A variable scale seamless power prediction method and system based on a meteorological large model

Abstract

The application discloses a variable-scale seamless power prediction method and system based on a meteorological large model, and the method comprises the following steps: inputting a prediction task into a multi-scale power prediction model, decomposing the prediction task, and obtaining task components of different frequencies; obtaining a meteorological initial field, inputting the meteorological initial field and the task components into a meteorological prediction large model, and obtaining a meteorological element prediction set corresponding to each task component; inputting the meteorological element prediction set into the multi-scale power prediction model, and outputting a power prediction result of each task component; and fusing and correcting the power prediction result of each task component, eliminating the jump of different task components on a power curve, and obtaining a continuous power prediction curve. The variable-scale seamless power prediction method and system based on the meteorological large model are proposed, the meteorological prediction large model and the multi-scale power prediction model are fused, and then the self-adaptive capability under the variable scale and the seamless continuity of the prediction result are realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a variable-scale seamless power prediction method and system based on a large meteorological model. Background Technology

[0002] High-precision power generation forecasting is a key technology supporting grid dispatching, electricity market transactions, and system safety operation. With the rapid development of artificial intelligence technology and the big data industry, technologies that predict power generation based on artificial intelligence and combined with meteorological data have been widely used. For example, relying on numerical weather prediction data (NWP) and historical power data, and establishing the mapping relationship between meteorological conditions and power generation through machine learning models (such as gradient boosting trees, LSTM, etc.), is the current mainstream power generation forecasting method.

[0003] However, the above-mentioned power generation prediction methods still have the following problems:

[0004] 1. Traditional NWP lacks high-resolution, high-frequency, and more accurate meteorological elements related to new energy (such as radiation and wind speed). Therefore, the above prediction method excludes meteorological elements that are deeply related to power generation in terms of data source. As a result, the mapping relationship obtained by this method cannot reflect the deep correlation between power generation and meteorological elements. That is, the prediction results obtained by the above prediction method will always be limited to the correlation between power generation and surface meteorological data, and its prediction accuracy cannot be guaranteed.

[0005] 2. Power generation forecasting tasks often involve multi-scale requirements, such as "slope event" forecasting (predicting large fluctuations in power generation over a short period of time) and power generation cycle change forecasting (predicting changes in power generation over a certain period of time). The requirements for meteorological elements for these two tasks are quite different. However, the above forecasting methods have a single working mode and cannot adaptively adjust the model's calculation path, feature processing method, or attention focus according to the task requirements. As a result, when facing multi-scale tasks, it is necessary to stitch together or switch between multiple models to process the task, which is complex to manage and has poor results.

[0006] 3. When dealing with multi-scale tasks, the prediction results obtained by splicing or switching multiple traditional models can only be prediction results at multiple different scales. They cannot be directly applied, lack cross-scale consistency, and cannot directly guide the actual dispatching and market transactions of the power grid. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is to propose a variable-scale seamless power prediction method and system based on a large meteorological model. This method can achieve adaptive capability and seamless continuity of prediction results under variable scales by integrating upstream data sources (large meteorological prediction model) and multi-scale power prediction models.

[0008] One technical solution adopted in this invention is to provide a variable-scale seamless power prediction method based on a large meteorological model, which includes the following steps:

[0009] S1: Input the prediction task into the multi-scale power prediction model, decompose the prediction task, and obtain task components of different frequencies;

[0010] S2: Obtain the initial meteorological field, input the initial meteorological field and task components into the large meteorological prediction model, and obtain the meteorological element prediction set corresponding to each task component;

[0011] S3: Input the meteorological element prediction set into the multi-scale power prediction model and output the power prediction results for each task component;

[0012] S4: Perform fusion correction on the power prediction results of each task component to eliminate jumps in the power curve of different task components and obtain a continuous power prediction curve.

[0013] Furthermore, the decomposition of the prediction task includes the following sub-steps:

[0014] S11: Input prediction task, which includes geographic location, prediction start time, prediction end time and desired time resolution;

[0015] S12: Time feature extraction, calculate the task duration based on the predicted start time and predicted end time;

[0016] S13: Perform frequency domain decomposition and frequency band division on the task duration;

[0017] S14: Calculate the task components based on frequency domain decomposition and frequency band division.

[0018] Furthermore, step S13 includes the following sub-steps:

[0019] S131: Perform frequency domain decomposition of the task duration based on Fourier transform and output the spectral characteristics:

[0020]

[0021] Where S(f) represents the spectral characteristics, Let δ(t) represent the Fourier transform, and let δ(t) represent the time impulse function.

