decomposition, and nonlinear mapping is performed through a feedforward network to obtain the temporal feature embedding of the current layer, where the th... The layer operations are as follows:
[0045] ,
[0046] ,
[0047] in, The output is the seasonal term after further sequence decomposition. The output is the trend term after further sequence decomposition. Enter the seasonal item for the previous level. For sequence decomposition, For the first Embedding of temporal features from layer output, It is a feedforward network;
[0048] The temporal feature embedding is iterated through multiple layers to obtain the final temporal feature embedding:
[0049] ,
[0050] in, The temporal feature embedding is the output of the temporal feature extraction unit. This indicates that the embedding of temporal features is performed Layer iteration, The dimension is The set of real numbers, For batch size, For sequence length, For feature dimensions.
[0051] Optionally, methods for performing multi-scale convolutional encoding on numerical weather forecast grid image sequence data using a CNN encoder, and dimensional alignment of multi-scale features with temporal feature embeddings using a global average pooling layer, include:
[0052] pass Stacked convolutional layers are used to extract deep spatial semantic features from the input numerical weather forecast grid image sequence data, resulting in a high-level feature map containing rich spatial structure information. The layer 1 is... The layer operations are as follows:
[0053] ,
[0054] in, For the first The convolutional features output by the layer, For the first The convolutional features output by the layer, This represents a two-dimensional convolution operation. and For learnable parameters, For batch normalization, For activation functions;
[0055] Compressed features are obtained by performing global average pooling on the high-level feature maps using a global average pooling layer.
[0056] ,
[0057] in, For the first High-level feature maps output by the layer. For compression features, For global average pooling, The height of the high-level feature map. The width of the high-level feature map. This is the high-level feature map output by the last convolutional layer. For height indexing, For width index;
[0058] By mapping the compressed features to the latent spatial dimension using a linear projection matrix, we obtain spatial feature embeddings aligned with the temporal feature dimension.
[0059] ,
[0060] in, Embed the spatial features output by the spatial feature extraction unit. It is a linear projection matrix.
[0061] Optionally, methods for bidirectional feature interaction between spatial attention streams and temporal attention streams through multi-head attention mechanisms include:
[0062] For time-guided spatial attention flow, the first The calculation method for each attention point is as follows:
[0063] The input features are mapped to a subspace using a linear projection matrix:
[0064] ,
[0065] ,
[0066] ,
[0067] in, The temporal feature embedding is the output of the temporal feature extraction unit. Embed the spatial features output by the spatial feature extraction unit. , , It is a linear projection matrix. The query matrix is constructed based on time-series features. The key matrix is constructed based on spatial feature embedding. The value matrix is constructed based on spatial feature embedding;
[0068] The attention score is calculated using the scaled dot product and then normalized using Softmax to obtain the weighted output of the corresponding attention head:
[0069] ,
[0070] ,
[0071] in, Attention score for spatial attention flow guided by time. For each head dimension, The weighted output of the attention head in a time-guided spatial attention flow;
[0072] For spatially guided temporal attention streams, the first The calculation method for each attention point is as follows:
[0073] The input features are mapped to a subspace using a linear projection matrix:
[0074] ,
[0075] ,
[0076] ,
[0077] in, The query matrix is constructed based on spatial feature embedding. The key matrix is constructed based on temporal features. The value matrix is constructed based on the embedding of temporal features;
[0078] The attention score is calculated using the scaled dot product and then normalized using Softmax to obtain the weighted output of the corresponding attention head:
[0079] ,
[0080] ,
[0081] in, Attention score for spatially guided temporal attention flow. The weighted output of the attention head in a spatially guided temporal attention stream;
[0082] By concatenating and linearly projecting all attention heads generated by the two attention streams, the interactive features are obtained:
[0083] ,
[0084] ,
[0085] in, For time-guided spatial attention flow output, For splicing, The first in the spatial attention stream guided by time The output of each attention head, Output weights for attention. For spatially guided temporal attention stream output, The first in the temporal attention stream guided by space The output of each attention head.
