New energy output prediction method and system based on LSTM-Transform hybrid architecture
By processing meteorological data using the LSTM-Transformer hybrid architecture, extracting cloud movement features and constructing a meteorological node graph structure, the problem of insufficient accuracy of traditional regression models in predicting new energy output is solved, achieving higher prediction accuracy and generalization ability.
Patent Information
- Application Number
- CN202511579108.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-31
- Publication Date
- 2026-02-13
AI Technical Summary
Traditional regression models cannot effectively capture complex nonlinearity, time-varying nature, and coupling in the prediction of new energy output, resulting in insufficient prediction accuracy and poor generalization ability.
A hybrid LSTM-Transformer architecture is adopted, which uses Transformer network and LSTM network to process the long and short time series features of meteorological data respectively. ConvLSTM and optical flow method are combined to extract cloud system motion features, construct meteorological node graph structure and perform graph embedding, and generate continuous weather forecasts to predict the output of new energy.
It improves the accuracy and generalization ability of new energy output prediction, and can automatically learn nonlinear relationships and implicit patterns to adapt to the dynamic changes of multi-source spatiotemporal data.
Smart Images
Figure CN121525936A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of new energy output prediction, and more particularly relates to a new energy output prediction method and system based on an LSTM-Transformer hybrid architecture. BACKGROUND
[0002] In the prior art, regression methods have been widely used in scientific research, engineering applications and intelligent systems as an important means of data modeling and prediction. Traditional regression analysis mainly includes linear regression, multiple regression and polynomial regression, etc. The basic idea is to establish a functional relationship between input variables and output variables to realize the prediction of unknown results. Due to the simple principle and small amount of calculation, this kind of method has played an important role in early data analysis and prediction tasks, such as using linear regression to fit the relationship between stress and displacement in geological exploration, using multiple regression to model the dependence relationship between load and meteorological factors in power systems, and using regression methods to predict price trends in economic research. However, with the continuous increase of data size and complexity, traditional regression models have gradually exposed obvious limitations. First, linear or low-order polynomial models assume that the relationship between variables is relatively simple, but in real applications, there are often high nonlinearity, time-varying and coupling between data, and traditional methods cannot effectively capture complex dynamic characteristics, resulting in insufficient prediction accuracy. Second, regression analysis relies on prior feature selection and model assumptions. If the input data contains noise, missing values or high dimensions, it will often lead to overfitting or underfitting, seriously affecting the generalization ability of the model.
[0003] Therefore, it is urgent to propose a more adaptive and intelligent technical solution. SUMMARY
[0004] To solve the above technical problems, the application provides a new energy output prediction method based on an LSTM-Transformer hybrid architecture, which comprises: Step 101, acquiring multi-source meteorological data, processing the time series observation data features in the meteorological data through a Transformer network and an LSTM network, and fusing the hidden representations of the Transformer network and the LSTM network at each time, to generate fused hidden representations; Step 102, combining a ConvLSTM network with an optical flow method to extract cloud system motion features of remote sensing images in the meteorological data, and according to the cloud system motion features, aligning the squall line / rain belt deformation through deformable convolution, and then extracting squall line / rain belt deformation features; Step 103, according to the fused hidden representation, cloud system movement characteristics and numerical weather forecast, a meteorological node graph structure is constructed through a dynamic graph neural network, and a graph embedding of the meteorological node graph structure is extracted; Step 104, taking the graph embedding as the leading input, the fused hidden representation, the cloud system movement characteristics and the squall line / rain band deformation characteristics as auxiliary inputs, inputting into a task head to generate a predicted continuous weather forecast, and comparing the predicted continuous weather forecast with historical weather data to predict new energy output under corresponding weather.
[0005] Further, the time series observation data features in the meteorological data are processed through the Transformer network and the LSTM network respectively, including: modeling the long time scale sequence in the meteorological data by the Transformer network according to a slow change window; and modeling the short time scale sequence in the meteorological data by the LSTM network according to a fast change window.