[0022] S132: Extract dominant frequencies from spectral features:

[0023]

[0024] Among them, f d Indicates the dominant frequency;

[0025] S133: Frequency band division based on dominant frequency:

[0026]

[0027] Among them, B j This indicates the frequency range of frequency band j.

[0028] Furthermore, step S14 includes the following sub-steps:

[0029] S141: Calculate the time resolution of task components based on frequency band division:

[0030]

[0031] Where, Δt j Indicates component temporal resolution. ω represents the upper bound frequency of frequency band j. j E represents the weight of frequency band j. j Describes the band energy of band j and E represents the total energy across the entire frequency band, and

[0032] S142: Construct task component description vectors:

[0033]

[0034] Where loc represents geolocation code, P time This indicates the time and location encoding.

[0035] Furthermore, step S2 includes the following sub-steps:

[0036] S21: Obtain the initial meteorological field and spatially encode it.

[0037]

[0038] Where I0 represents the initial meteorological field and L represents the number of vertical layers, H×W represents the spatial grid size, and C represents the number of meteorological elements. CONV3D enc Represents a 3D convolutional encoder, d emb Indicates the embedding dimension;

[0039] S22: The task component description vector is fused as a spatial coding feature of the initial conditional meteorological field to generate conditional meteorological features;

[0040] S23: Based on the fused conditional meteorological characteristics, predict future meteorological elements through spatiotemporal evolution.

[0041] Furthermore, step S22 includes the following sub-steps:

[0042] S221: Project the task component description vectors into the same embedding space as the initial meteorological field.

[0043]

[0044] Among them, M d MLP represents the projected feature vector. proj This represents a multilayer sensing projection network;

[0045] S222: Broadcast the projected feature vector to every spatial grid point:

[0046]

[0047] Where Re represents M d A spatial copying operation that replicates H×W times to form a spatial feature map;

[0048] S223: Fuse the task component features obtained after broadcasting with the spatial coding features of the initial meteorological field:

[0049] F in =g⊙E s +(1-g)⊙M b ;

[0050] g=σ(W g ·[E s M b ]+b g );

[0051] Where g represents the gate value, ranging from [0,1]; σ represents the sigmoid activation function; W g Represents the gate weight matrix; [E s M b ] indicates that the spatial coding features of the initial meteorological field and the task component features after broadcasting are concatenated along the feature dimension; b g ⊙ indicates gating bias; ⊙ indicates element-wise multiplication.

[0052] Furthermore, step S23 includes the following sub-steps:

[0053] S231: Predict time step based on the temporal resolution evolution of task components:

[0054] Δt=min(Δt j ,Δt p );

[0055] Where, Δt p Indicates the maximum step size of the physical constraint;

[0056] S232: Spatiotemporal evolution using 3DTransformer:

[0057] Y t+Δt =Y t +Δt·f θ (Y t ,F in );

[0058] Among them, Y t+Δt Y represents the predicted state at time t+Δt; t f represents the state of the meteorological field at time t; θ Represents the 3DTransformer evolution function;

[0059] S233: Based on the prediction start time and prediction end time, iteratively execute the evolution function to generate a meteorological element prediction set aligned with the task components.

[0060] Furthermore, step S3 includes the following sub-steps:

[0061] S31: Input the meteorological element prediction set and task component description vector into the multi-scale power prediction model, and perform spatial focusing based on the geographical location of the task components:

[0062]

[0063] Where, α xy p represents the spatial attention weights. loc W indicates the geographical location of the task. loc Let b represent the weight matrix. loc Indicates the bias term. Represents the weather field at time t. This represents the meteorological sequence after spatial focusing. This represents the meteorological sequence of grid point (x, y) at time t;

[0064] S32: Divide long task sequences into fixed-length time blocks:

[0065]

[0066] Among them, B jk Let τ represent the k-th weather block. jλ represents the fixed time block length, λ represents the update rate, and : represents the full column slice operation.

[0067] S33: Dynamically select the computation path based on the time block features, and calculate the feature aggregation of all paths:

[0068]

[0069] Among them, h j K represents the aggregated features of task component j. j The total number of blocks represents task component j, MeanPool represents average pooling, and h jk This represents the fusion feature of the k-th time block of task component j;

[0070] S34: Output Task Component Power Prediction:

[0071]

[0072] in, This represents the power prediction sequence for component j. H represents the power component decoder. j Describing component aggregation features, d j This represents the task component description vector. Indicates the concatenation operation, T j The total time length of component j.