[0086] Optionally, methods for adaptively optimizing the joint loss function of multiple tasks using a dynamic uncertainty weighting strategy include:
[0087] The dynamic weights of each meteorological element learning task are obtained based on the historical average loss and the current instantaneous loss of each meteorological element learning task:
[0088] ,
[0089] ,
[0090] ,
[0091] in, For meteorological element learning task index, For the task At any moment Dynamic weights, For index, For traversing the index of meteorological elements, It is a set of meteorological elements, including continuous variables and sparse variables. For the task At any moment The ratio of the current instantaneous loss to the historical average loss. For the task At any moment The ratio of the current instantaneous loss to the historical average loss. For the task At any moment Historical average loss For the task At any moment Instantaneous loss, It is a constant. As a smoothing factor, For the task At any moment The historical average loss;
[0092] Based on the continuous variable learning task loss, the sparse variable learning task loss, and the dynamic weights of each meteorological element learning task, a multi-task joint loss function is constructed, and the model parameters are optimized through backpropagation. The expression of the multi-task joint loss function is as follows:
[0093] ,
[0094] ,
[0095] ,
[0096] ,
[0097] in, This represents the joint loss function for multiple tasks. For the task The true systematic error between meteorological observations and the original numerical weather prediction grid forecasts. For the task The corrected residual predictions For continuous variable learning task loss, Learning task loss for sparse variables, Duration For time step indexing, the continuous variable features include temperature. ,humidity Wind speed The sparse variable elements include precipitation. .
[0098] Optionally, methods for obtaining meteorological correction sequences based on numerical weather prediction grid image sequence data and multi-factor meteorological correction residuals include:
[0099] The multi-factor meteorological correction residuals are superimposed with the baseline forecast values to obtain the meteorological correction sequence:
[0100] ,
[0101] in, For meteorological correction sequences, As the baseline forecast value, Correcting residuals for multi-factor meteorological data.
[0102] Secondly, the present invention provides a power load forecasting device based on multimodal data fusion, comprising:
[0103] Multimodal meteorological data acquisition module: used to acquire the multimodal meteorological data to be measured in the target area, wherein the multimodal meteorological data includes time-consistent meteorological station observation sequence data and numerical weather prediction grid image sequence data; the meteorological station observation sequence data and the numerical weather prediction grid image sequence data contain a variety of corresponding meteorological elements;
[0104] The multi-element meteorological correction residual acquisition module is used to input the multi-modal meteorological data to be measured into a pre-trained multi-modal meteorological correction model to obtain multi-element meteorological correction residuals. The multi-modal meteorological correction model extracts temporal features based on meteorological station observation sequence data and spatial features based on numerical weather prediction grid image sequence data. It then deeply fuses the temporal and spatial features through a bidirectional cross-attention mechanism to obtain fused features. Based on these fused features, it predicts the correction residuals of each meteorological element using a multi-task learning framework and adaptively optimizes the multi-task joint loss function using a dynamic uncertainty weighting strategy.
[0105] Meteorological correction sequence acquisition module: used to acquire meteorological correction sequences based on numerical weather prediction grid image sequence data and multi-factor meteorological correction residuals;
[0106] The power load forecasting result acquisition module is used to input meteorological correction sequences and historical load data into the power load forecasting model and output the power load forecasting results for future times.
[0107] Thirdly, the present invention provides a computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the power load forecasting method based on multimodal data fusion as described in any of the first aspects.
[0108] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: By extracting the temporal features of meteorological station observation sequence data and the spatial features of numerical weather prediction grid image sequence data in parallel, and achieving deep fusion of temporal and spatial features based on a bidirectional cross-attention mechanism, this invention can capture the temporal evolution patterns of meteorological elements and identify wide-area spatial context information, thereby achieving refined matching in complex non-stationary time series. Furthermore, by combining a multi-task learning framework to jointly predict the correction residuals of various meteorological elements, and introducing a dynamic uncertainty weighting strategy to adaptively adjust the loss weights of each task, this invention solves the problem of inconsistent convergence speeds among multiple tasks, ensuring that meteorological data are physically and logically consistent. The consistency of the sequence is ultimately based on the corrected high-precision meteorological sequence for power load forecasting, which significantly improves the accuracy and robustness of power load forecasting in complex terrain areas. By constructing a nonlinear systematic bias correction mechanism based on deep residual learning, instead of directly predicting absolute meteorological values, it learns the systematic error between NWP forecast values and actual observation values, and constructs a nonlinear mapping function from macro forecasts to local realities. This solves the nonlinear systematic bias caused by the coarse spatial resolution of NWP and the limitation of physical parameterization schemes in complex terrains such as high mountains and canyons. It can effectively correct the local micro-meteorological errors caused by terrain forcing and significantly improve the forecast robustness under extreme weather conditions. Attached Figure Description
[0109] Figure 1 The diagram shown is a flowchart of a power load forecasting method based on multimodal data fusion in one embodiment of the present invention.
[0110] Figure 2 The diagram shown is an architecture diagram of a multimodal meteorological data fusion method in one embodiment of the present invention;
[0111] Figure 3 The diagram shown is an architecture diagram of a multi-task joint loss optimization method in one embodiment of the present invention;
[0112] Figure 4 The diagram shown is a flowchart of a multi-task joint loss weight calculation method in one embodiment of the present invention. Detailed Implementation
[0113] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.
