Multi-station wind speed long-term prediction method and system based on cross-station spatio-temporal statistical correlation self-supervised learning

Abstract

This application provides a method and system for long-term multi-station wind speed prediction based on cross-station spatiotemporal statistical correlation self-supervised learning, relating to the interdisciplinary fields of atmospheric science and artificial intelligence. The method includes: acquiring and preprocessing historical meteorological data from multiple stations; constructing a cross-station spatiotemporal heterogeneous graph containing temporal, spatial, and spatiotemporal edges; extracting deep spatiotemporal representations using a spatiotemporal graph neural network encoder; performing temporal decomposition to obtain trend, periodic, and residual components; constructing a self-supervised contrastive learning mechanism based on the trend component and the cross-station spatiotemporal autocorrelation function; optimizing the encoder using cross-station spatiotemporal contrastive loss; constructing a decomposed long-term prediction architecture corresponding to the three components; optimizing the model using a two-stage strategy of self-supervised pre-training and supervised fine-tuning; and outputting multi-station wind speed prediction results for the next 30 to 90 days. This application effectively improves the accuracy and physical consistency of long-term multi-station wind speed prediction by integrating cross-station spatiotemporal self-supervised learning with a decomposed prediction architecture.

Description

Technical Field

[0001] This application relates to the interdisciplinary fields of atmospheric science and artificial intelligence, and in particular to a method and system for long-term prediction of multi-station wind speed based on cross-station spatiotemporal statistical correlation self-supervised learning. Background Technology

[0002] Current long-term wind speed forecasting technologies mainly revolve around three major directions: physical model-driven, data-driven, and fusion learning. However, all of these have significant limitations, as follows: Traditional wind speed forecasting methods mainly fall into two categories: numerical weather prediction models and statistical machine learning methods. Numerical weather prediction models solve for wind field evolution based on atmospheric dynamic equations, but these methods have significant limitations: firstly, numerical simulations are extremely sensitive to initial conditions, and errors accumulate exponentially with increasing forecast duration, leading to a sharp decline in medium- and long-term forecast accuracy; secondly, numerical models are limited by computational resources and grid resolution, making it difficult to accurately characterize local wind field features under complex terrain conditions and lacking the ability to simulate the spatial distribution of wind speeds at multiple stations. Statistical machine learning methods build forecasting models by mining statistical patterns from historical wind speed data, but these methods mainly rely on linear assumptions or shallow feature extraction, limiting their ability to capture the strong nonlinearity and non-stationarity of wind speed time series. When faced with seasonal variations, extreme weather events, and long-term evolution patterns, the model's generalization performance significantly decreases, making it difficult to meet the accuracy requirements for medium- and long-term forecasts of 30 to 90 days.

[0003] In recent years, deep learning technology has made significant progress in the field of time series forecasting, with neural network models such as LSTM, GRU, and Transformer being widely used in wind speed forecasting research. However, in the scenario of long-term wind speed forecasting at multiple stations, existing deep learning methods still face three core challenges: First, the limited sliding window length. Most time series models use a fixed-length historical observation window as input. Due to limitations in computational complexity and gradient propagation stability, the window length is difficult to significantly expand, resulting in the model's inability to effectively capture long-term dependencies such as seasonal cycles that extend beyond the window range. Second, insufficient modeling of cross-station spatiotemporal dependencies. Traditional methods often treat each wind station as an independent entity for single-station forecasting, ignoring the spatial correlations and time-delay propagation patterns formed between different stations due to factors such as weather system movement and terrain coupling effects, failing to fully utilize the collaborative information of multi-station data. Third, strong dependence on labeled data. Existing supervised learning frameworks heavily rely on a large amount of high-quality labeled data, but the acquisition of long-term forecast labels is costly and time-consuming, and the sample sparsity problem is prominent, resulting in insufficient generalization ability of the model in data-constrained scenarios.

[0004] To address the problem of cross-site spatiotemporal dependency modeling, some studies have attempted to introduce spatiotemporal graph neural network technology. However, existing spatiotemporal graph modeling methods still have significant shortcomings in wind speed prediction applications: the graph structure construction lacks physical constraints, and most methods set static connection weights based on geographical distance or data similarity, failing to effectively reflect the dynamic time-lag characteristics of wind field propagation and the evolutionary laws of weather systems. Furthermore, self-supervised learning, as a representation learning paradigm driven by unlabeled data, still lacks domain specificity in the field of meteorological time-series forecasting: the division of positive and negative samples is mostly based on similarity calculations after data augmentation, without incorporating meteorological physical laws, making it difficult for the model to learn physically meaningful long-term spatiotemporal correlation features.

[0005] Existing technologies for long-term multi-station wind speed forecasting generally suffer from core problems such as insufficient capture of long-term temporal dependencies, lack of physical constraints in cross-station spatiotemporal correlation modeling, and limited utilization of unlabeled data. Therefore, there is an urgent need to develop a long-term multi-station wind speed forecasting method that integrates cross-station spatiotemporal statistical correlation constraints with self-supervised representation learning to address the aforementioned technical bottlenecks. Summary of the Invention

[0006] To address the challenges of strong time series non-stationarity in long-term multi-station wind speed forecasting, the limitation of forecast models by the sliding time window length, and the difficulty in effectively characterizing cross-site spatiotemporal dependencies, this application provides a multi-station wind speed long-term forecasting method and system based on cross-site spatiotemporal statistical correlation self-supervised learning. This method solves the technical difficulties of traditional forecasting methods, such as insufficient long-term structure modeling capability outside the window, significant accumulation of forecast errors over time, and poor model generalization stability. As a result, it improves the forecasting accuracy and robustness of multi-station wind speed on medium- to long-term forecasting scales such as 30 days to 90 days.

