Intelligent wind field forecasting method and system based on combined application of ST-LSTM and Swin-Transform

Abstract

The invention discloses an intelligent wind field forecasting method and an intelligent wind field forecasting system based on combined application of ST-LSTM (Long Short Term Memory) and Swin-Transform. The method belongs to the technical field of data processing, and comprises the following steps: constructing a time sequence prediction model based on ST-LSTM; inputting the historical data of the wind field into a time sequence prediction model based on ST-LSTM to obtain future wind field prediction data; the method comprises the following steps: constructing a wind field space downscaling model based on a Swin Transform architecture; and inputting the wind field forecast data into the wind field space downscaling model based on the Swin Transform architecture to obtain wind field refined forecast data. According to the method, the STM-LSTM is adopted as the extension of the LSTM, the long-term and short-term dependency relationship in time and space can be better captured when time sequence data is processed, and the Swindow-Transform can extract features in a global range through a self-attention mechanism of the Swindow-Transform, effectively capture the mutual influence between different positions in a wind field and further improve the accuracy of forecasting.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to an intelligent wind farm forecasting method and system based on the combined application of ST-LSTM and Swin-Transformer. Background Art

[0002] In the process of wind power generation grid connection, in order to improve stability and reduce wind abandonment, researchers have been hoping to accurately predict wind speed / power. In the past few decades, a large number of related models have been developed and applied, including statistical models, physical models and artificial intelligence models.

[0003] Statistical models are often also referred to as time series-based forecasting methods. They establish a relationship between historical wind speed data and the actual wind speed. This method is easy to model and low-cost, but it is difficult to predict the nonlinear and non-stationary characteristics of wind speed and has limited accuracy. The physical model is mainly numerical weather forecasting. It inputs an initial field containing a large number of meteorological elements such as air pressure, temperature and wind field, and predicts future weather conditions by solving the partial differential equations of atmospheric dynamics. This method is currently the most important means used in weather forecasting services. It has high accuracy and strong interpretability, but the calculation process is very complex and requires a lot of computing resources and time costs. The artificial intelligence model is actually a better alternative to the traditional statistical model. It also uses historical wind field data to predict future wind field changes. It has high accuracy and high computational efficiency, but weak interpretability, and the cumulative error of the forecast results will gradually increase during the model iteration process. The forecast results will gradually become smooth and the spatial resolution will decrease.

[0004] Based on the artificial intelligence model, in recent years, more and more institutions have launched AI meteorological big models, gradually solving the global problem that the forecast accuracy of traditional artificial intelligence models is not as good as that of numerical model forecasts. Compared with numerical model forecasts, it has a higher running speed and can complete the second-level forecast of global weather. As the world's first global mesoscale AI meteorological big model that is completely data-driven, FourCastNet can predict key variables such as ground wind speed and precipitation for up to a week with high accuracy. The "Pangu" big model developed by the Huawei team surpassed IFS in forecast accuracy for the first time by using 3D Earth-Specific Transformer and hierarchical forecasting technology, and supports a wide range of application scenarios, such as real-time forecasting of typhoon paths and ensemble forecasting. GraphCast developed by the DeepMind team uses the autoregressive model of graph neural networks, and surpassed the forecast results of PangGu-Weather with an accuracy of 99.2% in 252 forecast tasks. The "Wind and Cloud" model developed by the Shanghai Institute of Artificial Intelligence is based on multimodal and multi-task deep learning methods, and has achieved effective forecasts of key atmospheric variables for more than 10 days at high resolution, and has surpassed GraphCast on 80% of the evaluation indicators. These data-driven global weather forecast models provide medium-term forecasts of conventional global meteorological elements based on massive meteorological data, with a spatial resolution of up to 0.25°. Compared with traditional artificial intelligence models, AI meteorological large models have higher accuracy and computational efficiency, but their forecast range covers the entire world, and their spatial resolution is appropriate on a global scale, but it is obviously insufficient to meet the needs of refined forecasts for smaller local areas; although the wind field forecast results have good indicators, there is a significant smoothing phenomenon in the spatial distribution of the forecast field, making it difficult to predict strong wind events and extreme events. Summary of the invention

[0005] The purpose of the present invention is to overcome the problem of insufficient accuracy of wind farm forecast results, and proposes a wind farm intelligent forecasting method and system based on the joint application of ST-LSTM and Swin-Transformer.