[0073] Furthermore, step S4 includes the following sub-steps:

[0074] S41: Time alignment of power predictions for different task components:

[0075]

[0076] in, This represents the aligned power sequence, where Line represents the linear interpolation function, and t... j The timestamp vector representing task component j, t u Indicates a unified timeline;

[0077] S42: Locate the time interval between different task components and mark areas of data transition risk.

[0078]

[0079] Among them, Ω j Indicates the time range of the risk zone. δt represents the end time of task component j, and δt represents the transition bandwidth.

[0080] S43: Optimize the curvature of the risk area to eliminate curvature abrupt changes.

[0081]

[0082] Where i represents the number of iterations, and η represents the learning rate. Represents the loss function. This represents the power sequence of task component j after i+1 iterations. This represents the power sequence of task component j after i iterations;

[0083] S44: Weighted fusion of the power sequences of each task component yields the final power prediction curve:

[0084]

[0085] in, This represents the final power prediction curve, ω. i (t) represents the time decay weight.

[0086] Another approach adopted in this invention is: a variable-scale seamless power prediction system based on a large meteorological model, which includes the following modules:

[0087] The task decomposition module is used to input the prediction task into the multi-scale power prediction model, decompose the prediction task, and obtain task components of different frequencies.

[0088] The weather forecasting module is used to acquire the initial weather field. The initial weather field and task components are input into the large weather forecasting model to obtain the weather element forecast set corresponding to each task component.

[0089] The power prediction module is used to input the meteorological element prediction set into the multi-scale power prediction model and output the power prediction results for each task component.

[0090] The component fusion module is used to fuse and correct the power prediction results of each task component, eliminate the jumps in the power curves of different task components, and obtain a continuous power prediction curve.

[0091] The present invention provides a variable-scale seamless power prediction method based on a meteorological large-scale model, which has at least the following beneficial effects: 1. By using a meteorological large-scale model, customized forecast results are dynamically generated and output according to the specific needs of the downstream "power prediction large-scale model". These outputs not only provide key meteorological elements with high resolution and high accuracy that are not available in traditional NWP data sources, but also improve the accuracy of power prediction; 2. By using the ideas of scale segmentation and path selection, the calculation path and attention focus of the power prediction model are dynamically adjusted to achieve intelligent switching of working modes, ensuring that the best prediction effect can be achieved in the most efficient way under any scenario; 3. Through fusion and correction mechanisms and cross-scale consistency design, the power prediction curve can smoothly transition at any scale transformation and time domain connection point, making the prediction results more conducive to the actual dispatching of the power grid and market transactions. Attached Figure Description

[0092] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0093] Figure 1 This is a flowchart of one embodiment of a variable-scale seamless power prediction method based on a large meteorological model according to the present invention.

[0094] Figure 2 For the present invention Figure 1 The sub-flowchart of step S1.

[0095] Figure 3 For the present invention Figure 2 The sub-flowchart of step S13.

[0096] Figure 4 For the present invention Figure 2 The sub-flowchart of step S14.

[0097] Figure 5 For the present invention Figure 1 The sub-flowchart of step S2.

[0098] Figure 6 For the present invention Figure 5 The sub-flowchart of step S22.

[0099] Figure 7 For the present invention Figure 5 The sub-flowchart of step S23.

[0100] Figure 8 For the present invention Figure 1 The sub-flowchart of step S3 in the middle.

[0101] Figure 9For the present invention Figure 1 Sub-flowchart of step S4

[0102] Figure 10 This is a structural block diagram of one embodiment of a variable-scale seamless power prediction system based on a large meteorological model according to the present invention. Detailed Implementation

[0103] The invention will now be further described with reference to the accompanying drawings.

[0104] Please see Figure 1 This is a flowchart illustrating an embodiment of the deep learning-based artificial intelligence network optimization training method of the present invention. The method includes the following steps:

[0105] S1: Input the prediction task into the multi-scale power prediction model, decompose the prediction task, and obtain task components of different frequencies. In this S1 step, the multi-scale power prediction model does not rigidly execute a single process, but first deeply analyzes the three key characteristics inherent in each prediction task: the prediction time scale, the resolution of the result, and the specific prediction target type. This mechanism enables the model to act like an expert, selecting the most appropriate analysis tools and perspectives for different problems, thereby achieving optimal resource allocation and maximizing prediction performance.

[0106] When the model receives ultra-short-term, high-resolution forecasting instructions such as "predict minute-level slope events," it immediately switches to high-frequency response mode. In this mode, the model prioritizes high-frequency signals representing dramatic fluctuations through frequency decomposition and activates short computational paths designed to capture instantaneous changes. Simultaneously, its focus narrows dramatically, assigning the majority of weights to the most recent input data to ensure a rapid response to sudden market or weather changes. Conversely, if the task is to "predict total power generation over the next few days," the model enters long-term trend analysis mode. It then focuses on analyzing low-frequency signals representing periodic patterns, employing deeper and more complex computational paths to establish long-term dependencies and expanding its attention to the global picture to accurately grasp overall trends.