[0007] Firstly, this application provides a long-term multi-station wind speed prediction method based on cross-station spatiotemporal statistical correlation self-supervised learning, the method comprising: Get the target area M The historical meteorological monitoring datasets of each wind measurement station were preprocessed to obtain multi-station time-series meteorological data with a unified time reference and continuous and complete data. Based on the multi-station time-series meteorological data, a cross-station spatiotemporal heterogeneous map is constructed. Each node in the node set corresponds to the observation status of the wind measurement station at a certain moment, and the edge set includes the time edge set. Spatial edge set and spacetime edge set ; A spatiotemporal graph neural network encoder is constructed. Through a layer-by-layer message passing mechanism, the three types of information from each node—temporal neighbors, spatial neighbors, and spatiotemporal neighbors—are aggregated to obtain a deep spatiotemporal representation. The deep spatiotemporal representation is decomposed into trend components, periodic components, and residual components using a time-series decomposition module. Based on the trend components and the pre-calculated cross-site spatiotemporal autocorrelation function, positive and negative sample pairs for self-supervised contrastive learning are constructed. The parameters of the spatiotemporal graph neural network encoder and the time-series decomposition module are optimized using a cross-site spatiotemporal contrastive loss function. A decompositional long-term forecasting architecture is constructed, which takes trend component, periodic component and residual component as input and outputs the future wind speed forecast results of each wind measurement station in the target area. The spatiotemporal graph neural network encoder, the time series decomposition module and the decompositional long-term forecasting architecture are combined to form a forecasting model. The forecasting model is trained using a two-stage training strategy of self-supervised pre-training and supervised fine-tuning. Based on the trained forecasting model, long-term wind speed forecasting for multiple stations is realized.

[0008] As a preferred technical solution, within the target area M Historical meteorological monitoring dataset from [number] wind measurement stations, [number] Each station at time The observation vector is: in, u m,t , v m,t They are time points t Wind measurement station m The east-west component and the north-south component of the wind speed. T m,t For a moment t Wind measurement station m temperature P m,t For a moment t Wind measurement station m air pressure, φ m , λ m , h m These are wind measurement stations m Latitude, longitude, and altitude; The preprocessing of historical meteorological monitoring datasets includes sequential steps such as station screening and time alignment, outlier repair and missing value completion, wind direction coding and feature standardization, and data format regularization. Among these steps, station screening ensures that the maximum distance between wind measurement stations does not exceed a preset threshold. Time alignment ensures that all data timestamps are in the same time zone and have the same time resolution.

[0009] As a preferred technical solution, the time edge set Used to model the dynamic evolution relationship between different times within the same wind measurement station, for wind measurement stations m Connection time and The time-domain weights are based on the time-domain autocorrelation coefficient of the wind speed sequence at that wind measurement station. Sure, The statistical correlation strength of wind speed sequences at the same wind measurement station at a lag time τ is quantified, with the time edge weights being the autocorrelation exponential decay function. The set of spatial edges Used to model the spatial correlation structure between different wind measurement stations at the same time, for any two wind measurement stations m and n ,time t Spatial boundary weighting of the great circle distance between two wind measurement stations The Pearson correlation coefficient with historical wind speed series was determined, where The results were obtained from the latitude and longitude coordinates of the two wind measurement stations using the spherical trigonometry formula. The set of spatiotemporal edges To capture the wind field propagation effect between different wind measurement stations at different times, the weights of the spatiotemporal edges are calculated using a pre-computed cross-station spatiotemporal autocorrelation function. The absolute value, Historical data statistics reflect the wind measurement stations m With wind measurement stations n Time lag Spatial distance Wind speed correlation strength under certain conditions.

[0010] As a preferred technical solution, in each layer of the spatiotemporal graph neural network encoder, each node simultaneously receives messages from three types of neighbors: temporal neighbors, spatial neighbors, and spatiotemporal neighbors. For each type of neighbor, an attention mechanism is used to calculate aggregation weights. Through a learnable linear transformation, the representations of the query node and neighboring nodes are mapped to the same feature space. A similarity score is calculated and normalized to obtain the attention coefficient. The three types of messages are fused after independent linear transformations, and then nonlinear transformations are performed through a multilayer perceptron. The updated representation is added to the original representation using residual connections. After stacking L layers of the network, the deep spatiotemporal representation of the node is output.

[0011] As a preferred technical solution, the trend component is extracted by multi-scale moving average operation, and the original depth spatiotemporal representation sequence is smoothed by multiple sliding windows with different time scales, and the average multi-scale smoothing result is obtained; the periodic component is obtained by fitting the Fourier basis function with a preset typical period; the residual component is the remaining part after subtracting the trend component and the periodic component from the original depth spatiotemporal representation.

[0012] As a preferred technical solution, based on the trend component and the pre-calculated cross-site spatiotemporal autocorrelation function, positive and negative sample pairs for self-supervised contrastive learning are constructed, specifically including: Random sampling from multi-station time-series meteorological data N Each spatiotemporal window is used as a training batch; Calculate any two spatiotemporal windows and Time distance Spatial distance between stations The correlation strength is obtained by querying the pre-calculated cross-site spatiotemporal autocorrelation function. : in, This is the cross-site spatiotemporal autocorrelation function obtained from global statistics of the training data; For anchoring windows Its positive sample set Includes batches with a correlation higher than the threshold Other windows; For each anchor window, all relevance levels not exceeding [a certain threshold] are [not specified]. All other windows are considered negative samples.

[0013] As a preferred technical solution, the cross-site spatiotemporal comparison loss function is: in, This represents the total number of windows in the batch. For window The number of positive samples; For window and The cross-site spatiotemporal autocorrelation intensity, as a weighting factor, makes the highly correlated positive samples contribute more; To characterize the cosine similarity between vectors; z i and z j For window and The fixed-dimensional representation vector is obtained by performing time-averaged pooling on the trend representation sequence within the window. The temperature coefficient hyperparameter controls the sharpness of the similarity distribution; For window The corresponding set of negative samples; exp is the natural exponential function; This represents the loss in cross-site spatiotemporal comparison.

[0014] As a preferred technical solution, the decomposed long-term forecasting architecture includes a trend forecasting branch, a cycle forecasting branch, and a residual forecasting branch, wherein: The trend prediction branch takes the trend component as input, aggregates spatiotemporal information through global pooling operation, maps it to the prediction time through a multilayer perceptron, and outputs the trend prediction value of each wind measurement station. The periodic prediction branch performs analytical extrapolation based on the Fourier basis function coefficients obtained from the periodic components, extending the sine and cosine basis functions to future times to generate seasonal prediction values. The residual prediction branch takes the recent sequence of the residual component as input, performs short-time extrapolation using a lightweight linear model, and outputs the residual correction value. The final wind speed forecast is obtained by adding and fusing the trend forecast, seasonal forecast, and residual correction values.