[0006] In order to achieve the above object, the present invention adopts the following technical scheme:

[0007] An intelligent wind farm forecasting method based on the joint application of ST-LSTM and Swin-Transformer includes the following steps:

[0008] Build a time series prediction model based on ST-LSTM;

[0009] Input the historical data of wind farm into the time series prediction model based on ST-LSTM to obtain the future wind farm forecast data;

[0010] Construct a wind field spatial downscaling model based on the Swin Transformer architecture;

[0011] The wind field forecast data is input into the wind field spatial downscaling model based on the Swin Transformer architecture to obtain the wind field refined forecast data.

[0012] Furthermore, the method for constructing a time series prediction model based on ST-LSTM includes the following steps:

[0013] The time series prediction model is constructed by using the reanalysis wind field data sample set through ST-LSTM and GRU. The optimal model obtained by training the time series prediction model using the training architecture of the hierarchical time aggregation strategy is the time series prediction model based on ST-LSTM.

[0014] The method for constructing a wind field spatial downscaling model based on the Swin Transformer architecture includes the following steps:

[0015] The wind farm spatial downscaling model was constructed using the Swin Transformer architecture using the numerical simulation data sample set. The optimal model obtained by training the spatial downscaling model using the loss function was the wind farm spatial downscaling model based on the Swin Transformer architecture.

[0016] Furthermore, the data in the reanalysis wind farm data sample set adopts ERA5 reanalysis wind farm data, and the reanalysis wind farm data sample set is divided into a reanalysis test set, a reanalysis validation set and a reanalysis training set; the reanalysis test set is the data of two random days each month in the reanalysis wind farm data sample set, the reanalysis validation set is the data of two random days each month in the reanalysis wind farm data sample set excluding the data of the reanalysis test set, and the reanalysis training set is the data of two random days each month in the reanalysis wind farm data sample set excluding the data of the reanalysis validation set.

[0017] Furthermore, the training architecture of the hierarchical time aggregation strategy uses a greedy algorithm to limit the minimum number of iterations for the prediction target; the loss function when training the ST-LSTM-based time series prediction model uses mse, mae and perceptual loss, and the perceptual loss uses VGG16 to extract features; the normalization and LayerNorm functions are used for normalization when training the ST-LSTM-based time series prediction model; the Adam optimizer is used for optimization when training the ST-LSTM-based time series prediction model; the learning rate when training the ST-LSTM-based time series prediction model is dynamically adjusted using the ReduceLROnPlateau learning rate scheduler.

[0018] Furthermore, the numerical simulation data sample set is divided into a numerical simulation test set, a numerical simulation verification set, and a numerical simulation training set. The numerical simulation test set is the data of random days in the numerical simulation data sample set, the numerical simulation verification set is the data of random days in the numerical simulation data sample set excluding the data in the numerical simulation test set, and the numerical simulation training set is the data of random days in the numerical simulation data sample set excluding the data in the numerical simulation verification set.

[0019] Furthermore, the wind field spatial downscaling model based on the Swin Transformer architecture includes shallow feature extraction, deep feature extraction and high-resolution wind field reconstruction. The shallow feature extraction uses a convolutional layer to extract shallow features and directly transmits them to high-resolution wind field reconstruction through long jump connections.

[0020] Deep feature extraction uses RSTB blocks to extract deep features and transmit them to high-resolution wind field reconstruction. Each RSTB block uses multiple Swin Transformer layers for local attention calculation and cross-window interaction. A convolutional layer is connected at the end of each RSTB block using residual connection;

[0021] High-resolution wind field reconstruction integrates shallow and deep features.

[0022] Furthermore, the loss function when training the wind farm spatial downscaling model based on the Swin Transformer architecture adopts L1 loss; the optimization when training the wind farm spatial downscaling model based on the Swin Transformer architecture adopts Adam optimizer; the learning rate when training the wind farm spatial downscaling model based on the Swin Transformer architecture is dynamically adjusted using ReduceLROnPlateau learning rate scheduler.