[0107] For mixed-scale tasks such as routine day-to-day or intraday rolling scheduling, the model demonstrates high flexibility, adopting a parallel working mode. It simultaneously processes high-frequency and low-frequency information, activating parallel hybrid computation paths. One branch focuses on modeling baseline trends, while another captures immediate fluctuations, ultimately intelligently fusing the analysis results near the output layer. Its standard multi-head attention mechanism fully utilizes this mode; different "attention heads" automatically learn and focus on features at different levels and dimensions, achieving comprehensive and detailed modeling of complex tasks. It is through this dynamic and logically rigorous set of association rules that the model truly achieves intelligent switching between working modes, ensuring optimal prediction results in the most efficient way under any scenario.

[0108] Please see Figure 2 This is a sub-flowchart for step S1 of this solution. Step S1 may also include the following sub-steps:

[0109] S11: Input the prediction task, which includes the geographical location, prediction start time, prediction end time, and desired time resolution. Since this scheme requires prediction of power generation based on multiple scales, the prediction task input in step S11 needs to include the geographical location, prediction start time, prediction end time, and desired time resolution of the task. This resolution is actually used to distinguish between short-term and long-term tasks. For example, high resolution corresponds to the short-term "predicting minute-level slope events" mentioned above, while low resolution corresponds to the long-term "predicting the total power generation in the next few days" mentioned above.

[0110] S12: Time feature extraction, calculate the task duration based on the predicted start time and predicted end time;

[0111] S13: Perform frequency domain decomposition and frequency band division of the task duration. This step S13 mainly uses Fourier transform to convert the time-domain prediction task into the frequency domain, thereby enabling the prediction task to be decomposed from a frequency perspective.

[0112] Please see Figure 3 Here is a sub-flowchart for step S13 of this solution. Step S13 may include the following sub-steps:

[0113] S131: Perform frequency domain decomposition of the task duration based on Fourier transform and output the spectral characteristics:

[0114]

[0115] Where S(f) represents the spectral characteristics, Let δ(t) represent the Fourier transform, and let δ(t) represent the time impulse function.

[0116] S132: Extract dominant frequencies from spectral features:

[0117]

[0118] Among them, f d Indicates the dominant frequency;

[0119] S133: Frequency band division based on dominant frequency:

[0120]

[0121] Among them, B j This represents the frequency range of frequency band j. Steps S131-S133 are the specific processes of converting the time-domain prediction task into the frequency domain through Fourier transform and dividing the prediction task according to frequency. This process can generate the frequency domain representation of the prediction task, thereby providing a data basis for distinguishing between short-term and long-term prediction tasks.

[0122] S14: Calculate the task components based on frequency domain decomposition and frequency band division. This S14 step divides the prediction task from a frequency perspective, thereby enabling the subsequent generation of meteorological element prediction sets to better meet the task requirements.

[0123] Please see Figure 4 Here is a sub-flowchart for step S14 of this solution. Step S14 may include the following sub-steps:

[0124] S141: Calculate the time resolution of task components based on frequency band division:

[0125]

[0126] Where, Δt j Indicates component temporal resolution. ω represents the upper bound frequency of frequency band j. j E represents the weight of frequency band j. j Describes the band energy of band j and E represents the total energy across the entire frequency band, and

[0127] S142: Construct task component description vectors:

[0128]

[0129] Where loc represents geolocation code, P timeThis represents the temporal location encoding. S141-S142 above describes the specific process of dividing the prediction task components. This process first calculates the temporal resolution of the task components using frequency bands. This temporal resolution represents the type of prediction task. Then, it combines geographic location encoding and temporal location encoding to construct a task component description vector, ensuring that each task component contains its task type and spatiotemporal information, thus providing a data foundation for subsequent weather and power prediction.

[0130] S2: Obtain the initial meteorological field. Input the initial meteorological field and task components into the large-scale meteorological prediction model to obtain the meteorological element prediction set corresponding to each task component. In the data acquisition and model training phase of this S2 step, the large-scale meteorological prediction model uses the ECMWF (European Centre for Medium-Range Weather Forecasts) ERA5 global atmospheric reanalysis dataset as the training basis. From this high-quality historical dataset spanning decades, core meteorological elements including temperature, wind speed, geopotential, humidity, and solar radiation, which is crucial for new energy sources, are extracted. Through deep learning training on massive historical data, the large-scale meteorological prediction model can autonomously learn and master the complex laws of atmospheric physical evolution, laying the foundation for accurate prediction.