[0006] Further, the outputs of the Transformer network and the LSTM network are fused at each time to generate fused time series features, including: setting a cross-scale fusion gate to dynamically weight fuse the outputs of the Transformer network and the LSTM network, and adjusting the contribution proportion of the Transformer network and the LSTM network according to an extreme event.
[0007] Further, the cross-scale fusion gate includes: , wherein, is the time fused hidden representation, is a Sigmoid function, is a weight of the Transformer network hidden representation, is the hidden representation of the Transformer network, is a weight of the LSTM network hidden representation, is the hidden representation of the LSTM network, is a bias term, is an element-wise multiplication.
[0008] Further, before step 102, a terrain forced attention is further set to enhance the expression of meteorological features on the windward and leeward surfaces, and the terrain forced attention is multiplied with the image features of the remote sensing image.
[0009] Further, the terrain forced attention includes: , wherein, is a remote sensing image position The terrain there compels attention. Scaling factor Location of remote sensing images Topographic gradient at the location, For time Location of remote sensing images The wind speed vector at that location.
[0010] Furthermore, in the meteorological node diagram structure, the nodes represent: meteorological observation locations; The edge weights are: , in, For the first The node to the first The edge weights of each node. For the first Features of each node With the Features of each node cosine similarity, The weight of water vapor flux intensity. For the first The node to the first Water vapor flux intensity at each node As a weight for geographical distance, For the first The node to the first Geographical distance between nodes For spatial attenuation scale, The weights of the changes For the first The node to the first The available potential energy difference for convection at each node For the first The node to the first The wind shear variation at each node is poor.
[0011] This invention also proposes a new energy output prediction system based on an LSTM-Transformer hybrid architecture, comprising: The fusion module is used to acquire multi-source meteorological data. It processes the time-series observation data features in the meteorological data through Transformer network and LSTM network respectively, and fuses the hidden representations of Transformer network and LSTM network at time step to generate fused hidden representation. The alignment module is used to extract cloud movement features from remote sensing images in meteorological data by combining ConvLSTM network with optical flow method. Based on the cloud movement features, the squall line / rainband deformation is aligned by deformable convolution and then the squall line / rainband deformation features are extracted. The graph embedding module is used to construct a meteorological node graph structure through a dynamic graph neural network based on the fused hidden representation, cloud movement characteristics, and numerical weather forecast, and to extract the graph embedding of the meteorological node graph structure. The prediction module uses graph embedding as the primary input, and the fused hidden representation, cloud movement features, and squall line / rainband deformation features as auxiliary inputs. These are input into the task header to generate a predicted continuous weather forecast. The predicted continuous weather forecast is then compared with historical weather data to predict the output of new energy sources under the corresponding weather conditions.
[0012] Furthermore, the time-series observation data features in meteorological data are processed using Transformer networks and LSTM networks respectively, including: modeling long-term time-scale sequences in meteorological data segmented by slow-changing windows using Transformer networks; and modeling short-term time-scale sequences in meteorological data segmented by fast-changing windows using LSTM networks.
[0013] Furthermore, the time-series fusion of the outputs of the Transformer network and the LSTM network to generate fused temporal features includes: setting a cross-scale fusion gating to dynamically weight and fuse the outputs of the Transformer network and the LSTM network, and adjusting the contribution ratio of the Transformer network and the LSTM network according to extreme events.
[0014] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art: This invention's technical solution introduces a regression method to map complex multi-source spatiotemporal data into continuous numerical prediction results. This approach effectively improves the model's fitting accuracy to dynamic processes while fully exploring the inherent correlations within the data. Compared to traditional methods relying on linear models or single empirical formulas, this solution can automatically learn nonlinear relationships and implicit patterns, thereby achieving higher accuracy and generalization capabilities in scenarios such as weather forecasting, environmental monitoring, and energy dispatching. Attached Figure Description
[0015] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention; Figure 2 This is a system structure diagram of Embodiment 2 of the present invention. Detailed Implementation
[0016] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.