[0114] Example 1
[0115] like Figure 1As shown, this embodiment provides a power load forecasting method based on multimodal data fusion. A multimodal meteorological correction model is trained using historical meteorological station observation data and corresponding time-series NWP grid data. The trained multimodal meteorological correction model outputs multi-factor meteorological correction residuals based on the NWP grid data to be measured and the corresponding time-series meteorological station observation data. Then, spatial interpolation and benchmark construction are performed on the NWP grid data to be measured to generate a benchmark NWP forecast value. The benchmark NWP forecast value is then superimposed with the multi-factor meteorological correction residuals to obtain a meteorological correction sequence. Finally, historical power load data is input into the load forecasting model, and the circuit load forecast results for future times are output. Specifically, the method includes the following steps:
[0116] I. Acquisition and Processing of Multimodal Meteorological Data
[0117] Step 1: Acquire multimodal meteorological data. Obtain historical meteorological station observation sequence data for the target area. and the corresponding numerical weather prediction (NWP) grid image sequence data All of these include temperature ( ), relative humidity ), wind speed ( ) and precipitation ( );in, Indicates the length of the time series. and These represent the number of channels for the observed variable and the NWP physical quantity, respectively. and These represent the height and width of the grid, respectively.
[0118] Step 2: Define the set of target meteorological elements to be corrected as follows These correspond to temperature, relative humidity, wind speed, and precipitation, respectively; a spatiotemporal interpolation operator is defined. , will be gridded Mapped to site geographic coordinates Then, by aligning the time series data with historical meteorological station observation sequences using linear interpolation, a preliminary baseline forecast value is obtained. The calculation formula is:
[0119] ,
[0120] in, .
[0121] II. Obtaining meteorological correction sequences based on multimodal meteorological data using a multimodal meteorological correction model.
[0122] like Figure 2 and Figure 3As shown, in this embodiment, the structure of the multimodal meteorological correction model includes a multimodal feature extraction module, a bidirectional cross-modal attention fusion module, a shared feature fusion module, and a multi-task residual prediction module, all sequentially. The multimodal feature extraction module includes parallel temporal feature extraction units and spatial feature extraction units. The temporal feature extraction unit includes a sequential Autoformer encoder and an autocorrelation mechanism, while the spatial feature extraction unit includes a sequential CNN encoder and a global average pooling layer. The bidirectional cross-modal attention fusion module includes a sequential bidirectional attention extraction unit and a cross-modal attention fusion unit. The attention extraction unit includes parallel time-guided spatial attention streams and spatially guided temporal attention streams. The cross-modal attention fusion unit includes a sequential multi-head attention mechanism, a splicing layer, a feedforward neural network, and layer normalization. The shared feature fusion module includes a global average pooling layer. The multi-task residual prediction module includes independent multilayer perceptron regression heads for each meteorological element.
[0123] Obtaining multi-factor meteorological correction residuals through a multimodal meteorological correction model includes the following specific steps:
[0124] Step 3: Preliminary sequence decomposition of time-series data, performed using the sequence decomposition module in the Autoformer time-series feature encoder. This involves processing historical meteorological station observation sequences. The input to the Autoformer encoder is first processed by moving average pooling. Smoothing periodic fluctuations to extract trend terms ,filling The operation is used to adapt the pooling kernel size to achieve integer division of the sequence length, preserve the trend features of the time series boundaries, and subtract the trend term from the input sequence to obtain the seasonal term. The calculation formula is:
[0125] ,
[0126] .
[0127] Step 4: Frequency Domain-Based Periodic Detection and Autocorrelation Calculation. The autocorrelation mechanism is used to handle the seasonal term. First, through the linear projection matrix... , , Map seasonal items to query matrix Key matrix Sum matrix :
[0128] ,
[0129] ,
[0130] ,
[0131] Subsequently, the query matrix is processed using Fast Fourier Transform (FFT) and Inverse Fast Fourier Transform (IFFT). Bond matrix Transform to the frequency domain and calculate the periodic correlation strength of the sequence itself. :
[0132] ,
[0133] in, Indicates time delay. This represents element-wise multiplication. Indicates complex conjugation.
[0134] Step 5: Top-k based latency information aggregation. Based on the obtained autocorrelation... Select the top-k most significant time delays The autocorrelation coefficients normalized by Softmax are used as attention weights to adjust the value matrix. Perform the corresponding time-delayed rolling roll aggregation to obtain the output of the autocorrelation mechanism. :
[0135] ,
[0136] in, For the first The autocorrelation output of the layer, For time-delay index, For the amount of time delay, Indicates the first A time delay. This step utilizes the detected periodicity to align the historical subsequence with the current time step, thereby extracting information.