[0131] In the data processing and inference prediction phase, the core of the model is a deep neural network based on an advanced 3D Transformer architecture. When performing real-time prediction (i.e., inference) tasks, it doesn't predict from scratch, but rather uses the latest high-resolution forecast data publicly released by ECMWF as the initial meteorological field for the large-scale meteorological prediction model. Based on this, deeper and more targeted extrapolations and refined calculations are performed to generate customized meteorological forecasts with higher resolution and greater accuracy, specifically optimized for downstream power prediction tasks. This is a key step in achieving results that surpass general forecasts.

[0132] Please see Figure 5 This is a sub-flowchart for step S2 of this solution. Step S2 also includes the following sub-steps:

[0133] S21: Obtain the initial meteorological field and spatially encode it.

[0134]

[0135] Where I0 represents the initial meteorological field and L represents the number of vertical layers, H×W represents the spatial grid size, and C represents the number of meteorological elements. CONV3D enc Represents a 3D convolutional encoder, d emb The embedding dimension is represented by the following step S21: encoding the input initial meteorological field, extracting spatial and channel features, and outputting an embedding representation.

[0136] S22: The task component description vector is fused with the spatial coding features of the initial conditional meteorological field to generate conditional meteorological features. This step S22 effectively fuses the task component description vector with the coding features of the initial meteorological field to generate conditional meteorological features, enabling the meteorological forecasting model to generate customized meteorological forecasts according to different task component requirements.

[0137] Please see Figure 6 Here is a sub-flowchart for step S22 of this solution. Step S22 also includes the following sub-steps:

[0138] S221: Project the task component description vectors into the same embedding space as the initial meteorological field.

[0139]

[0140] Among them, M d MLP represents the projected feature vector. proj This represents a multilayer sensing projection network;

[0141] S222: Broadcast the projected feature vector to every spatial grid point:

[0142]

[0143] Where Re represents M d A spatial copying operation that replicates H×W times to form a spatial feature map;

[0144] S223: Fuse the task component features obtained after broadcasting with the spatial coding features of the initial meteorological field:

[0145] F in =g⊙E s +(1-g)⊙M b ;

[0146] g=σ(W g ·[E s M b +b g );

[0147] Where g represents the gate value, ranging from [0,1]; σ represents the sigmoid activation function; W g Represents the gate weight matrix; [E s M b ] indicates that the spatial coding features of the initial meteorological field and the task component features after broadcasting are concatenated along the feature dimension; b g⊙ represents gating bias; ⊙ represents element-wise multiplication. S221-S223 above describes the specific process of fusing the description vector of the task component with the coding features of the initial meteorological field. Based on this fused feature, a customized meteorological element prediction set that meets the conditions of the task component can be output using the large meteorological prediction model.

[0148] S23: Based on the fused conditional meteorological features, predict future meteorological elements through spatiotemporal evolution. This S23 step predicts the future meteorological field based on the fused conditional features through 3DTransformer spatiotemporal evolution. This step adopts an adaptive time step, dynamically adjusting the prediction step size according to the temporal resolution of the task components, while ensuring physical consistency.

[0149] Please see Figure 7 Here is a sub-flowchart for step S23 of this scheme. This scheme S23 also includes the following sub-steps:

[0150] S231: Predict time step based on the temporal resolution evolution of task components:

[0151] Δt=min(Δt j ,Δt p );

[0152] Where, Δt p Indicates the maximum step size of the physical constraint;

[0153] S232: Spatiotemporal evolution using 3DTransformer:

[0154] Y t+Δt =Y t +Δt·f θ (Y t ,F in );

[0155] Among them, Y t+Δt Y represents the predicted state at time t+Δt; t f represents the state of the meteorological field at time t; θ Represents the 3DTransformer evolution function;

[0156] S233: Based on the prediction start time and prediction end time, iteratively execute the evolution function to generate a meteorological element prediction set aligned with the task components.

[0157] S3: Input the meteorological element forecast set into the multi-scale power prediction model and output the power prediction results for each task component. The scale power prediction model in this S3 step can be understood as a hybrid deep learning network with built-in dynamic routing and multi-scale analysis modules. Its core is not a single, fixed structure, but an intelligent system that can change its internal data flow and processing focus according to task instructions. Its basic framework may be a powerful time-series Transformer model to capture long-term dependencies, but its key innovation lies in the integrated "variable-scale prediction module," which controls the entire model's operation.