[0017] The method provided by this invention can be implemented in a terminal environment that may include one or more of the following components: a processor, a storage medium, and a display screen. The storage medium stores at least one instruction, which is loaded and executed by the processor to implement the method described in the following embodiments.
[0018] A processor may include one or more processing cores. The processor uses various interfaces and lines to connect various parts of the terminal, and performs various functions and processes data by running or executing instructions, programs, code sets or instruction sets stored in the storage medium, and by calling data stored in the storage medium.
[0019] Storage media can include random access memory (RAM) or read-only memory (ROM). Storage media can be used to store instructions, programs, code, code sets, or instructions.
[0020] The display screen is used to show the user interface of each application.
[0021] In addition, those skilled in the art will understand that the structure of the terminal described above does not constitute a limitation on the terminal. The terminal may include more or fewer components, or combine certain components, or have different component arrangements. For example, the terminal may also include radio frequency circuits, input units, sensors, audio circuits, power supplies, and other components, which will not be described in detail here.
[0022] Example 1 like Figure 1 As shown, this embodiment proposes a new energy output prediction method based on an LSTM-Transformer hybrid architecture. The new energy sources in this embodiment include photovoltaic power generation and / or wind power generation, including: Step 101: Acquire multi-source meteorological data, process the time-series observation data features in the meteorological data using Transformer network and LSTM network respectively, and fuse the hidden representations of Transformer network and LSTM network time by time to generate the fused hidden representation; Preferably, multi-source meteorological data can include: time-series observation data characteristics (e.g., meteorological observation stations / buoys), remote sensing images (stationary meteorological satellite cloud images, radar mosaics, etc.), static surface factors (elevation DEM, topographic slope / aspect, underlying surface type, etc.), numerical weather prediction, and disaster labels (typhoon track / intensity, ice thickness, extreme threshold events, etc.).
[0023] Specifically, the processing of time-series observation data features in meteorological data using Transformer networks and LSTM networks includes: modeling long-term time-scale sequences in meteorological data segmented by slow-changing windows using Transformer networks; and modeling short-term time-scale sequences in meteorological data segmented by fast-changing windows using LSTM networks.
[0024] Preferably, the timeline is divided into a slow-change window of 5–10 days and a fast-change window of 1–3 hours: For slowly varying sequences, sparse attention (such as log-sparse) is used to model low-frequency modalities such as seasons and weekends / weekdays through Long-Horizon Transformer (LHT); For rapidly changing sequences, a multi-layer gating method is used to capture high-frequency perturbations at the convective and boundary layers by employing a Gated LSTM Stack (GLS).
[0025] Specifically, the time-series fusion of the outputs of the Transformer network and the LSTM network to generate fused temporal features includes: setting a cross-scale fusion gating to dynamically weight the fusion of the outputs of the Transformer network and the LSTM network, and adjusting the contribution ratio of the Transformer network and the LSTM network according to extreme events.
[0026] Specifically, setting cross-scale fusion gating includes: , in, For time Hidden representation after temporal fusion For the Sigmoid function, The weights of the hidden representation for the Transformer network, For the hidden representation of the Transformer network, The weights of the hidden representation for the LSTM network. For the hidden representation of the LSTM network, For bias terms, This is element-wise multiplication.
[0027] Step 102: By combining ConvLSTM network with optical flow method, cloud movement features of remote sensing images in meteorological data are extracted. Based on cloud movement features, squall line / rainband deformation is aligned by deformable convolution and then squall line / rainband deformation features are extracted. Preferably, the remote sensing image (e.g., satellite cloud image) is used to extract motion texture using ConvLSTM and superimposed with optical flow to obtain cloud system motion features (cloud system displacement field); Deformable convolution guided by cloud system motion features is used to align squall line / rainband deformation and extract squall line / rainband deformation features to reduce blurring caused by translation / deformation.