[0023] An intelligent wind farm forecasting system based on the joint application of ST-LSTM and Swin-Transformer, including a time series forecasting model building module for building a time series forecasting model based on ST-LSTM;

[0024] The future wind farm forecast module is used to input the historical wind farm data into the ST-LSTM-based time series prediction model to obtain future wind farm forecast data;

[0025] Wind farm spatial downscaling model construction module, used to build a wind farm spatial downscaling model based on the Swin Transformer architecture;

[0026] The wind field refined forecast module is used to input wind field forecast data into the wind field spatial downscaling model based on the Swin Transformer architecture to obtain wind field refined forecast data.

[0027] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, an intelligent wind farm forecasting method based on the combined application of ST-LSTM and Swin-Transformer is implemented.

[0028] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for intelligent wind farm forecasting based on the combined application of ST-LSTM and Swin-Transformer is implemented.

[0029] Compared with the prior art, the present invention has the following beneficial technical effects:

[0030] The present invention proposes an intelligent wind farm forecasting method based on the joint application of ST-LSTM and Swin-Transformer. The ST-LSTM used is an extension of LSTM, which can better capture long-term and short-term dependencies in time and space when processing time series data. This is particularly important for wind farm forecasting, because changes in meteorological elements such as wind speed and wind direction are often affected by factors of multiple time scales. Through its self-attention mechanism, Swin-Transformer can extract features on a global scale, effectively capture the mutual influence between different positions in the wind farm, and further improve the accuracy of the forecast.

[0031] Furthermore, a layered time aggregation strategy is adopted instead of continuous iterative forecasting to avoid the problem of the forecast wind field becoming gradually blurred with the increase of time length, and to reduce the cumulative error. The wind field forecast results are spatially downscaled to further improve the spatial resolution of the forecast wind field, providing new products for operational wind field forecasting and meeting the refined needs of wind power generation.

[0032] Furthermore, the combined application of ST-LSTM and Swin-Transformer can fuse meteorological data from multiple sources (such as wind speed, wind direction, temperature, humidity, etc.). Through feature extraction and fusion, the model can more comprehensively understand the complex changes in the wind field, thereby improving generalization ability. The cross-domain attention mechanism of Swin-Transformer enables the model to better process data from different fields, making it possible for wind field forecasts to be applied under different geographical and climatic conditions. Swin-Transformer adopts a hierarchical structure to divide the input data into multiple small blocks for processing, reducing the amount of calculation and improving the processing speed. This is of great significance for the real-time forecasting of large-scale wind field data.

[0033] Furthermore, compared with RNN and LSTM, the Transformer model (including Swin-Transformer) has better parallel processing capabilities and can complete the training and inference process faster. The self-attention mechanism of Swin-Transformer makes the model decision process visualized, which helps to understand how the model makes predictions based on the input data and enhances the interpretability of the model. Through the visualization results, meteorological experts and decision makers can more intuitively understand the changing trends of the wind field and make more scientific and reasonable decisions. The joint model can quickly process real-time data and give immediate forecast results, meeting the real-time requirements of meteorological forecasts. By optimizing the model structure and parameters, it can ensure that the model can maintain stable forecast performance under different conditions and reduce forecast errors caused by model fluctuations. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The drawings in the specification are used to provide further understanding of the present invention and constitute a part of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0035] Figure 1 A schematic diagram of a flow chart of an intelligent wind farm forecasting method based on the combined application of ST-LSTM and Swin-Transformer according to the present invention;

[0036] Figure 2 This is a structural schematic diagram of a wind farm intelligent forecasting system based on the combined application of ST-LSTM and Swin-Transformer in the present invention;

[0037] Figure 3 This is a schematic diagram of the electronic device structure of an intelligent wind farm forecasting method based on the combined application of ST-LSTM and Swin-Transformer of the present invention;

[0038] Figure 4 A schematic diagram of a sample of a time series prediction model of a wind farm intelligent forecasting method based on the combined application of ST-LSTM and Swin-Transformer in the present invention;

[0039] Figure 5 The present invention provides a training framework for a time series prediction model of a wind farm intelligent forecasting method based on the combined application of ST-LSTM and Swin-Transformer;

[0040] Figure 6 This is a structure diagram of an intelligent spatial downscaling model of a wind farm intelligent forecasting method based on the combined application of ST-LSTM and Swin-Transformer in the present invention;

[0041] Figure 7 The present invention is a flowchart of an embodiment of a method for intelligent wind farm forecasting based on the combined application of ST-LSTM and Swin-Transformer. DETAILED DESCRIPTION

[0042] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0043] The following detailed description is an exemplary description, which is intended to provide further detailed description of the present invention. Unless otherwise specified, all technical terms used in the present invention have the same meaning as those generally understood by those skilled in the art to which the present application belongs. The terms used in the present invention are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention.