[0158] Furthermore, it can dynamically adjust the model's computational path and attention focus through multi-scale partitioning and path selection. For the aforementioned minute-level slope event prediction task, this module allocates more computational resources and attention to the sub-networks processing high-frequency components, and focuses the attention mechanism on data within the most recent time window. Conversely, if the task is to "predict the total power generation in the next few days," the module activates and strengthens the computational path processing low-frequency components, while adjusting the attention mechanism to cover a longer time span and capture the overall trend. This ability to "intelligently" change the working mode within a single model is fundamentally different from the rigid approach of traditional techniques that require splicing or switching multiple independent models.

[0159] Please see Figure 8 This S3 step includes the following sub-steps:

[0160] S31: Input the meteorological element prediction set and task component description vector into the multi-scale power prediction model, and perform spatial focusing based on the geographical location of the task components:

[0161]

[0162] Where, α xy p represents the spatial attention weights. loc W indicates the geographical location of the task. loc Let b represent the weight matrix. loc Indicates the bias term. Represents the weather field at time t. This represents the meteorological sequence after spatial focusing. This represents the meteorological sequence of grid point (x,y) at time t; this step S31 can focus on meteorological data of key areas based on the geographical location of the task, reducing the computational burden on irrelevant areas.

[0163] S32: Divide long task sequences into fixed-length time blocks:

[0164]

[0165] Among them, Bjk Let τ represent the k-th weather block. j λ represents the fixed time block length, λ represents the update rate, and : represents the full column slice operation; this S32 step adapts the input size of Pathformer by dividing the long time sequence into fixed-length blocks.

[0166] S33: Dynamically select the computation path based on the time block features, and calculate the feature aggregation of all paths:

[0167]

[0168] Among them, h j K represents the aggregated features of task component j. j The total number of blocks represents task component j, MeanPool represents average pooling, and h jk This represents the fusion feature of the k-th time block of task component j; step S33 dynamically selects the calculation path (global / local) based on the meteorological block features to optimize and improve computational efficiency.

[0169] S34: Output Task Component Power Prediction:

[0170]

[0171] in, This represents the power prediction sequence for component j. H represents the power component decoder. j Describing component aggregation features, d j This represents the task component description vector. Indicates the concatenation operation, T j The total time length of component j. Step S34 decodes the meteorological features into component power prediction values ​​using a power component decoder.

[0172] S4: The power prediction results for each task component are fused and corrected to eliminate jumps in the power curve between different task components, resulting in a continuous power prediction curve. There are jump points in the power predictions of different task components on the power curve. Suppose the power grid dispatch center (or a new energy power plant) initiates a prediction task at 8:00 AM, generating a power prediction curve for the next 24 hours. According to this curve, the prediction value at 9:00 AM is 100 MW. However, at 9:00 AM, the system obtains the latest meteorological data and the power plant's own operating data, automatically triggering a new rolling prediction task. Due to the updated input information, the prediction value of this newly generated prediction curve at the 9:00 AM starting point becomes 105 MW. Without seamless processing, the curve seen by the dispatcher or plant personnel will instantly "jump" from 100 MW to 105 MW at 9:00 AM, and this discontinuity will severely interfere with judgment.

[0173] Step S4 is specifically designed to address the issue of a smooth transition at this transition point. Instead of directly replacing 100 MW with 105 MW after the new prediction result is generated, a rapid fusion and correction mechanism is implemented. This module comprehensively considers the actual power values ​​of the last few points before point 9, the endpoint trend of the old prediction curve, and the starting trend of the new prediction curve. Through weighted averaging, spline interpolation, or a lightweight filtering algorithm, a smooth transition segment is generated near point 9. Thus, the curve ultimately presented to the user smoothly and naturally transitions from the original trend to the new prediction trajectory, ensuring the continuity of the prediction curve at the time-domain transition point.

[0174] Please see Figure 9 Here is a sub-flowchart for step S4 of this solution. Step S4 includes the following sub-steps:

[0175] S41: Time alignment of power predictions for different task components:

[0176]

[0177] in, This represents the aligned power sequence, where Line represents the linear interpolation function, and t... j The timestamp vector representing task component j, t u This indicates a unified timeline; the purpose of step S41 is to unify the time coordinates of different task components.

[0178] S42: Locate the time interval between different task components and mark areas of data transition risk.

[0179]

[0180] Among them, Ω j Indicates the time range of the risk zone. The end time of task component j is indicated by δt, and the transition bandwidth is indicated by δt. This step S42 is used to mark the transition area that needs special processing.

[0181] S43: Optimize the curvature of the risk area to eliminate curvature abrupt changes.