[0137] Step 6: Deep feature extraction and encoder output, through... The encoder network consists of stacked layers. In each layer, the autocorrelation output of the current layer is concatenated with the seasonal term input residual of the previous layer, followed by sequence decomposition to eliminate intermediate trends. Finally, a nonlinear mapping is performed using a feedforward neural network (FFN). The layer's output is calculated as follows:
[0138] ,
[0139] ,
[0140] in, The output is the seasonal term after further sequence decomposition. The output is the trend term after further sequence decomposition. Enter the seasonal item for the previous level. For sequence decomposition, For the first Embedding of temporal features from layer output.
[0141] go through After layer iteration, the th The output of the layer is the final temporal feature embedding:
[0142] ,
[0143] in, The temporal feature embedding is the output of the temporal feature extraction unit. This indicates that the embedding of temporal features is performed Layer iteration, For batch size, For sequence length, For feature dimensions.
[0144] Step 7: Multi-scale spatial convolutional coding of NWP data. Definition For the first The output feature map of the convolutional layer, where, That is, input data .pass Layer-by-layer convolutional operations extract deep spatial semantics, the first layer... Layer computation includes convolution, batch normalization, and ReLU activation:
[0145] ,
[0146] in, This represents a two-dimensional convolution operation. and For learnable parameters, As weight, For bias. After After layer feature extraction, a high-level feature map containing rich spatial structure information is obtained. .
[0147] Step 8: Global pooling and dimensionality alignment of spatial features. This involves processing the high-level feature maps output from the last layer of the CNN. Global Average Pooling (GAP) is performed to compress the data into a feature vector. Then, a linear projection matrix is used. Map it to the latent spatial dimension This yields spatial feature embeddings aligned with the temporal feature dimension. :
[0148] ,
[0149] ,
[0150] in, For compression features, The height of the high-level feature map. The width of the high-level feature map. This is the high-level feature map output by the last convolutional layer. For height indexing, This step, indexed by width, transforms the high-dimensional grid image data into a low-dimensional feature vector, preparing for subsequent bidirectional cross-modal attention interactions.
[0151] Step 9: Parallel Construction of Bidirectional Cross-Attention Streams. Construct a bidirectional cross-attention module, establishing two attention streams in parallel: one is a time-guided spatial attention stream, embedded with temporal features. For the query vector, embedding with spatial features The first is the key vector and the second is the space-guided temporal attention stream, which consists of key vectors and value vectors. For query vector, with These are the key vector and value vector. This step aims to simultaneously retrieve relevant NWP spatial structure information using time trends, and to reweight historical observation sequences using space meteorological conditions.
[0152] Step 10: For any attention stream in Step 9, use a multi-head attention mechanism for feature interaction. Taking a time-guided spatial attention stream as an example, the first... The calculation process for each attention head is as follows:
[0153] First, through the linear projection matrix , , Mapping input features to a subspace:
[0154] ,
[0155] ,
[0156] ,
[0157] in, The query matrix is constructed based on time-series features. The key matrix is constructed based on spatial feature embedding. The value matrix is constructed based on spatial feature embedding;
[0158] Next, the attention score is calculated using the scaled dot product. Then, the weighted output of the header is obtained by Softmax normalization. :
[0159] ,
[0160] ,
[0161] in, Attention score for spatial attention flow guided by time. For each head dimension, The weighted output of the attention head in a time-guided spatial attention flow.
[0162] Similarly, for spatially guided temporal attention flows, exchange and , The input source is used to calculate the corresponding output. :
[0163] The input features are mapped to a subspace using a linear projection matrix:
[0164] ,
[0165] ,
[0166] ,
[0167] in, The query matrix is constructed based on spatial feature embedding. The key matrix is constructed based on temporal features. The value matrix is constructed based on the embedding of temporal features;
[0168] The attention score is calculated using the scaled dot product and then normalized using Softmax to obtain the weighted output of the corresponding attention head:
[0169] ,
[0170] ,
[0171] in, Attention score for spatially guided temporal attention flow. The weighted output of the attention head in a spatially guided temporal attention stream.
[0172] Step 11: Deep fusion of cross-modal features. This involves fusing all the features generated by the two attention streams in Step 10. Each feature is concatenated and linearly projected to obtain the interactive features. and :
[0173] ,
[0174] ,
[0175] in, For time-guided spatial attention flow output, The first in the spatial attention stream guided by time The output of each attention head, Output weights for attention. For spatially guided temporal attention stream output, The first in the temporal attention stream guided by space The output of each attention head.
[0176] Subsequently, these two interactive features are concatenated again and nonlinearly fused with a feedforward neural network (FFN) and a layer normalized (LayerNorm) to generate a fused feature sequence containing spatiotemporal complementary information. :
[0177] ,
[0178] ,
[0179] in, These are the interactive features after being spliced together.