[0044] Embodiment 1

[0045] See also Figure 1 , an intelligent wind farm forecasting method based on the joint application of ST-LSTM and Swin-Transformer, comprising the following steps:

[0046] Build a time series prediction model based on ST-LSTM;

[0047] Input the historical data of wind farm into the time series prediction model based on ST-LSTM to obtain the future wind farm forecast data;

[0048] Construct a wind field spatial downscaling model based on the Swin Transformer architecture;

[0049] The wind field forecast data is input into the wind field spatial downscaling model based on the Swin Transformer architecture to obtain the wind field refined forecast data.

[0050] The intelligent wind farm forecasting method based on the joint application of ST-LSTM and Swin-Transformer can significantly improve the forecasting accuracy, enhance the model generalization ability, improve the computational efficiency, enhance the visualization and interpretability, and meet the requirements of real-time and stability. It is conducive to providing a more reliable, efficient and intelligent solution for wind farm forecasting.

[0051] The construction method of the time series prediction model based on ST-LSTM includes the following steps:

[0052] The time series prediction model is constructed by using the reanalysis wind field data sample set through ST-LSTM and GRU. The optimal model obtained by training the time series prediction model using the training architecture of the hierarchical time aggregation strategy is the time series prediction model based on ST-LSTM.

[0053] The method for constructing a wind field spatial downscaling model based on the Swin Transformer architecture includes the following steps:

[0054] The wind farm spatial downscaling model was constructed using the Swin Transformer architecture using the numerical simulation data sample set. The optimal model obtained by training the spatial downscaling model using the loss function was the wind farm spatial downscaling model based on the Swin Transformer architecture.

[0055] The data in the reanalysis wind farm data sample set adopts ERA5 reanalysis wind farm data, and the reanalysis wind farm data sample set is divided into reanalysis test set, reanalysis validation set and reanalysis training set; the reanalysis test set is the data of two random days of each month in the reanalysis wind farm data sample set, the reanalysis validation set is the data of two random days of each month in the reanalysis wind farm data sample set excluding the data of the reanalysis test set, and the reanalysis training set is the data of two random days of each month in the reanalysis wind farm data sample set excluding the data of the reanalysis validation set.

[0056] The training architecture of the hierarchical time aggregation strategy uses a greedy algorithm to limit the minimum number of iterations for the prediction target; the loss function when training the ST-LSTM-based time series prediction model uses mse, mae and perceptual loss, and the perceptual loss uses VGG16 to extract features; the normalize and LayerNorm functions are used for normalization when training the ST-LSTM-based time series prediction model; the Adam optimizer is used for optimization when training the ST-LSTM-based time series prediction model; the learning rate when training the ST-LSTM-based time series prediction model is dynamically adjusted using the ReduceLROnPlateau learning rate scheduler.

[0057] The numerical simulation data sample set is divided into a numerical simulation test set, a numerical simulation verification set, and a numerical simulation training set. The numerical simulation test set is the data of random days in the numerical simulation data sample set, the numerical simulation verification set is the data of random days in the numerical simulation data sample set excluding the data in the numerical simulation test set, and the numerical simulation training set is the data of random days in the numerical simulation data sample set excluding the data in the numerical simulation verification set.

[0058] The wind field spatial downscaling model based on the Swin Transformer architecture includes shallow feature extraction, deep feature extraction and high-resolution wind field reconstruction. Shallow feature extraction uses convolutional layers to extract shallow features and directly transmits them to high-resolution wind field reconstruction through long jump connections.

[0059] Deep feature extraction uses RSTB blocks to extract deep features and transmit them to high-resolution wind field reconstruction. Each RSTB block uses multiple Swin Transformer layers for local attention calculation and cross-window interaction. A convolutional layer is connected at the end of each RSTB block using residual connection;

[0060] High-resolution wind field reconstruction integrates shallow and deep features.

[0061] The loss function used when training the wind farm spatial downscaling model based on the Swin Transformer architecture is L1 loss; the optimization used when training the wind farm spatial downscaling model based on the Swin Transformer architecture is Adam optimizer; the learning rate used when training the wind farm spatial downscaling model based on the Swin Transformer architecture is dynamically adjusted using the ReduceLROnPlateau learning rate scheduler.