[0015] As a preferred technical solution, the two-stage training strategy includes a first stage and a second stage, wherein: The first stage is a self-supervised pre-training stage, which only optimizes the parameters of the spatiotemporal graph neural network encoder and the temporal decomposition module, with the cross-site spatiotemporal contrast loss function as the optimization target; The second stage is the supervised fine-tuning stage, in which the weights of the spatiotemporal graph neural network encoder and the temporal decomposition module, which have been completed by self-supervised pre-training, are loaded as initialization. The branch parameters of the decompositional long-term prediction architecture are jointly optimized, and the optimization objective is a weighted combination of the prediction mean square error and the cross-station spatiotemporal comparison loss. In the fine-tuning stage, the backbone parameters of the spatiotemporal graph neural network encoder are frozen, or the parameters of the spatiotemporal graph neural network encoder are fine-tuned using a learning rate that is smaller than that in the self-supervised pre-training stage.

[0016] Secondly, this application provides a multi-station long-term wind speed forecasting system based on cross-station spatiotemporal statistical correlation self-supervised learning, used to implement the method described above, the system comprising: The data preprocessing module is configured to acquire data within the target area. M The historical meteorological monitoring datasets of each wind measurement station were preprocessed to obtain multi-station time-series meteorological data with a unified time reference and continuous and complete data. The heterogeneous graph construction module is configured to construct a cross-station spatiotemporal heterogeneous graph based on the multi-station time-series meteorological data. Each node in the node set corresponds to the observation status of the wind measurement station at a certain moment, and the edge set includes the time edge set. Spatial edge set and spacetime edge set ; The encoder module is configured to build a spatiotemporal graph neural network encoder. Through a layer-by-layer message passing mechanism, it aggregates three types of information from each node's temporal neighbors, spatial neighbors, and spatiotemporal neighbors to obtain a deep spatiotemporal representation. The self-supervised learning module is configured to perform temporal decomposition on the deep spatiotemporal representation using the temporal decomposition module to obtain trend components, periodic components, and residual components; based on the trend components and the pre-calculated cross-site spatiotemporal autocorrelation function, positive and negative sample pairs for self-supervised contrastive learning are constructed, and the parameters of the spatiotemporal graph neural network encoder and the temporal decomposition module are optimized through the cross-site spatiotemporal contrastive loss function; The long-term forecasting module is configured to construct a decompositional long-term forecasting architecture. The decompositional long-term forecasting architecture takes trend components, periodic components, and residual components as inputs and outputs the future wind speed forecast results for each wind measurement station in the target area. The spatiotemporal graph neural network encoder, the time series decomposition module, and the decompositional long-term forecasting architecture are combined to form a forecasting model. The forecasting model is trained using a two-stage training strategy of self-supervised pre-training and supervised fine-tuning. Based on the trained forecasting model, long-term wind speed forecasting for multiple stations is realized.

[0017] The multi-station long-term wind speed prediction method and system based on cross-station spatiotemporal statistical correlation self-supervised learning provided in this application have at least the following beneficial effects: 1) To address the issues of strong time series non-stationarity, limited window length, and difficulty in characterizing cross-station dependencies in long-term wind speed forecasting at multiple stations, this application designs a cross-station spatiotemporal self-supervised learning framework. Based on historical wind speed and meteorological data from multiple stations, it significantly improves the accuracy and physical consistency of long-term forecasts by constructing a cross-station spatiotemporal heterogeneous graph and a self-supervised comparative learning mechanism.

[0018] 2) In the data preprocessing stage, standardization, spatiotemporal interpolation, and quality control are used to fully ensure the integrity, consistency, and reliability of the input data, laying a high-quality data foundation for subsequent model training. In the cross-station spatiotemporal graph construction stage, three types of edges (temporal edges, spatial edges, and spatiotemporal edges) are designed to capture the dynamic evolution within stations, the synchronous correlation between stations, and the cross-station and cross-period wind field propagation effects, respectively, embedding physical constraints at the graph structure level. In the feature learning stage, a spatiotemporal graph neural network encoder is used to extract deep spatiotemporal features of the long-term evolution of wind speed at multiple stations. Combined with the time-series decomposition module, trend, period, and residual components are separated. The feature space is optimized through cross-station spatiotemporal comparison loss, which clusters spatiotemporally similar wind speed patterns and separates spatiotemporally dissimilar patterns, thereby strengthening the feature discrimination power. In the long-term prediction stage, the pre-trained spatiotemporal encoder is frozen to preserve the spatiotemporal consistency of wind speed at multiple stations, and the decomposition prediction branch is fine-tuned to achieve high-precision prediction of wind speed at multiple stations for the next 30 to 90 days.

[0019] 3) In the long-term wind speed prediction task at multiple stations, the method of this application significantly reduces the RMSE and MAE indices, thus significantly improving the accuracy of long-term prediction. At the same time, the cross-station spatiotemporal consistency constraint ensures the physical rationality of the prediction results, and the spatial visualization results verify the good generalization ability of the model in different terrain regions. It can provide highly reliable technical support for wind power prediction, grid dispatch and energy management. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0021] Figure 1 A flowchart of a long-term multi-station wind speed prediction method based on cross-station spatiotemporal statistical correlation self-supervised learning provided for embodiments of this application; Figure 2 A network structure diagram of a multi-station long-term wind speed prediction framework that integrates a cross-station spatiotemporal self-supervised learning mechanism is provided in the embodiments of this application; Figure 3 This is a structural diagram of a multi-station long-term wind speed prediction system based on cross-station spatiotemporal statistical correlation self-supervised learning, provided as an embodiment of this application.

[0022] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0023] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0024] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0025] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0026] In the field of atmospheric science and meteorological forecasting, long-term multi-station wind speed forecasting has received widespread attention as a key technology with significant economic and social impact. However, due to the diversity and complexity of multi-station wind speed spatiotemporal evolution models, existing technologies struggle to perform in-depth spatiotemporal feature mining and effective correlation modeling of multi-station meteorological data, exhibiting significant limitations.