[0180] This embodiment constructs two encoders in parallel. One encoder utilizes the autocorrelation mechanism of Autoformer to process the time series data of meteorological stations, capturing long-term dependencies and periodicity. The other encoder uses CNN to process high-resolution NWP grid images, extracting spatial structural features such as cloud distribution and wind direction. Furthermore, an innovative bidirectional cross-attention mechanism is designed to overcome the bottleneck of single-modal information utilization, achieving deep alignment between discrete station time series data and continuous NWP image data. This allows the model to dynamically focus on the most valuable local regions in the NWP image for prediction based on the current time trend, recovering the wide-area spatial context information lost by traditional methods.
[0181] Step 12: Generation of globally shared feature representations. This involves processing the fused feature sequence. Global average pooling is performed over the time dimension to obtain globally shared feature representations. As a unified input for multi-task learning:
[0182] ,
[0183] in, For timing length, To fuse feature dimensions.
[0184] like Figure 4 As shown, based on the continuous variable loss or sparse variable loss of each task, the ratio of task uncertainty estimate to instantaneous loss is calculated. Then, dynamic weights are calculated using Softmax and weighted summation is performed to obtain the final multi-task joint loss function. The specific steps are as follows:
[0185] Step 13: Construction of the multi-task regression head and generation of corrected residuals. Based on the globally shared feature representation obtained in Step 12. Construct a set of meteorological elements An independent multi-task regression head. Definition In order to target the An independent multilayer perceptron for each meteorological element (such as temperature and humidity), the perceptron consisting of fully connected layers and activation functions. These values are input into the regression header of each task, and the corrected residual values of each meteorological element at the current time are obtained. The calculation formula is:
[0186] ,
[0187] in, , For the first The learnable parameters of each regression head. The set of output residuals. This refers to the systematic deviation between the model-predicted NWP values and the actual observed values.
[0188] Step 14: Define the loss function for each task. For temperature, humidity, and wind speed (continuous variables), Calculate the loss of continuous variables For precipitation (sparse variable). Calculate the loss of sparse variables :
[0189] ,
[0190] ,
[0191] in, For continuous variable learning task loss, Learning task loss for sparse variables, For timing length, For the task The corrected residual true value, For the task The corrected residual predictions For the task The corrected residual true value, For the task The corrected residual prediction value.
[0192] Step 15: Estimation of task training uncertainty. It includes variables such as temperature, humidity, wind speed, and precipitation. The exponential moving average of the historical average loss for each task is calculated. And calculate the ratio of the current instantaneous loss to the historical average loss. To measure the training difficulty and uncertainty of the task:
[0193] ,
[0194] ,
[0195] in, For the task At any moment The ratio of the current instantaneous loss to the historical average loss. For the task At any moment Historical average loss For the task At any moment Instantaneous loss, It is a constant. As a smoothing factor, For the task At any moment The historical average loss.
[0196] Step 16: Adaptive calculation of dynamic loss weights. Based on ratio. Calculate the dynamic weights of each task. By using Softmax normalization, tasks that are difficult to learn are given greater weight:
[0197] ,
[0198] in, For the task At any moment Dynamic weights, For index, For traversing the index of meteorological elements, For the task At any moment The ratio of the current instantaneous loss to the historical average loss.
[0199] Step 17: Construct the final multi-task joint loss function We optimize the model parameters by weighted summation of the losses from all subtasks and using backpropagation:
[0200] ,
[0201] in, For the task The true systematic error between meteorological observations and the original numerical weather prediction grid forecasts. For the task The corrected residual prediction value.
[0202] Step 18: Output corrected residuals using the trained model. , Add it to the baseline forecast value The final high-precision meteorological correction sequence was obtained. :
[0203] .
[0204] This embodiment proposes a multi-task joint learning strategy based on hard parameter sharing and dynamic uncertainty weighting. The corrections for four elements—temperature, humidity, wind speed, and precipitation—are treated as related tasks. The model shares a feature extraction layer and a fusion layer at the bottom layer, outputting the residuals of each element only at the output through an independent multilayer perceptron regression head. To address the varying convergence difficulties of different meteorological variables, a weighted joint loss function based on dynamic uncertainty is designed. By calculating the relative loss ratio, the weights of each task in the total loss are dynamically adjusted in real time. This solves the problem of meteorological data violating physical laws caused by traditional independent modeling. Through shared representations, the model can uncover the inherent physical dependencies between variables, ensuring the consistency of the generated correction data in terms of physical logic and avoiding input noise misleading downstream load forecasts.
[0205] III. Obtaining power load forecasting results based on meteorological correction sequences and historical conformity data.