[0062] Embodiment 2

[0063] See also Figure 2 , an intelligent wind farm forecasting system based on the joint application of ST-LSTM and Swin-Transformer, including a time series forecasting model building module for building a time series forecasting model based on ST-LSTM;

[0064] The future wind farm forecast module is used to input the historical wind farm data into the ST-LSTM-based time series prediction model to obtain future wind farm forecast data;

[0065] Wind farm spatial downscaling model construction module, used to build a wind farm spatial downscaling model based on the Swin Transformer architecture;

[0066] The wind field refined forecast module is used to input wind field forecast data into the wind field spatial downscaling model based on the Swin Transformer architecture to obtain wind field refined forecast data.

[0067] Embodiment 3

[0068] See also Figure 3 , an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for intelligent wind farm forecasting based on the combined application of ST-LSTM and Swin-Transformer is implemented.

[0069] Embodiment 4

[0070] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the computer program implements the wind farm intelligent forecasting method based on the combined application of ST-LSTM and Swin-Transformer.

[0071] Embodiment 5

[0072] An intelligent wind farm forecasting method based on the joint application of ST-LSTM and Swin-Transformer, the specific steps are as follows:

[0073] (1) Step 1: Dataset creation.

[0074] ERA5 is the fifth generation of global climate reanalysis dataset developed by the European Centre for Medium-Range Weather Forecasts (ECMWF). This dataset provides high-resolution (including temporal and spatial resolution) estimates of meteorological variables in the atmosphere, ocean, land surface and subsurface from 1950 to the present. The ERA5 dataset is an integrated result based on a large amount of observational data (such as satellite observations, radiosonde, ground station observations, etc.) and advanced numerical weather prediction models (NWP). ERA5 provides data with a variety of different spatial and temporal resolutions, such as horizontal resolution of 0.25° (about 31 kilometers) or higher, and temporal resolution of every hour or less. The dataset covers a large number of meteorological and climate variables, including but not limited to temperature, humidity, wind speed, wind direction, precipitation, radiation, air pressure, sea surface temperature, etc. ERA5 covers data from 1950 (some variables may be earlier or later) to the present, providing valuable data resources for climate research, historical weather analysis, environmental science and other fields. The ERA5 dataset is a dynamically updated system. New observational data and model prediction results are continuously integrated into the dataset to provide more accurate and up-to-date climate information.

[0075] For the time series prediction model, this embodiment uses ERA5 reanalysis wind field data to create a sample set. A total of two years of data from 2017 to 2018 are used for sampling. Two days are randomly selected each month as test set samples TS1, and two days are randomly selected from the remaining samples each month as validation set samples VL1. The remaining samples at this time are training set samples TR1. In order to meet the needs of the hierarchical time aggregation strategy, the sample set of the time series prediction model is divided into three types. The input data are all wind field data in the past 6 hours. Forecast for the 1st, 3rd and 6th hours respectively like Figure 1 D represents wind field data, the subscript represents time, and the superscript represents spatial resolution.

[0076] For the intelligent spatial downscaling model, this embodiment uses numerical simulation data to make a sample set. A total of two years of data from 2017 to 2018 are used for sampling. 120 days are randomly selected as test set samples TS2, and 90 days are randomly selected from the remaining samples as validation set samples VL2. The remaining samples at this time are training set samples TR1. The input data is 0.25° numerical simulation wind field data The output data is the numerical simulation wind field data of 0.125°

[0077] (2) Step 2: Construction and training of time series prediction model.

[0078] ST-LSTM is an extension of RNN (Recurrent Neural Network). Specifically, ST-LSTM introduces the processing capability of spatial dimension on the basis of LSTM (Long Short-Term Memory), so that it can capture both temporal and spatial features in data. RNN is a type of neural network structure specially used for processing sequence data. It can capture the temporal information and contextual information in sequence data through the loop structure. However, traditional RNN is prone to the problem of gradient vanishing or gradient exploding when processing long sequences, which makes the model difficult to train. In order to overcome these problems, researchers have proposed improved models such as LSTM. LSTM controls the transmission and forgetting process of information by introducing structures such as forget gate, input gate and output gate, thereby effectively alleviating the problem of gradient vanishing and gradient exploding, and performs well in long sequence tasks. ST-LSTM is a further extension of LSTM, introducing the processing capability of spatial dimension. This enables ST-LSTM to process spatiotemporal sequence data simultaneously, that is, to consider the temporal changes and spatial distribution of data at the same time. In practical applications, ST-LSTM has been used in a variety of spatiotemporal prediction tasks, such as traffic flow prediction, weather forecast, etc., and has achieved good results.