[0028] Specifically, before step 102, the steps also include: setting terrain forced attention to enhance the expression of meteorological features on the windward and leeward sides, and multiplying the terrain forced attention with the image features of the remote sensing image.
[0029] Specifically, terrain-forced attention includes: , in, Location of remote sensing images The terrain there compels attention. Scaling factor Location of remote sensing images Topographic gradient at the location, For time Location of remote sensing images The wind speed vector at that location.
[0030] Preferably, regarding the scaling factor The value of is given in the following example in this embodiment: = 0.5: The effect of orographic lifting / sinking is weakened, and the moderating effect of wind on precipitation or icing is relatively mild.
[0031] = 1.0: Baseline value, neither magnified nor over-compressed, maintaining the natural proportions of terrain and wind action.
[0032] = 2.0: Enhances the windward slope lifting effect and the leeward slope sinking effect, making the model more sensitive.
[0033] = 5.0: Significantly amplifies the windward / leeward difference in terrain, typically used to simulate local weather under strong terrain control (e.g., blizzards in mountainous areas, icing on windward slopes).
[0034] Step 103: Based on the fused hidden representation, cloud system motion characteristics and numerical weather forecast, a meteorological node graph structure is constructed using a dynamic graph neural network (DGN), and the graph embedding (spatial-temporal features after message passing of nodes / edges) of the meteorological node graph structure is extracted. Preferably, node characteristics are constructed by providing physical quantities such as wind field, temperature field, and humidity field through NWP (Numerical Weather Prediction).
[0035] Step 104: Using graph embedding as the primary input, and the fused hidden representation, cloud motion features, and squall line / rainband deformation features as auxiliary inputs, the data is input into the task header to generate a predicted continuous weather forecast. The predicted continuous weather forecast is then compared with historical weather data to predict the output of new energy sources under the corresponding weather conditions.
[0036] Preferably, the fused hidden representation, cloud motion features, and squall line / rainband deformation features are mapped to the same dimension as the graph embedding through 1x1 convolution or MLP, and then fused into the graph embedding through multimodal fusion technology.
[0037] Preferably, the graph embedding is used as the primary input, and the fused hidden representation, cloud motion features, and squall line / rainband deformation features are used as auxiliary inputs. These are input into a fully connected regression layer, and the output generates a predicted continuous weather forecast.
[0038] Specifically, in the meteorological node diagram structure, the nodes represent: meteorological observation locations; The edge weights are: , in, For the first The node to the first The edge weights of each node. For the first Features of each node With the Features of each node cosine similarity, The weight of water vapor flux intensity. For the first The node to the first Water vapor flux intensity at each node As a weight for geographical distance, For the first The node to the first Geographical distance between nodes For spatial attenuation scale, The weights of the changes For the first The node to the first The available potential energy difference for convection at each node For the first The node to the first The wind shear variation at each node is poor.
[0039] Example 2 like Figure 2 As shown, this embodiment proposes a new energy output prediction system based on an LSTM-Transformer hybrid architecture, including: The fusion module is used to acquire multi-source meteorological data. It processes the time-series observation data features in the meteorological data through Transformer network and LSTM network respectively, and fuses the hidden representations of Transformer network and LSTM network at time step to generate fused hidden representation. Preferably, multi-source meteorological data can include: time-series observation data characteristics (e.g., meteorological observation stations / buoys), remote sensing images (stationary meteorological satellite cloud images, radar mosaics, etc.), static surface factors (elevation DEM, topographic slope / aspect, underlying surface type, etc.), numerical weather prediction, and disaster labels (typhoon track / intensity, ice thickness, extreme threshold events, etc.).
[0040] Specifically, the processing of time-series observation data features in meteorological data using Transformer networks and LSTM networks includes: modeling long-term time-scale sequences in meteorological data segmented by slow-changing windows using Transformer networks; and modeling short-term time-scale sequences in meteorological data segmented by fast-changing windows using LSTM networks.