[0027] Current research on long-term prediction of multi-station wind speed often suffers from limitations in representing complex spatiotemporal patterns or failing to adequately capture implicit cross-station relationships among data. There is a lack of efficient methods for modeling the complex nonlinear spatiotemporal relationships between multidimensional meteorological elements and multi-station wind speed. Existing numerical weather prediction models and traditional machine learning methods, when capturing key spatiotemporal features influencing multi-station wind speed and their interactions, often fail to effectively mine deep spatiotemporal similarities and differences in the data, thus limiting a comprehensive understanding and prediction of the long-term evolution of multi-station wind speed. This limitation makes it difficult to achieve accurate long-term predictions of multi-station wind speed under complex and variable meteorological conditions in different regions and terrains. Furthermore, existing technologies lack the ability to effectively utilize unlabeled cross-station data and discriminate spatiotemporal features when dealing with the diversity of spatiotemporal variation patterns of multi-station wind speed, failing to fully utilize the inherent spatiotemporal structural information of large amounts of historical multi-station meteorological data, thereby affecting the accuracy and reliability of long-term prediction results. Therefore, developing a new method capable of deeply mining the spatiotemporal characteristics of multi-station meteorological data and accurately predicting long-term wind speed has become an urgent task in current research.

[0028] In view of the above problems, this application provides a method for long-term multi-station wind speed prediction based on cross-station spatiotemporal statistical correlation self-supervised learning. First, a cross-station spatiotemporal heterogeneous graph is constructed based on historical wind speed and meteorological data of multiple stations, and a cross-station spatiotemporal autocorrelation calculation mechanism is designed to learn and characterize the spatiotemporal dependencies of the long-term evolution of wind speed at multiple stations, providing a physically consistent spatiotemporal structural foundation for model training. Second, a cross-station self-supervised pre-training-long-term prediction fine-tuning architecture is established. During the cross-station comparative pre-training process, the model is driven to learn discriminative spatiotemporal feature representations related to wind speed at multiple stations, and fine-tuning is performed in conjunction with the long-term prediction task to achieve high-precision prediction of wind speed at multiple stations for the next 30 to 90 days, improving the accuracy, generalization ability and physical rationality of the model, and providing reliable support for applications such as wind power prediction, grid dispatching and energy management.

[0029] Specifically, this method first analyzes the periodicity, non-stationarity, and spatiotemporal dependence characteristics of multi-station wind speed data, and constructs a cross-station spatiotemporal heterogeneous map by combining meteorological and physical laws. Then, a spatiotemporal graph neural network encoder is used to extract the spatiotemporal pattern features of the long-term evolution of multi-station wind speed. The feature space is optimized through a cross-station spatiotemporal contrast loss function, causing spatiotemporally similar wind speed patterns to cluster and spatiotemporally dissimilar patterns to separate. Finally, the parameters of the pre-trained spatiotemporal encoder are frozen to preserve the learned spatiotemporal consistency of multi-station wind speed, and the long-term prediction branch parameters are fine-tuned, ultimately achieving long-term accurate prediction of multi-station wind speed. Figure 1 As shown, the multi-station long-term wind speed prediction method based on cross-station spatiotemporal statistical correlation self-supervised learning specifically includes the following steps S10-S50.

[0030] S10: Obtain the target area M Historical meteorological monitoring datasets from several wind measurement stations were collected and preprocessed to obtain multi-station time-series meteorological data with a unified time reference and continuous and complete data.

[0031] In some embodiments, data collection and preprocessing can be performed in the following manner: A multi-station historical meteorological monitoring dataset is taken as input, the dataset containing... The multidimensional meteorological variable series of several wind measurement stations covers observational data of a certain period in the target area. Let the first... Each station at time The observation vector is: in, u m,t , v m,t They are time points t Wind measurement station m The east-west component and the north-south component of the wind speed. T m,t For a moment t Wind measurement station m temperature P m,t For a moment t Wind measurement station m air pressure, φ m , λ m , h m These are wind measurement stations m Latitude, longitude, and altitude.

[0032] Preprocessing is carried out through a four-stage core process: First, sites are selected based on the geographical range of the target area, ensuring that the maximum distance between sites does not exceed a preset threshold. (100km) Simultaneously, all data timestamps are unified to UTC+8 time zone, with a time resolution of [missing information]. (1 hour) Ensure time reference consistency; then, repair outliers according to meteorological and physical laws, and use linear interpolation to complete missing data to ensure data continuity; next, convert wind direction into a direction cosine vector to avoid angular periodicity issues, and standardize all core features to eliminate the influence of dimensional differences; finally, organize the processed data into a unified CSV format of "site ID-timestamp-feature vector". This preprocessing effectively removes data noise while preserving the long-term temporal details of wind speeds from multiple stations, providing a high-quality data foundation for subsequent cross-station spatiotemporal map construction.

[0033] S20: Based on the aforementioned multi-station time-series meteorological data, construct a cross-station spatiotemporal heterogeneous map. Each node in the node set corresponds to the observation status of the wind measurement station at a certain moment, and the edge set includes the time edge set. Spatial edge set and spacetime edge set .

[0034] In some embodiments, cross-site spatiotemporal heterogeneous graph construction can be achieved in the following manner: based on preprocessed multi-site time series data, a cross-site spatiotemporal heterogeneous graph is constructed. Each node in the diagram The size of the node set corresponds to the observation status of a specific station at a specific time. The initial features of each node are the low-dimensional representation graph obtained by linear projection of the preprocessed observation vector. The edges are divided into three types to capture different spatiotemporal dependencies.

[0035] The first category is time edge sets. Used to model the dynamic evolution relationship between adjacent time points within the same site. For a site Connection time and The edge weights are determined based on the time-domain autocorrelation coefficient of the wind speed sequence at that site. (Time-domain autocorrelation coefficient) Quantify the wind speed sequence at the same site in lag The statistical correlation strength at time t, its value range is: A larger absolute value indicates a stronger correlation. The edge weights are designed as an autocorrelation-based exponential decay function to ensure that closely spaced connections with high autocorrelation have stronger weights. Specifically, this includes a temperature coefficient hyperparameter. Used to adjust the decay rate.