[0182]

[0183] Where i represents the number of iterations, and η represents the learning rate. Represents the loss function. This represents the power sequence of task component j after i+1 iterations. This represents the power sequence of task component j after i iterations; step S43 is used to perform conjugate iteration on the power curves in the transition region, thereby eliminating the jump points of the power curves.

[0184] S44: Weighted fusion of the power sequences of each task component yields the final power prediction curve:

[0185]

[0186] in, This represents the final power prediction curve, ω. i (t) represents the time decay weight. Step S44 generates the final smooth power prediction curve by weighted averaging.

[0187] This invention proposes a variable-scale seamless power prediction method and system based on a meteorological large-scale model. It can dynamically generate and output customized forecast results according to the specific needs of the downstream "power prediction large-scale model" using a meteorological large-scale model. These outputs not only provide key meteorological elements with high resolution and high accuracy not available in traditional NWP data sources, thus improving power prediction accuracy, but also dynamically adjust the computational path and focus of the power prediction model through scale partitioning and path selection, achieving intelligent switching of working modes and ensuring optimal prediction results in the most efficient way under any scenario. Furthermore, through fusion and correction mechanisms and cross-scale consistency design, it ensures a smooth transition of the power prediction curve at any scale transformation and time-domain transition point, making the prediction results more conducive to actual grid dispatch and market transactions.

[0188] Please see Figure 6 This application also proposes a variable-scale seamless power prediction system based on a large meteorological model. The system is implemented based on the aforementioned variable-scale seamless power prediction method based on a large meteorological model. The system includes a task decomposition module 100, a meteorological prediction module 200, a power prediction module 300, and a component fusion module 400. Specifically:

[0189] The task decomposition module 100 is used to input the prediction task into the multi-scale power prediction model, decompose the prediction task, and obtain task components of different frequencies.

[0190] The meteorological forecasting module 200 is used to acquire the initial meteorological field. The initial meteorological field and the task components are input into the large meteorological forecasting model to obtain the meteorological element forecast set corresponding to each task component.

[0191] The power prediction module 300 is used to input the meteorological element prediction set into the multi-scale power prediction model and output the power prediction results for each task component.

[0192] The component fusion module 400 is used to fuse and correct the power prediction results of each task component, eliminate the jumps in the power curves of different task components, and obtain a continuous power prediction curve.

[0193] The above description merely illustrates preferred embodiments of the present invention and is quite specific and detailed; however, it should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A variable-scale seamless power prediction method based on a large meteorological model, characterized in that, The method includes the following steps: S1: Input the prediction task into the multi-scale power prediction model, decompose the prediction task, and obtain task components of different frequencies; S2: Obtain the initial meteorological field, input the initial meteorological field and task components into the large meteorological prediction model, and obtain the meteorological element prediction set corresponding to each task component; S3: Input the meteorological element prediction set into the multi-scale power prediction model and output the power prediction results for each task component; S4: Perform fusion correction on the power prediction results of each task component to eliminate jumps in the power curve of different task components and obtain a continuous power prediction curve. The decomposition of the prediction task includes the following sub-steps: S11: Input prediction task, which includes geographic location, prediction start time, prediction end time and desired time resolution; S12: Time feature extraction, calculate the task duration based on the predicted start time and predicted end time; S13: Perform frequency domain decomposition and frequency band division on the task duration; S14: Calculate the task components based on frequency domain decomposition and frequency band division.

2. The variable-scale seamless power prediction method based on a large meteorological model as described in claim 1, characterized in that, Step S13 includes the following sub-steps: S131: Perform frequency domain decomposition of the task duration based on Fourier transform and output the spectral characteristics: ； in, Indicates spectral characteristics, Indicates Fourier transform, Represents a time impulse function; S132: Extract dominant frequencies from spectral features: ； in, Indicates the dominant frequency; S133: Frequency band division based on dominant frequency: ； in, This indicates the frequency range of frequency band j.

3. The variable-scale seamless power prediction method based on a large meteorological model as described in claim 2, characterized in that, S14 includes the following sub-steps: S141: Calculate the time resolution of task components based on frequency band division: ； ； in, Indicates component temporal resolution. This represents the upper bound frequency of frequency band j. This indicates the weight of frequency band j. Describes the band energy of band j and , Represents the total energy across the entire frequency band and ; S142: Construct task component description vectors: in, Indicates geographic location code, This indicates the time and location encoding.