[0206] Step 19: The corrected meteorological sequence As an exogenous variable, it is related to historical load data. Input the power load forecasting model (such as Autoformer) and output the power load forecast results for future time periods. .
[0207] ,
[0208] in, This represents the power load forecasting model.
[0209] Example 2
[0210] This embodiment provides a power load forecasting device based on multimodal data fusion, including:
[0211] Multimodal meteorological data acquisition module: used to acquire the multimodal meteorological data to be measured in the target area, wherein the multimodal meteorological data includes time-consistent meteorological station observation sequence data and numerical weather prediction grid image sequence data; the meteorological station observation sequence data and the numerical weather prediction grid image sequence data contain a variety of corresponding meteorological elements;
[0212] The multi-element meteorological correction residual acquisition module is used to input the multi-modal meteorological data to be measured into a pre-trained multi-modal meteorological correction model to obtain multi-element meteorological correction residuals. The multi-modal meteorological correction model extracts temporal features based on meteorological station observation sequence data and spatial features based on numerical weather prediction grid image sequence data. It then deeply fuses the temporal and spatial features through a bidirectional cross-attention mechanism to obtain fused features. Based on these fused features, it predicts the correction residuals of each meteorological element using a multi-task learning framework and adaptively optimizes the multi-task joint loss function using a dynamic uncertainty weighting strategy.
[0213] Meteorological correction sequence acquisition module: used to acquire meteorological correction sequences based on numerical weather prediction grid image sequence data and multi-factor meteorological correction residuals;
[0214] The power load forecasting result acquisition module is used to input meteorological correction sequences and historical load data into the power load forecasting model and output the power load forecasting results for future times.
[0215] The device provided in this embodiment can execute the power load forecasting method based on multimodal data fusion provided in any step of Embodiment 1, and has the corresponding functional modules and beneficial effects of the execution method.
[0216] Example 3
[0217] This embodiment provides a computer storage medium storing a computer program. When the computer program is executed by a processor, it implements the power load forecasting method based on multimodal data fusion as provided in any step of Embodiment 1.
[0218] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0219] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0220] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0221] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0222] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A power load forecasting method based on multimodal data fusion, characterized in that, include: Acquire multimodal meteorological data for the target area, wherein the multimodal meteorological data includes time-consistent meteorological station observation sequence data and numerical weather prediction grid image sequence data; the meteorological station observation sequence data and the numerical weather prediction grid image sequence data contain multiple corresponding meteorological elements; The multimodal meteorological data to be tested is input into a pre-trained multimodal meteorological correction model to obtain multi-element meteorological correction residuals. The multimodal meteorological correction model extracts temporal features from meteorological station observation sequence data and spatial features from numerical weather prediction grid image sequence data. It then deeply fuses the temporal and spatial features using a bidirectional cross-attention mechanism to obtain fused features. Based on these fused features, a multi-task learning framework is used to predict the correction residuals for each meteorological element. Finally, a dynamic uncertainty weighting strategy is employed to adaptively optimize the multi-task joint loss function. Meteorological correction sequences are obtained based on gridded image sequence data from numerical weather prediction and multi-element meteorological correction residuals. The meteorological correction sequence and historical load data are input into the power load forecasting model, which outputs the power load forecast results for future times.
2. The power load forecasting method based on multimodal data fusion according to claim 1, characterized in that, The structure of the multimodal meteorological correction model includes a multimodal feature extraction module, a bidirectional cross-modal attention fusion module, a shared feature fusion module, and a multi-task residual prediction module, which are sequentially connected. The multimodal feature extraction module includes a parallel temporal feature extraction unit and a spatial feature extraction unit; The temporal feature extraction unit includes a serial Autoformer encoder and an autocorrelation mechanism, and the spatial feature extraction unit includes a serial CNN encoder and a global average pooling layer. The bidirectional cross-modal attention fusion module includes a serial bidirectional attention extraction unit and a cross-modal attention fusion unit; The attention extraction unit includes a parallel time-guided spatial attention stream and a spatially guided time attention stream; The cross-modal attention fusion unit includes a serial multi-head attention mechanism, a splicing layer, a feedforward neural network, and layer normalization. The shared feature fusion module includes a global average pooling layer; The multi-task residual prediction module includes independent multilayer sensor regression heads for each meteorological element.