[0036] The second category is spatial edge sets. This is used to model the spatial correlation structure between different stations at the same time. For any two stations... and At any moment Establishing spatial connections involves a comprehensive consideration of two aspects of geographic and meteorological information: firstly, the great circle distance between the two stations. The first factor is the shortest path length between two points on the Earth's surface, calculated from latitude and longitude coordinates using spherical trigonometry. The second factor is the Pearson correlation coefficient of the historical wind speed sequences of the two stations, which quantifies the synchronicity of wind speed change trends between the two stations. The edge weights have a negative exponential relationship with distance and a positive power relationship with correlation, and include a distance attenuation scale parameter. and correlation index parameters Used to adjust the relative contributions of the two. The third type is the set of spatiotemporal edges. This is used to capture the wind field propagation effect between different stations at different times. Considering the moving characteristics of weather systems, the stations... At any moment The wind field conditions may be related to the site Lag The states at any given time exhibit statistical correlations, reflecting the propagation of the wind field from upstream to downstream stations. The weights of the spatiotemporal edges are directly derived from the pre-calculated cross-site spatiotemporal autocorrelation function. The absolute value of the function, obtained from historical data statistics, reflects the site's... With the site Time lag Spatial distance The correlation strength of wind speed under certain conditions. By querying this function and selecting the time delay corresponding to the correlation peak, cross-site connections with optimal time alignment can be established, thereby embedding the physical laws of wind field propagation into the graph structure.

[0037] S30: Construct a spatiotemporal graph neural network encoder, and through a layer-by-layer message passing mechanism, aggregate three types of information from each node's temporal neighbors, spatial neighbors, and spatiotemporal neighbors to obtain a deep spatiotemporal representation.

[0038] This spatiotemporal graph neural network encoder employs a multi-layer spatiotemporal graph neural network to perform deep encoding of node features. Through a layer-by-layer message passing mechanism, the encoder aggregates neighbor information for each node to update its representation, thereby embedding the local spatiotemporal structure into the deep feature space.

[0039] In some embodiments, in each layer of the spatiotemporal graph neural network encoder, each node simultaneously receives messages from three types of neighbors: temporal neighbors (same site, different times), spatial neighbors (same time, different sites), and spatiotemporal neighbors (across sites and times). For each type of neighbor, an attention mechanism is used to calculate aggregation weights. Through a learnable linear transformation, the representations of the query node and neighboring nodes are mapped to the same space. A similarity score is calculated and normalized to obtain the attention coefficient, adaptively identifying important neighbors and suppressing noise interference.

[0040] The three types of messages are fused after independent transformation, then subjected to a nonlinear transformation via a multilayer perceptron. Finally, residual connections are used to add the updated representation to the original representation to alleviate the gradient vanishing problem. After layer stacking, the network outputs a depth spatiotemporal representation. This representation encodes the local spatiotemporal neighborhood structure information centered on the node.

[0041] S40: The deep spatiotemporal representation is decomposed into trend components, periodic components, and residual components using the temporal decomposition module; based on the trend components and the pre-calculated cross-site spatiotemporal autocorrelation function, positive and negative sample pairs for self-supervised contrastive learning are constructed, and the parameters of the spatiotemporal graph neural network encoder and the temporal decomposition module are optimized through the cross-site spatiotemporal contrastive loss function.

[0042] Step S40 is used to implement temporal decomposition and cross-site self-supervised contrastive learning: In order to separate the change patterns of multiple time scales and learn discriminative long-term features, this embodiment performs temporal decomposition on the deep representation and designs a cross-site self-supervised contrastive learning mechanism based on the decomposed components.

[0043] In some embodiments, the time-series decomposition module separates the representation sequence of each site into three components with explicit physical meaning. Trend component To reflect the long-term trend of wind speed variation, a multi-scale moving average operation is used to extract the data: the original representation sequence is smoothed by using sliding windows of different time scales (such as 3 days, 7 days, 14 days, 30 days, and 90 days), and the results of multiple scales are averaged to eliminate short-term fluctuations and highlight long-term evolution patterns.

[0044] Periodic components To reflect the seasonal and periodic changes in wind speed, a Fourier basis function fitting method is used: a set of typical cycles (such as daily, weekly, and annual cycles) is pre-defined, and the coefficients of the corresponding sine and cosine basis functions for each cycle are learned to reconstruct the seasonal pattern. The residual component is the part remaining after subtracting the trend and cycle from the original representation, reflecting random fluctuations and extreme events.

[0045] Self-supervised contrastive learning is based on trend components, utilizing long-term statistical correlations of wind speeds from multiple stations to construct a self-supervised signal. Specifically, random sampling is performed from multi-station time-series data. The first spatiotemporal window is used as the training batch, and the second... Each spatiotemporal window is represented as Includes site identifier Start time and window length The trend representation sequence within each window is subjected to time-averaged pooling to obtain a fixed-dimensional representation vector for that window. .

[0046] The key step is to construct a measure of the spatiotemporal correlation between windows. For any two spatiotemporal windows... and Calculate the time distance at its center time. Spatial distance between stations The correlation strength is obtained by querying the pre-calculated cross-site spatiotemporal autocorrelation function: in, The cross-site spatiotemporal autocorrelation function is obtained from global statistics of the training data. This function takes time lag and spatial distance as input and outputs the Pearson correlation coefficient of the wind speed sequences of two stations, reflecting the statistical correlation strength under a specific spatiotemporal interval.

[0047] Based on this correlation metric, a relative ranking strategy is used to separate positive and negative samples. For the anchoring window... Its positive sample set Includes batches with a correlation higher than the threshold Other windows, which have strong long-term spatiotemporal correlations with the anchor window, may be in similar seasonal phases or be affected by the same weather systems. The negative sample partitioning employs a dynamic relative strategy: for each positive sample, all correlations not exceeding [a certain threshold] are considered. All other windows are considered as negative samples corresponding to the anchor sample. The advantage of this relative strategy is that it eliminates the need to set an absolute threshold and adapts to the diversity of data distribution. The cross-site spatiotemporal contrast loss function is designed as follows: in, This represents the total number of windows in the batch. Anchor window The number of positive samples; For window and The cross-site spatiotemporal autocorrelation intensity, as a weighting factor, makes the highly correlated positive samples contribute more; To characterize the cosine similarity between vectors; The temperature coefficient hyperparameter controls the sharpness of the similarity distribution; For window The corresponding negative sample set; summing the denominator by traversing the negative sample set and the positive sample set. The loss function, by maximizing the ratio of similarity between positive and negative sample pairs, forces the model to cluster spatiotemporally relevant samples and push away spatiotemporally unrelated samples in the feature space, thereby learning discriminative long-term spatiotemporal features.

[0048] S50: Construct a decompositional long-term forecasting architecture, which takes trend component, periodic component and residual component as input, and outputs the future wind speed forecast results of each wind measurement station in the target area; combine the spatiotemporal graph neural network encoder, time series decomposition module and decompositional long-term forecasting architecture to form a forecasting model, and train the forecasting model with a two-stage training strategy of self-supervised pre-training and supervised fine-tuning, and realize multi-station long-term wind speed forecasting based on the trained forecasting model.