4. The variable-scale seamless power prediction method based on a large meteorological model as described in claim 1, characterized in that, Step S2 includes the following sub-steps: S21: Obtain the initial meteorological field and spatially encode it. ； in, Indicates the initial meteorological field and L represents the number of vertical layers. This indicates the spatial grid size, and C represents the number of meteorological elements. This represents a 3D convolutional encoder. Indicates the embedding dimension; S22: The task component description vector is fused as a spatial coding feature of the initial conditional meteorological field to generate conditional meteorological features; S23: Based on the fused conditional meteorological characteristics, predict future meteorological elements through spatiotemporal evolution.

5. The variable-scale seamless power prediction method based on a large meteorological model as described in claim 4, characterized in that, Step S22 includes the following sub-steps: S221: Project the task component description vectors into the same embedding space as the initial meteorological field. ； in, This represents the projected feature vector. This represents a multilayer sensing projection network; Represents the task component description vector; S222: Broadcast the projected feature vector to every spatial grid point: ； Where Re represents the copy This involves a spatial copying operation to form a spatial feature map; S223: Fuse the task component features obtained after broadcasting with the spatial coding features of the initial meteorological field: ； ； in, This represents the gate value, which ranges from [0,1]. This represents the sigmoid activation function; Represents the gate weight matrix; This means concatenating the spatial coding features of the initial meteorological field with the task component features after broadcasting along the feature dimension; Indicates gating bias; This indicates element-wise multiplication.

6. The variable-scale seamless power prediction method based on a large meteorological model as described in claim 5, characterized in that, Step S23 includes the following sub-steps: S231: Predict time step based on the temporal resolution evolution of task components: ； in, Indicates the maximum step size of the physical constraint; S232: Spatiotemporal evolution using 3DTransformer: ； in, express The predicted state at any given moment; This represents the state of the meteorological field at time t; Represents the 3DTransformer evolution function; S233: Based on the prediction start time and prediction end time, iteratively execute the evolution function to generate a meteorological element prediction set aligned with the task components.

7. The variable-scale seamless power prediction method based on a large meteorological model as described in claim 1, characterized in that, Step S3 includes the following sub-steps: S31: Input the meteorological element prediction set and task component description vector into the multi-scale power prediction model, and perform spatial focusing based on the geographical location of the task components: ； ； in, Represents spatial attention weights. Indicates the geographical location of the task. Represents the weight matrix. Indicates the bias term. Represents the weather field at time t. This represents the meteorological sequence after spatial focusing. This represents the meteorological sequence of grid point (x, y) at time t; S32: Divide long task sequences into fixed-length time blocks: ； in, This represents the k-th weather block. Indicates a fixed time block length; : indicates a full column slice operation; C indicates the number of meteorological elements. S33: Dynamically select the computation path based on the time block features, and calculate the feature aggregation of all paths: ； in, This represents the aggregation feature of task component j. This represents the total number of blocks in task component j. Indicates average pooling. This represents the fusion feature of the k-th time block of task component j; S34: Output Task Component Power Prediction: ； in, This represents the power prediction sequence for component j. Indicates a power component decoder. Indicates component aggregation features, This represents the task component description vector. This indicates a splicing operation. The total time length of component j.

8. The variable-scale seamless power prediction method based on a large meteorological model as described in claim 1, characterized in that, Step S4 includes the following sub-steps: S41: Time alignment of power predictions for different task components: ; in, This represents the aligned power sequence, where Line represents the linear interpolation function. This represents the timestamp vector of task component j. Indicates a unified timeline; S42: Locate the time interval between different task components and mark areas of data transition risk. ； in, Indicates the time range of the risk zone. This indicates the end time of task component j. Indicates the transition bandwidth; S43: Optimize the curvature of the risk area to eliminate curvature abrupt changes. ； Where i represents the iteration number. Indicates the learning rate. Represents the loss function. This represents the power sequence of task component j after i+1 iterations. This represents the power sequence of task component j after i iterations; S44: Weighted fusion of the power sequences of each task component yields the final power prediction curve: ； in, This represents the final power prediction curve. This represents the time decay weight.

9. A variable-scale seamless power prediction system based on a large meteorological model, characterized in that, The system for variable-scale seamless power prediction using the method of any one of claims 1 to 8 comprises: The task decomposition module is used to input the prediction task into the multi-scale power prediction model, decompose the prediction task, and obtain task components of different frequencies. The weather forecasting module is used to acquire the initial weather field. The initial weather field and task components are input into the large weather forecasting model to obtain the weather element forecast set corresponding to each task component. The power prediction module is used to input the meteorological element prediction set into the multi-scale power prediction model and output the power prediction results for each task component. The component fusion module is used to fuse and correct the power prediction results of each task component, eliminate the jumps in the power curves of different task components, and obtain a continuous power prediction curve.

Images