[0041] Preferably, the timeline is divided into a slow-change window of 5–10 days and a fast-change window of 1–3 hours: For slowly varying sequences, sparse attention (such as log-sparse) is used to model low-frequency modalities such as seasons and weekends / weekdays through Long-Horizon Transformer (LHT); For rapidly changing sequences, a multi-layer gating method is used to capture high-frequency perturbations at the convective and boundary layers by employing a Gated LSTM Stack (GLS).
[0042] Specifically, the time-series fusion of the outputs of the Transformer network and the LSTM network to generate fused temporal features includes: setting a cross-scale fusion gating to dynamically weight the fusion of the outputs of the Transformer network and the LSTM network, and adjusting the contribution ratio of the Transformer network and the LSTM network according to extreme events.
[0043] Specifically, setting cross-scale fusion gating includes: , in, For time Hidden representation after temporal fusion For the Sigmoid function, The weights of the hidden representation for the Transformer network, For the hidden representation of the Transformer network, The weights of the hidden representation for the LSTM network. For the hidden representation of the LSTM network, For bias terms, This is element-wise multiplication.
[0044] The alignment module is used to extract cloud movement features from remote sensing images in meteorological data by combining ConvLSTM network with optical flow method. Based on the cloud movement features, the squall line / rainband deformation is aligned by deformable convolution and then the squall line / rainband deformation features are extracted. Preferably, the remote sensing image (e.g., satellite cloud image) is used to extract motion texture using ConvLSTM and superimposed with optical flow to obtain cloud system motion features (cloud system displacement field); Deformable convolution guided by cloud system motion features is used to align squall line / rainband deformation and extract squall line / rainband deformation features to reduce blurring caused by translation / deformation.
[0045] Specifically, before step 102, the steps also include: setting terrain forced attention to enhance the expression of meteorological features on the windward and leeward sides, and multiplying the terrain forced attention with the image features of the remote sensing image.
[0046] Specifically, terrain-forced attention includes: , in, Location of remote sensing images The terrain there compels attention. Scaling factor Location of remote sensing images Topographic gradient at the location, For time Location of remote sensing images The wind speed vector at that location.
[0047] Preferably, regarding the scaling factor The value of is given in the following example in this embodiment: = 0.5: The effect of orographic lifting / sinking is weakened, and the moderating effect of wind on precipitation or icing is relatively mild.
[0048] = 1.0: Baseline value, neither magnified nor over-compressed, maintaining the natural proportions of terrain and wind action.
[0049] = 2.0: Enhances the windward slope lifting effect and the leeward slope sinking effect, making the model more sensitive.
[0050] = 5.0: Significantly amplifies the windward / leeward difference in terrain, typically used to simulate local weather under strong terrain control (e.g., blizzards in mountainous areas, icing on windward slopes).
[0051] The graph embedding module is used to construct a meteorological node graph structure through a dynamic graph neural network (DGN) based on the fused hidden representation, cloud movement characteristics, and numerical weather forecast, and to extract the graph embedding (spatiotemporal features of nodes / edges after message passing) of the meteorological node graph structure. Preferably, node characteristics are constructed by providing physical quantities such as wind field, temperature field, and humidity field through NWP (Numerical Weather Prediction).
[0052] The prediction module uses graph embedding as the primary input, and the fused hidden representation, cloud movement features, and squall line / rainband deformation features as auxiliary inputs. These are input into the task header to generate a predicted continuous weather forecast. The predicted continuous weather forecast is then compared with historical weather data to predict the output of new energy sources under the corresponding weather conditions.
[0053] Preferably, the fused hidden representation, cloud motion features, and squall line / rainband deformation features are mapped to the same dimension as the graph embedding through 1x1 convolution or MLP, and then fused into the graph embedding through multimodal fusion technology.