[0079] Based on the recurrent neural network (RNN), a time series prediction model is constructed, and the gated recurrent unit (GRU) is introduced to better control the information flow, allowing the model to learn long-term dependencies, which helps to improve the prediction accuracy of the model. Compared with the commonly used model ConvLSTM for spatiotemporal sequence prediction, the spatiotemporal long short-term memory unit (ST-LSTM) used in this case consists of two memory units that can simultaneously extract and memorize spatial and temporal representations. One is the temporal memory unit included in the traditional LSTM, and the other is the spatiotemporal memory unit that captures the spatial dependencies between adjacent frames. The spatiotemporal memory unit is designed to learn complex transformation functions in a short neighborhood by increasing the depth of nonlinear neurons between temporally adjacent RNN states. Unlike simple memory splicing, the ST-LSTM unit uses a shared output gate for the two memory types, achieving seamless memory fusion, which can effectively model deformation and motion trajectories in spatiotemporal sequences.

[0080] ST-LSTM (Spatiotemporal Long Short-Term Memory Unit) is a special variant of LSTM (Long Short-Term Memory Network) that is specifically designed to process spatiotemporal sequence data such as videos, radar images, etc. ST-LSTM captures spatial and temporal dependencies in data by introducing spatiotemporal memory units and temporal memory units, as well as the interaction mechanism between them. The Spatiotemporal Memory Unit is mainly responsible for capturing the spatial dependencies between adjacent frames. In video processing, this means that it is able to understand and model the spatial relationships between pixels or features within a frame, and how these relationships change over time. The Spatiotemporal Memory Unit achieves this through vertical propagation across layers (i.e., passing information between different RNN layers). It allows information to flow in the vertical direction, thereby sharing and integrating spatial information in network structures at different levels. This mechanism enables ST-LSTM to more fully understand the complex structure of spatiotemporal sequence data and improve the model's ability to model spatial changes and temporal evolution. The Temporal Memory Unit is responsible for capturing temporal dependencies in sequence data. It is similar to the memory unit in the traditional LSTM, updating and maintaining the state through horizontal propagation (i.e., passing information between time steps within the same RNN layer). The temporal memory unit ensures that the model can handle long-term temporal dependencies and avoids the gradient vanishing or gradient exploding problems that occur when processing long sequence data. Combined with the spatiotemporal memory unit, the temporal memory unit enables ST-LSTM to capture complex dependencies in space and time at the same time, thereby performing well when processing spatiotemporal sequence data. Shared Output Gate In ST-LSTM, the spatiotemporal memory unit and the temporal memory unit share an output gate. This means that the decision of the output gate takes into account the states of both the spatiotemporal memory and the temporal memory at the same time to generate a comprehensive output representation. The design of the shared output gate simplifies the model structure, reduces the number of parameters, and helps the model to comprehensively consider spatial and temporal information when generating output. By sharing the output gate, ST-LSTM can more effectively control the quantity and quality of information output from the memory unit, thereby improving the prediction accuracy and generalization ability of the model.

[0081] Hierarchical time aggregation strategy is a data management technology, which is particularly effective when processing time series data. The core idea of ​​this strategy is to aggregate time series data according to different time intervals to facilitate data analysis and storage at different levels. Hierarchical time aggregation strategy divides time into levels, such as year, quarter, month, week, day, etc. Each level contains the time of the next level, and the time range of each level is fixed. Through this hierarchical division, data analysis and query at multiple time granularities can be achieved. Implementing hierarchical time aggregation strategy usually includes the following steps: First, collect the time series data to be processed. Divide the time into different levels according to actual needs, such as year, quarter, month, week, day, etc. Aggregate the data according to the divided time levels. Analyze and query the aggregated data as needed.