3. The power load forecasting method based on multimodal data fusion according to claim 2, characterized in that, The step of inputting the multimodal meteorological data to be measured into a pre-trained multimodal meteorological correction model to obtain multi-factor meteorological correction residuals includes: The time-series feature extraction unit decomposes the meteorological station observation sequence data through the Autoformer encoder to extract the seasonal term, and then aggregates the time delay information of the seasonal term through the autocorrelation mechanism to obtain the time-series feature embedding. The spatial feature extraction unit performs multi-scale convolutional encoding on the numerical weather forecast grid image sequence data through a CNN encoder to extract multi-scale features. The multi-scale features are then aligned with the temporal feature embedding by a global average pooling layer to obtain the spatial feature embedding. Through a bidirectional attention extraction unit, a time-guided spatial attention stream is constructed using temporal feature embedding as the query and spatial feature embedding as the key and value; and a space-guided temporal attention stream is constructed using spatial feature embedding as the query and temporal feature embedding as the key and value. A multi-head attention mechanism is used to perform bidirectional feature interaction between spatial attention stream and temporal attention stream to obtain the interactive features. The interactive features are then concatenated through a concatenation layer and nonlinearly fused through a feedforward neural network and layer normalization to generate a fused feature sequence. The fused feature sequence is subjected to global average pooling by the shared feature fusion module to obtain a global shared feature representation; The multi-task residual prediction module inputs the globally shared feature representation into the independent multilayer perceptron regression head corresponding to each meteorological element, and maps it to obtain the corrected residuals of each meteorological element.
4. The power load forecasting method based on multimodal data fusion according to claim 3, characterized in that, The method of using a time-series feature extraction unit to decompose meteorological station observation sequence data through an Autoformer encoder, and then aggregating seasonal terms based on an autocorrelation mechanism, includes: Perform sequence decomposition on the input meteorological station observation sequence data and extract the trend term: ， in, For trend items, The input is the meteorological station observation sequence. For fill operation, For average pooling; The seasonal term is obtained by comparing the input meteorological station observation sequence data with the trend term: ， in, For seasonal items; The seasonal terms are mapped to the query matrix using a linear projection matrix. Key matrix Sum matrix : ， ， ， in, , , It is a linear projection matrix; The query matrix is transformed using the Fast Fourier Transform. Key matrix Transform to the frequency domain and calculate the autocorrelation of the seasonal term: ， in, For periodic correlation strength, For inverse fast Fourier transform, For Fast Fourier Transform, Indicates time delay. This represents element-wise multiplication. Indicates complex conjugation; Multiple time delays are selected based on the autocorrelation of the seasonal term, and attention weights are generated based on the autocorrelation to adjust the value matrix. Perform the corresponding time-delay rolling aggregation to obtain the autocorrelation output: ， in, For the first The autocorrelation output of the layer, For time-delay index, For the amount of time delay, For scrolling, For normalized probability, Indicates the first A time delay; The autocorrelation output of the current layer and the seasonal term input of the previous layer are subjected to sequence decomposition, and nonlinear mapping is performed through a feedforward network to obtain the temporal feature embedding of the current layer, where the th... The layer operations are as follows: ， ， in, The output is the seasonal term after further sequence decomposition. The output is the trend term after further sequence decomposition. Enter the seasonal item for the previous level. For sequence decomposition, For the first Embedding of temporal features from layer output, It is a feedforward network; The temporal feature embedding is iterated through multiple layers to obtain the final temporal feature embedding: ， in, The temporal feature embedding is the output of the temporal feature extraction unit. This indicates that the embedding of temporal features is performed Layer iteration, The dimension is The set of real numbers, For batch size, For sequence length, For feature dimensions.
5. The power load forecasting method based on multimodal data fusion according to claim 3, characterized in that, The method of performing multi-scale convolutional encoding on numerical weather forecast grid image sequence data using a CNN encoder, and then aligning the dimensions of multi-scale features with temporal feature embeddings using a global average pooling layer, includes: pass Stacked convolutional layers are used to extract deep spatial semantic features from the input numerical weather forecast grid image sequence data, resulting in a high-level feature map containing rich spatial structure information. The layer 1 is... The layer operations are as follows: ， in, For the first The convolutional features output by the layer, For the first The convolutional features output by the layer, This represents a two-dimensional convolution operation. and For learnable parameters, For batch normalization, For activation functions; Compressed features are obtained by performing global average pooling on the high-level feature maps using a global average pooling layer. ， in, For the first High-level feature maps output by the layer. For compression features, For global average pooling, The height of the high-level feature map. The width of the high-level feature map. This is the high-level feature map output by the last convolutional layer. For height indexing, For width index; By mapping the compressed features to the latent spatial dimension using a linear projection matrix, we obtain spatial feature embeddings aligned with the temporal feature dimension. ， in, Embed the spatial features output by the spatial feature extraction unit. It is a linear projection matrix.