[0054] Preferably, the graph embedding is used as the primary input, and the fused hidden representation, cloud motion features, and squall line / rainband deformation features are used as auxiliary inputs. These are input into a fully connected regression layer, and the output generates a predicted continuous weather forecast.
[0055] Specifically, in the meteorological node diagram structure, the nodes represent: meteorological observation locations; The edge weights are: , in, For the first The node to the first The edge weights of each node. For the first Features of each node With the Features of each node cosine similarity, The weight of water vapor flux intensity. For the first The node to the first Water vapor flux intensity at each node As a weight for geographical distance, For the first The node to the first Geographical distance between nodes For spatial attenuation scale, The weights of the changes For the first The node to the first The available potential energy difference for convection at each node For the first The node to the first The wind shear variation at each node is poor.
[0056] Example 3 This invention also proposes a storage medium storing multiple instructions for implementing a new energy output prediction based on an LSTM-Transformer hybrid architecture.
[0057] Optionally, in this embodiment, the storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals.
[0058] Optionally, in this embodiment, the storage medium is configured to store program code for performing the method steps of Embodiment 1.
[0059] Example 4 This invention also proposes an electronic device, including a processor and a storage medium connected to the processor. The storage medium stores multiple instructions, which can be loaded and executed by the processor to enable the processor to perform a new energy output prediction based on an LSTM-Transformer hybrid architecture.
[0060] Specifically, the electronic device in this embodiment can be a computer terminal, which may include one or more processors and a storage medium.
[0061] The storage medium can be used to store software programs and modules, such as the new energy power output prediction based on an LSTM-Transformer hybrid architecture in this embodiment of the invention. The processor executes the software programs and modules stored in the storage medium to perform various functional applications and data processing, thus realizing the aforementioned new energy power output prediction based on an LSTM-Transformer hybrid architecture. The storage medium may include high-speed random access storage media, and may also include non-volatile storage media, such as one or more magnetic storage systems, flash memory, or other non-volatile solid-state storage media. In some instances, the storage medium may further include storage media remotely configured relative to the processor, which can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.
[0062] The processor can execute the method steps of Embodiment 1 by calling the information and application stored in the storage medium through the transmission system.
[0063] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0064] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0065] In the several embodiments provided by this invention, it should be understood that the disclosed technical content can be implemented in other ways. The system embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between units or modules, and may be electrical or other forms.
[0066] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0067] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0068] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, optical disks, and other media capable of storing program code.
[0069] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.
Claims
1. A new energy power output prediction method based on an LSTM-Transformer hybrid architecture, characterized in that, include: Step 101: Acquire multi-source meteorological data, process the time-series observation data features in the meteorological data using Transformer network and LSTM network respectively, and fuse the hidden representations of Transformer network and LSTM network time by time to generate the fused hidden representation; Step 102: By combining ConvLSTM network with optical flow method, cloud movement features of remote sensing images in meteorological data are extracted. Based on cloud movement features, squall line / rainband deformation is aligned by deformable convolution and then squall line / rainband deformation features are extracted. Step 103: Based on the fused hidden representation, cloud system motion characteristics, and numerical weather forecast, a meteorological node graph structure is constructed using a dynamic graph neural network, and the graph embedding of the meteorological node graph structure is extracted. Step 104: Using graph embedding as the primary input, and the fused hidden representation, cloud motion features, and squall line / rainband deformation features as auxiliary inputs, the data is input into the task header to generate a predicted continuous weather forecast. The predicted continuous weather forecast is then compared with historical weather data to predict the output of new energy sources under the corresponding weather conditions.
2. The renewable energy output prediction method based on an LSTM-Transformer hybrid architecture as described in claim 1, characterized in that, The characteristics of time-series observation data in meteorological data are processed by Transformer network and LSTM network respectively. The process includes: modeling long-term time-scale sequences of meteorological data divided into slow-changing windows using Transformer network; and modeling short-term time-scale sequences of meteorological data divided into fast-changing windows using LSTM network.