[0082] The training architecture using the layered temporal aggregation strategy is as follows Figure 2 As shown, this embodiment trains three independent models to make predictions for the first hour, third hour, and sixth hour in the future respectively; in the inference stage, given a prediction target, a greedy algorithm can be used to ensure the minimum number of iterations. For example, for the forecast for the 12th hour in the future, two 6-hour forecasts are performed; and for the forecast for the 10th hour in the future, one 6-hour forecast, one 3-hour forecast, and one 1-hour forecast are performed. The loss function combines mse, mae, and perceptual loss, where the perceptual loss uses VGG16 to extract features. The model is normalized using normalize and LayerNorm functions, and the Adam optimizer is used. The initial learning rate is 0.001, and the ReduceLROnPlateau learning rate scheduler is used to dynamically adjust the learning rate. The patience value of the scheduler is 10, the attenuation factor is 0.5, and the number of training cycles is 100. The optimal model after convergence is called M1.

[0083] (3) Step 3: Construction and training of intelligent spatial downscaling model.

[0084] The structure of the wind field spatial downscaling model is as follows: Figure 3As shown in the figure, it includes three parts: shallow feature extraction, deep feature extraction and high-resolution wind field reconstruction. Shallow feature extraction uses convolutional layers to extract shallow features and transfers them directly to the high-resolution wind field reconstruction part through long jump connections to retain low-frequency information. Deep feature extraction is mainly composed of Swin Transformer blocks (RSTB) with residual connections. Each block uses multiple Swin Transformer layers for local attention calculations and cross-window interactions. A convolutional layer is added to the end of each RSTB block to enhance the features, and residual connections are used to provide shortcuts for feature aggregation. Finally, shallow and deep features are fused in the reconstruction part to achieve high-resolution wind field reconstruction. The advantage of this model lies in the Swin Transformer architecture it uses, which can effectively capture long-term dependencies and global contextual information, which is crucial for spatial downscaling tasks.

[0085] When training the spatial downscaling model, the loss function is L1 loss, the Adam optimizer is used, the initial learning rate is 0.001, and the ReduceLROnPlateau learning rate scheduler is used to dynamically adjust the learning rate. The patience value of the scheduler is 5, the decay factor is 0.5, the number of training cycles is 200, and the optimal model after convergence is called M2.

[0086] (4) Step 4: Model verification and application.

[0087] After obtaining M1 and M2 through steps 2 and 3 respectively, the reserved test set can be used to test the accuracy of the two models respectively. M1 is tested and evaluated using TS1 in step 1. The average absolute error of the U component is 0.828m / s, and the average root mean square error is 1.140m / s; the average absolute error of the V component is 0.903m / s, and the average root mean square error is 1.266m / s. M2 is tested and evaluated using TS2 in step 1. The average absolute error of the U component is 0.067m / s, and the average root mean square error is 0.123m / s; the average absolute error of the V component is 0.071m / s, and the average root mean square error is 0.129m / s.

[0088] In the application of actual business scenarios, the method flow proposed in this embodiment is as follows Figure 4 As shown, firstly, the historical wind field data with a spatial resolution of 0.25° is input into M1 to obtain the wind field forecast data for the next 12 hours; then the wind field forecast data with a spatial resolution of 0.25° is input into M2 to obtain the refined wind field forecast data with a spatial resolution of 0.125°.

[0089] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented 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.

[0090] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0091] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0092] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A wind farm intelligent forecasting method based on the joint application of ST-LSTM and Swin-Transformer, characterized in that: The following steps are involved: Build a time series prediction model based on ST-LSTM; Input the historical data of wind farm into the time series prediction model based on ST-LSTM to obtain the future wind farm forecast data; Construct a wind field spatial downscaling model based on the Swin Transformer architecture; The wind field forecast data is input into the wind field spatial downscaling model based on the Swin Transformer architecture to obtain the wind field refined forecast data.

2. According to claim 1, a wind farm intelligent forecasting method based on the combined application of ST-LSTM and Swin-Transformer is characterized in that: The construction method of the time series prediction model based on ST-LSTM includes the following steps: The time series prediction model is constructed by using the reanalysis wind field data sample set through ST-LSTM and GRU. The optimal model obtained by training the time series prediction model using the training architecture of the hierarchical time aggregation strategy is the time series prediction model based on ST-LSTM. The method for constructing a wind field spatial downscaling model based on the Swin Transformer architecture includes the following steps: The wind farm spatial downscaling model was constructed using the Swin Transformer architecture using the numerical simulation data sample set. The optimal model obtained by training the spatial downscaling model using the loss function was the wind farm spatial downscaling model based on the Swin Transformer architecture.