[0049] In some embodiments, during the training of the prediction model, a decompositional long-term prediction architecture is constructed to predict and fuse the three components—trend, period, and residual—separately. A two-stage training strategy is employed to optimize the model parameters. The trend prediction branch takes the trend representation output by the deep encoder as input, aggregates spatiotemporal information through global pooling, maps it to the prediction timeline via a multilayer perceptron, and outputs the trend prediction values ​​for each station for the next few days. ,in Corresponding to different prediction durations; the periodic prediction branch directly performs analytical extrapolation based on the Fourier coefficients fitted from historical data, extending the sine and cosine basis functions to future times to generate seasonal predictions. This branch requires no network parameters, ensuring strict continuity of the periodicity. The residual prediction branch takes the recent residual sequence as input, performs short-time extrapolation using a lightweight linear model, and outputs the residual correction. The final prediction is obtained by adding the three components together: The advantage of this decompositional design is that each branch focuses on a specific time scale, avoiding mutual interference between signals of different scales. At the same time, the analytical extrapolation of the periodic branches ensures the physical rationality of the seasonal pattern.

[0050] In some embodiments, the training process is divided into two stages. The first stage is self-supervised pre-training, which optimizes only the spatiotemporal encoder and the temporal decomposition module, with the cross-station spatiotemporal contrast loss as the optimization objective, enabling the model to learn the long-term spatiotemporal features of wind speed at multiple stations under unlabeled conditions. This stage typically requires more epochs to fully structure the feature space. The second stage is supervised fine-tuning, which loads the pre-trained encoder weights as initialization and jointly optimizes the prediction branch parameters. The encoder backbone can be frozen or fine-tuned using a smaller learning rate. The optimization objective is a weighted combination of the prediction mean square error and the self-supervised loss. The advantage of this two-stage strategy is that the pre-training stage fully utilizes a large amount of unlabeled historical data to construct a general spatiotemporal representation, while the fine-tuning stage adapts to a specific prediction task, improving both data efficiency and prediction accuracy.

[0051] In some embodiments, the method further includes a model evaluation step, which comprehensively measures the model's predictive performance and physical plausibility through a series of metrics. First, deterministic evaluation metrics such as root mean square error (RMSE) and mean absolute error (MAE) are used to quantify the deviation between the model's predicted values ​​and observed values. Simultaneously, correlation metrics, including Pearson correlation coefficient and coefficient of determination, are used to assess the consistency between predicted and actual trends. Additionally, probabilistic metrics, including continuous hierarchical probability scores and calibration error, are used to assess the accuracy of uncertainty quantification. Furthermore, to evaluate the effectiveness of contrastive learning pre-training, cosine similarity in the feature space is calculated to measure the clustering of similar samples. Finally, spatial visualization techniques are used to visually display long-term prediction results for different geographical regions to verify the model's spatial generalization ability, analyze its performance differences under different terrain and climatic conditions, and explain potential identification biases.

[0052] In some embodiments, a method is constructed based on multi-station measured data (including wind speed sequences and related meteorological elements such as temperature, humidity, and air pressure). Figure 2 The framework shown is a cross-station spatio-temporal self-supervised wind prediction (CS-STSWP) framework that integrates a cross-station spatio-temporal self-supervised learning mechanism. This prediction framework can be used to implement the aforementioned long-term wind speed prediction method. It constructs physically meaningful spatio-temporal sample similarity constraints by mining the long-term statistical correlations formed between different wind measurement stations under different time lag conditions, and guides the prediction model to learn the long-term evolution characteristics of multi-station wind speed without manual annotation.

[0053] Specifically, the framework first constructs a cross-station spatiotemporal heterogeneous map based on multi-station historical observation data, and explicitly models three types of dependencies: intra-station temporal evolution, inter-station spatial correlation, and cross-station time-delay propagation. Then, it designs a contrastive learning mechanism based on global spatiotemporal autocorrelation to extend traditional single-station temporal modeling to multi-station collaborative representation learning, enabling the model to capture long-term variation patterns such as seasonal cycles, weather-scale processes, and topographic modulation effects that exceed the sliding window length.

[0054] Furthermore, this embodiment introduces a self-supervised contrastive learning mechanism based on spatiotemporal statistical correlation during model training. In the feature space, it explicitly constrains the clustering of sample features with strong long-term spatiotemporal correlations and the separation of sample features with weaker correlations, thereby enhancing the model's ability to represent seasonal scale changes, weather system evolution, and cross-station wind speed co-variation patterns. This mechanism quantifies the statistical correlation strength at different stations and time lags by pre-calculating the cross-station spatiotemporal autocorrelation function of multi-station wind speed sequences, and constructs dynamic positive and negative sample pairs accordingly, enabling the model to learn discriminative long-term spatiotemporal features under unlabeled conditions. Through deep self-supervised learning of historical meteorological data combined with spatiotemporal dependency modeling, this invention demonstrates significant advantages in long-term multi-station wind speed prediction accuracy, prediction stability, and engineering application reliability. It can be widely applied in wind power assessment, wind resource analysis, new energy grid-connected scheduling, and smart meteorological services, possessing good promotional value and application prospects.

[0055] In summary, this embodiment constructs a multi-station long-term wind speed prediction framework (CS-STSWP) based on cross-station spatiotemporal self-supervised learning. The advancements of the method in this application are reflected in the following three aspects: First, a cross-station spatiotemporal heterogeneous graph construction mechanism is proposed. Multi-station wind speed time-series data are modeled as a heterogeneous graph, defining three types of edges: temporal edges based on temporal autocorrelation to capture intra-station dynamic evolution; spatial edges based on the correlation between geographical distance and wind speed to capture inter-station synchronous correlation; and spatiotemporal edges based on cross-station spatiotemporal autocorrelation to capture wind field propagation effects. Through optimal propagation delay calculation, dynamically time-aligned cross-station connections are established, embedding the physical laws of wind field evolution at the graph structure level, providing a physically consistent spatiotemporal topological foundation for subsequent feature learning.