6. The power load forecasting method based on multimodal data fusion according to claim 3, characterized in that, Methods for bidirectional feature interaction between spatial attention streams and temporal attention streams through multi-head attention mechanisms include: For time-guided spatial attention flow, the first The calculation method for each attention point is as follows: The input features are mapped to a subspace using a linear projection matrix: ， ， ， in, Embedding of temporal features output by the temporal feature extraction unit. Embed the spatial features output by the spatial feature extraction unit. , , It is a linear projection matrix. The query matrix is constructed based on time-series features. The key matrix is constructed based on spatial feature embedding. The value matrix is constructed based on spatial feature embedding; The attention score is calculated using the scaled dot product and then normalized using Softmax to obtain the weighted output of the corresponding attention head: ， ， in, Attention score for spatial attention flow guided by time. For each head dimension, The weighted output of the attention head in a time-guided spatial attention flow; For spatially guided temporal attention streams, the first The calculation method for each attention point is as follows: The input features are mapped to a subspace using a linear projection matrix: ， ， ， in, The query matrix is constructed based on spatial feature embedding. The key matrix is constructed based on temporal features. The value matrix is constructed based on the embedding of temporal features; The attention score is calculated using the scaled dot product and then normalized using Softmax to obtain the weighted output of the corresponding attention head: ， ， in, Attention score for spatially guided temporal attention flow. The weighted output of the attention head in a spatially guided temporal attention stream; By concatenating and linearly projecting all attention heads generated by the two attention streams, the interactive features are obtained: ， ， in, For time-guided spatial attention flow output, For splicing, The first in the spatial attention stream guided by time The output of each attention head, Output weights for attention. For spatially guided temporal attention stream output, The first in the temporal attention stream guided by space The output of each attention head.
7. The power load forecasting method based on multimodal data fusion according to claim 1, characterized in that, Methods for adaptively optimizing the joint loss function of multiple tasks using a dynamic uncertainty weighting strategy include: The dynamic weights of each meteorological element learning task are obtained based on the historical average loss and the current instantaneous loss of each meteorological element learning task: ， ， ， in, For meteorological element learning task index, For the task At any moment Dynamic weights, For the index, For traversing the index of meteorological elements, It is a set of meteorological elements, including continuous variables and sparse variables. For the task At any moment The ratio of the current instantaneous loss to the historical average loss. For the task At any moment The ratio of the current instantaneous loss to the historical average loss. For the task At any moment Historical average loss For the task At any moment Instantaneous loss, It is a constant. As a smoothing factor, For the task At any moment The historical average loss; Based on the continuous variable learning task loss, the sparse variable learning task loss, and the dynamic weights of each meteorological element learning task, a multi-task joint loss function is constructed, and the model parameters are optimized through backpropagation. The expression of the multi-task joint loss function is as follows: ， ， ， ， in, This represents the joint loss function for multiple tasks. For the task The true systematic error between meteorological observations and the original numerical weather prediction grid forecasts. For the task The corrected residual predictions For continuous variable learning task loss, Learning task loss for sparse variables, Duration For time step indexing, the continuous variable features include temperature. ,humidity Wind speed The sparse variable elements include precipitation. .
8. The power load forecasting method based on multimodal data fusion according to claim 1, characterized in that, Methods for obtaining meteorological correction sequences based on numerical weather prediction grid image sequence data and multi-factor meteorological correction residuals include: The multi-factor meteorological correction residuals are superimposed with the baseline forecast values to obtain the meteorological correction sequence: ， in, For meteorological correction sequences, As the baseline forecast value, Correcting residuals for multi-factor meteorological data.
9. A power load forecasting device based on multimodal data fusion, characterized in that, include: Multimodal meteorological data acquisition module: used to acquire the multimodal meteorological data to be measured in the target area, wherein the multimodal meteorological data includes time-consistent meteorological station observation sequence data and numerical weather prediction grid image sequence data; the meteorological station observation sequence data and the numerical weather prediction grid image sequence data contain a variety of corresponding meteorological elements; The multi-element meteorological correction residual acquisition module is used to input the multi-modal meteorological data to be measured into a pre-trained multi-modal meteorological correction model to obtain multi-element meteorological correction residuals. The multi-modal meteorological correction model extracts temporal features based on meteorological station observation sequence data and spatial features based on numerical weather prediction grid image sequence data. It then deeply fuses the temporal and spatial features through a bidirectional cross-attention mechanism to obtain fused features. Based on these fused features, it predicts the correction residuals of each meteorological element using a multi-task learning framework and adaptively optimizes the multi-task joint loss function using a dynamic uncertainty weighting strategy. Meteorological correction sequence acquisition module: used to acquire meteorological correction sequences based on numerical weather prediction grid image sequence data and multi-factor meteorological correction residuals; The power load forecasting result acquisition module is used to input meteorological correction sequences and historical load data into the power load forecasting model and output the power load forecasting results for future times.
10. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the power load forecasting method based on multimodal data fusion as described in any one of claims 1-8.