3. The renewable energy output prediction method based on an LSTM-Transformer hybrid architecture as described in claim 2, characterized in that, The time-series fusion of the outputs of the Transformer network and the LSTM network is performed by setting a cross-scale fusion gating to dynamically weight the outputs of the Transformer network and the LSTM network, and adjusting the contribution ratio of the Transformer network and the LSTM network according to extreme events.
4. The renewable energy output prediction method based on an LSTM-Transformer hybrid architecture as described in claim 3, characterized in that, Setting cross-scale fusion gating includes: in, For time Hidden representation after temporal fusion For the Sigmoid function, The weights of the hidden representation for the Transformer network, For the hidden representation of the Transformer network, The weights of the hidden representation for the LSTM network. For the hidden representation of the LSTM network, For bias terms, This is element-wise multiplication.
5. The renewable energy output prediction method based on an LSTM-Transformer hybrid architecture as described in claim 1, characterized in that, Before step 102, the steps also include: setting terrain forced attention to enhance the expression of meteorological features on the windward and leeward sides, and multiplying the terrain forced attention with the image features of the remote sensing image.
6. The renewable energy output prediction method based on an LSTM-Transformer hybrid architecture as described in claim 5, characterized in that, Terrain-forced attention includes: in, Location of remote sensing images The terrain there compels attention. Scaling factor Location of remote sensing images Topographic gradient at the location, For time Location of remote sensing images The wind speed vector at that location.
7. The renewable energy output prediction method based on an LSTM-Transformer hybrid architecture as described in claim 1, characterized in that, In the meteorological node diagram structure, nodes represent meteorological observation locations. The edge weights are: in, For the first The node to the first The edge weights of each node. For the first Features of each node With the Features of each node cosine similarity, The weight of water vapor flux intensity. For the first The node to the first Water vapor flux intensity at each node As a weight for geographical distance, For the first The node to the first Geographical distance between nodes For spatial attenuation scale, The weights of the changes For the first The node to the first The available potential energy difference for convection at each node For the first The node to the first The wind shear variation at each node is poor.
8. A new energy output prediction system based on an LSTM-Transformer hybrid architecture, characterized in that, include: The fusion module is used to acquire multi-source meteorological data. It processes the time-series observation data features in the meteorological data through Transformer network and LSTM network respectively, and fuses the hidden representations of Transformer network and LSTM network at time step to generate fused hidden representation. The alignment module is used to extract cloud movement features from remote sensing images in meteorological data by combining ConvLSTM network with optical flow method. Based on the cloud movement features, the squall line / rainband deformation is aligned by deformable convolution and then the squall line / rainband deformation features are extracted. The graph embedding module is used to construct a meteorological node graph structure through a dynamic graph neural network based on the fused hidden representation, cloud movement characteristics, and numerical weather forecast, and to extract the graph embedding of the meteorological node graph structure. The prediction module uses graph embedding as the primary input, and the fused hidden representation, cloud movement features, and squall line / rainband deformation features as auxiliary inputs. These are input into the task header to generate a predicted continuous weather forecast. The predicted continuous weather forecast is then compared with historical weather data to predict the output of new energy sources under the corresponding weather conditions.
9. A new energy output prediction system based on an LSTM-Transformer hybrid architecture as described in claim 8, characterized in that, The characteristics of time-series observation data in meteorological data are processed by Transformer network and LSTM network respectively. The process includes: modeling long-term time-scale sequences of meteorological data divided into slow-changing windows using Transformer network; and modeling short-term time-scale sequences of meteorological data divided into fast-changing windows using LSTM network.
10. A new energy output prediction system based on an LSTM-Transformer hybrid architecture as described in claim 9, characterized in that, The time-series fusion of the outputs of the Transformer network and the LSTM network is performed by setting a cross-scale fusion gating to dynamically weight the outputs of the Transformer network and the LSTM network, and adjusting the contribution ratio of the Transformer network and the LSTM network according to extreme events.