[0056] Secondly, a cross-site spatiotemporal self-supervised contrastive learning mechanism is designed. Based on the cross-site spatiotemporal autocorrelation function, the correlation strength between windows is calculated. A relative ranking strategy is used to divide positive and negative samples: positive samples are window pairs with spatiotemporal autocorrelation exceeding a threshold, and negative samples are windows with relatively low autocorrelation within the batch. Spatial negative samples are also introduced to enhance cross-site discriminative ability. The trend term representation space is optimized through weighted contrastive loss, enabling the clustering of spatiotemporally similar long-term wind speed evolution patterns and the separation of dissimilar patterns. This allows for the learning of discriminative spatiotemporal characteristics of multi-station wind speeds without manual annotation.

[0057] Third, a decompositional long-term forecasting architecture is established. The spatiotemporal encoder is decoupled from the forecasting branch. During the pre-training phase, the encoder is optimized through self-supervised learning. During the fine-tuning phase, the encoder is frozen to preserve the spatiotemporal consistency of wind speed across multiple stations, and only the decompositional forecasting branch is optimized. The trend branch predicts long-term trends based on deep representations, the periodic branch extrapolates seasonal patterns based on Fourier basis, and the residual branch corrects short-term fluctuations based on recent observations. The fusion of these three components achieves high-precision, physically consistent forecasts of wind speed across multiple stations over 30-90 days.

[0058] The method proposed in this application combines cross-station spatiotemporal self-supervised learning with a decompositional prediction architecture, integrating multi-station spatial correlation constraints with long-term time series modeling. This not only breaks through the limitation of sliding window length but also makes full use of cross-station spatial information, significantly improving the accuracy, stability, and physical reliability of long-term multi-station wind speed prediction, and providing core technical support for new energy meteorological prediction.

[0059] This application also provides a multi-station long-term wind speed prediction system based on cross-station spatiotemporal statistical correlation self-supervised learning, such as... Figure 3 As shown, this multi-station long-term wind speed prediction system based on cross-station spatiotemporal statistical correlation self-supervised learning includes: Data preprocessing module 301 is configured to acquire data within the target area. M The historical meteorological monitoring datasets of each wind measurement station were preprocessed to obtain multi-station time-series meteorological data with a unified time reference and continuous and complete data. Heterogeneous graph construction module 302 is configured to construct a cross-station spatiotemporal heterogeneous graph based on the multi-station time-series meteorological data. Each node in the node set corresponds to the observation status of the wind measurement station at a certain moment, and the edge set includes the time edge set. Spatial edge set and spacetime edge set ; The encoder module 303 is configured to construct a spatiotemporal graph neural network encoder, which aggregates three types of information from each node’s temporal neighbors, spatial neighbors and spatiotemporal neighbors through a layer-by-layer message passing mechanism to obtain a deep spatiotemporal representation. The self-supervised learning module 304 is configured to perform temporal decomposition on the deep spatiotemporal representation using the temporal decomposition module to obtain trend components, periodic components and residual components; based on the trend components and the pre-calculated cross-site spatiotemporal autocorrelation function, positive and negative sample pairs for self-supervised contrastive learning are constructed, and the parameters of the spatiotemporal graph neural network encoder and the temporal decomposition module are optimized through the cross-site spatiotemporal contrastive loss function. The long-term prediction module 305 is configured to construct a decompositional long-term prediction architecture. The decompositional long-term prediction architecture takes trend components, periodic components, and residual components as inputs and outputs the future wind speed prediction results of each wind measurement station in the target area. The spatiotemporal graph neural network encoder, the time series decomposition module, and the decompositional long-term prediction architecture are combined to form a prediction model. The prediction model is trained using a two-stage training strategy of self-supervised pre-training and supervised fine-tuning. Based on the trained prediction model, long-term prediction of wind speed at multiple stations is realized.

[0060] It should be noted that the multi-station long-term wind speed prediction system based on cross-station spatiotemporal statistical correlation self-supervised learning mentioned above belongs to the same technical concept as the prior method and can achieve the same technical effect, so it will not be elaborated here.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A long-term multi-station wind speed prediction method based on cross-station spatiotemporal statistical correlation self-supervised learning, characterized in that, The method includes: Get the target area M The historical meteorological monitoring datasets of each wind measurement station were preprocessed to obtain multi-station time-series meteorological data with a unified time reference and continuous and complete data. Based on the multi-station time-series meteorological data, a cross-station spatiotemporal heterogeneous map is constructed. Each node in the node set corresponds to the observation status of the wind measurement station at a certain moment, and the edge set includes the time edge set. Spatial edge set and spacetime edge set ; A spatiotemporal graph neural network encoder is constructed. Through a layer-by-layer message passing mechanism, the three types of information from each node—temporal neighbors, spatial neighbors, and spatiotemporal neighbors—are aggregated to obtain a deep spatiotemporal representation. The deep spatiotemporal representation is decomposed into trend components, periodic components, and residual components using a time-series decomposition module. Based on the trend components and the pre-calculated cross-site spatiotemporal autocorrelation function, positive and negative sample pairs for self-supervised contrastive learning are constructed. The parameters of the spatiotemporal graph neural network encoder and the time-series decomposition module are optimized using a cross-site spatiotemporal contrastive loss function. A decompositional long-term forecasting architecture is constructed, which takes trend component, periodic component and residual component as input and outputs the future wind speed forecast results of each wind measurement station in the target area. The spatiotemporal graph neural network encoder, the time series decomposition module and the decompositional long-term forecasting architecture are combined to form a forecasting model. The forecasting model is trained using a two-stage training strategy of self-supervised pre-training and supervised fine-tuning. Based on the trained forecasting model, long-term wind speed forecasting for multiple stations is realized.

2. The method according to claim 1, characterized in that, Within the target area M Historical meteorological monitoring dataset from [number] wind measurement stations, [number] Each station at time The observation vector is: in, u m,t , v m,t They are time points t Wind measurement station m The east-west component and the north-south component of the wind speed. T m,t For a moment t Wind measurement station m temperature P m,t For a moment t Wind measurement station m air pressure, φ m , λ m , h m These are wind measurement stations m Latitude, longitude, and altitude; The preprocessing of historical meteorological monitoring datasets includes sequential steps such as station screening and time alignment, outlier repair and missing value completion, wind direction coding and feature standardization, and data format regularization. Among these steps, station screening ensures that the maximum distance between wind measurement stations does not exceed a preset threshold. Time alignment ensures that all data timestamps are in the same time zone and have the same time resolution.