3. According to claim 2, a wind farm intelligent forecasting method based on the combined application of ST-LSTM and Swin-Transformer is characterized in that: The data in the reanalysis wind farm data sample set adopts ERA5 reanalysis wind farm data, and the reanalysis wind farm data sample set is divided into reanalysis test set, reanalysis validation set and reanalysis training set; the reanalysis test set is the data of two random days of each month in the reanalysis wind farm data sample set, the reanalysis validation set is the data of two random days of each month in the reanalysis wind farm data sample set excluding the data of the reanalysis test set, and the reanalysis training set is the data of two random days of each month in the reanalysis wind farm data sample set excluding the data of the reanalysis validation set.

4. According to claim 2, a wind farm intelligent forecasting method based on the combined application of ST-LSTM and Swin-Transformer is characterized in that: The training architecture of the hierarchical temporal aggregation strategy uses a greedy algorithm to limit the minimum number of iterations for the prediction target; The loss function for training the ST-LSTM-based time series prediction model uses mse, mae and perceptual loss. The perceptual loss uses VGG16 to extract features. The normalize and LayerNorm functions are used for normalization when training the ST-LSTM-based time series prediction model; the Adam optimizer is used for optimization when training the ST-LSTM-based time series prediction model; the learning rate when training the ST-LSTM-based time series prediction model is dynamically adjusted using the ReduceLROnPlateau learning rate scheduler.

5. According to claim 2, a wind farm intelligent forecasting method based on the combined application of ST-LSTM and Swin-Transformer is characterized in that: The numerical simulation data sample set is divided into a numerical simulation test set, a numerical simulation verification set, and a numerical simulation training set. The numerical simulation test set is the data of random days in the numerical simulation data sample set, the numerical simulation verification set is the data of random days in the numerical simulation data sample set excluding the data in the numerical simulation test set, and the numerical simulation training set is the data of random days in the numerical simulation data sample set excluding the data in the numerical simulation verification set.

6. The method for intelligent wind farm forecasting based on the combined application of ST-LSTM and Swin-Transformer according to claim 2 is characterized in that: The wind field spatial downscaling model based on the Swin Transformer architecture includes shallow feature extraction, deep feature extraction and high-resolution wind field reconstruction. Shallow feature extraction uses convolutional layers to extract shallow features and directly transmits them to high-resolution wind field reconstruction through long jump connections. Deep feature extraction uses RSTB blocks to extract deep features and transmit them to high-resolution wind field reconstruction. Each RSTB block uses multiple Swin Transformer layers for local attention calculation and cross-window interaction. A convolutional layer is connected at the end of each RSTB block using residual connection; High-resolution wind field reconstruction integrates shallow and deep features.

7. The method for intelligent wind farm forecasting based on the combined application of ST-LSTM and Swin-Transformer according to claim 2 is characterized in that: The loss function used when training the wind farm spatial downscaling model based on the Swin Transformer architecture is L1 loss; the optimization used when training the wind farm spatial downscaling model based on the Swin Transformer architecture is Adam optimizer; the learning rate used when training the wind farm spatial downscaling model based on the Swin Transformer architecture is dynamically adjusted using the ReduceLROnPlateau learning rate scheduler.

8. An intelligent wind farm forecasting system based on the joint application of ST-LSTM and Swin-Transformer, characterized in that: Includes a time series prediction model building module, which is used to build a time series prediction model based on ST-LSTM; The future wind farm forecast module is used to input the historical wind farm data into the ST-LSTM-based time series prediction model to obtain future wind farm forecast data; Wind farm spatial downscaling model construction module, used to build a wind farm spatial downscaling model based on the Swin Transformer architecture; The wind field refined forecast module is used to input wind field forecast data into the wind field spatial downscaling model based on the Swin Transformer architecture to obtain wind field refined forecast data.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for intelligent wind farm forecasting based on the combined application of ST-LSTM and Swin-Transformer as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements a wind farm intelligent forecasting method based on the combined application of ST-LSTM and Swin-Transformer as described in any one of claims 1 to 7.

Images