3. The method according to claim 1, characterized in that, The time edge set Used to model the dynamic evolution relationship between different times within the same wind measurement station, for wind measurement stations m Connection time and The time-domain weights are based on the time-domain autocorrelation coefficient of the wind speed sequence at that wind measurement station. Sure, The statistical correlation strength of wind speed sequences at the same wind measurement station at a lag time τ is quantified, with the time edge weights being the autocorrelation exponential decay function. The set of spatial edges Used to model the spatial correlation structure between different wind measurement stations at the same time, for any two wind measurement stations m and n ,time t Spatial boundary weighting of the great circle distance between two wind measurement stations The Pearson correlation coefficient with historical wind speed series was determined, where The results were obtained from the latitude and longitude coordinates of the two wind measurement stations using the spherical trigonometry formula. The set of spatiotemporal edges To capture the wind field propagation effect between different wind measurement stations at different times, the weights of the spatiotemporal edges are calculated using a pre-computed cross-station spatiotemporal autocorrelation function. The absolute value, Historical data statistics reflect the weather measurement stations m With wind measurement stations n Time lag Spatial distance Wind speed correlation strength under certain conditions.

4. The method according to claim 1, characterized in that, In each layer of the spatiotemporal graph neural network encoder, each node simultaneously receives messages from three types of neighbors: temporal neighbors, spatial neighbors, and spatiotemporal neighbors. For each type of neighbor, an attention mechanism is used to calculate the aggregation weight. Through a learnable linear transformation, the representations of the query node and the neighbor nodes are mapped to the same feature space. The similarity score is calculated and normalized to obtain the attention coefficient. The three types of messages are fused after independent linear transformations, and then nonlinearly transformed by a multilayer perceptron. The updated representation is added to the original representation using residual connections. After stacking L layers of network, the depth spatiotemporal representation of the output node is obtained.

5. The method according to claim 1, characterized in that, The trend component is extracted through multi-scale moving average operation, and the original depth spatiotemporal representation sequence is smoothed by multiple sliding windows with different time scales, and the average multi-scale smoothing result is obtained; the periodic component is obtained by fitting the Fourier basis function with a preset typical period; the residual component is the remaining part after subtracting the trend component and the periodic component from the original depth spatiotemporal representation.

6. The method according to claim 1, characterized in that, Based on the trend component and the pre-calculated cross-site spatiotemporal autocorrelation function, positive and negative sample pairs for self-supervised contrastive learning are constructed, specifically including: Random sampling from multi-station time-series meteorological data N One spatiotemporal window is used as the training batch; Calculate any two spatiotemporal windows and Time distance Spatial distance between stations The correlation strength is obtained by querying the pre-calculated cross-site spatiotemporal autocorrelation function. : in, This is the cross-site spatiotemporal autocorrelation function obtained from global statistics of the training data; For anchoring windows Its positive sample set Includes batches with a correlation higher than the threshold Other windows; For each anchor window, all relevance levels not exceeding [a certain threshold] are [not specified]. All other windows are considered negative samples.

7. The method according to claim 6, characterized in that, The cross-site spatiotemporal comparison loss function is: in, This represents the total number of windows in the batch. For window The number of positive samples; For window and The cross-site spatiotemporal autocorrelation intensity, as a weighting factor, makes the highly correlated positive samples contribute more; To characterize the cosine similarity between vectors; z i and z j For window and The fixed-dimensional representation vector is obtained by performing time-averaged pooling on the trend representation sequence within the window. The temperature coefficient hyperparameter controls the sharpness of the similarity distribution; For window The corresponding set of negative samples; exp is the natural exponential function; This represents the loss in cross-site spatiotemporal comparison.

8. The method according to claim 5, characterized in that, The decompositional long-term forecasting architecture includes a trend forecasting branch, a cycle forecasting branch, and a residual forecasting branch, wherein: The trend prediction branch takes the trend component as input, aggregates spatiotemporal information through global pooling operation, maps it to the prediction time through a multilayer perceptron, and outputs the trend prediction value of each wind measurement station. The periodic prediction branch performs analytical extrapolation based on the Fourier basis function coefficients obtained from the periodic components, extending the sine and cosine basis functions to future times to generate seasonal prediction values. The residual prediction branch takes the recent sequence of the residual component as input, performs short-time extrapolation using a lightweight linear model, and outputs the residual correction value. The final wind speed forecast is obtained by adding and fusing the trend forecast, seasonal forecast, and residual correction values.

9. The method according to claim 1, characterized in that, The two-stage training strategy includes a first stage and a second stage, wherein: The first stage is a self-supervised pre-training stage, which only optimizes the parameters of the spatiotemporal graph neural network encoder and the temporal decomposition module, with the cross-site spatiotemporal contrast loss function as the optimization target; The second stage is the supervised fine-tuning stage, in which the weights of the self-supervised pre-trained spatiotemporal graph neural network encoder and the temporal decomposition module are loaded as initialization, and the branch parameters of the decompositional long-term prediction architecture are jointly optimized. The optimization objective is a weighted combination of the prediction mean square error and the cross-station spatiotemporal comparison loss. In the fine-tuning stage, the backbone parameters of the spatiotemporal graph neural network encoder are frozen, and only the decompositional prediction branch is optimized.

10. A multi-station long-term wind speed forecasting system based on cross-station spatiotemporal statistical correlation self-supervised learning, used to implement the method as described in any one of claims 1 to 9, characterized in that, The system includes: The data preprocessing module is configured to acquire data within the target area. M The historical meteorological monitoring datasets of each wind measurement station were preprocessed to obtain multi-station time-series meteorological data with a unified time reference and continuous and complete data. The heterogeneous graph construction module is configured to construct a cross-station spatiotemporal heterogeneous graph based on the multi-station time-series meteorological data. Each node in the node set corresponds to the observation status of the wind measurement station at a certain moment, and the edge set includes the time edge set. Spatial edge set and spacetime edge set ; The encoder module is configured to build a spatiotemporal graph neural network encoder. Through a layer-by-layer message passing mechanism, it aggregates three types of information from each node's temporal neighbors, spatial neighbors, and spatiotemporal neighbors to obtain a deep spatiotemporal representation. The self-supervised learning module is configured to perform temporal decomposition on the deep spatiotemporal representation using the temporal decomposition module to obtain trend components, periodic components, and residual components; based on the trend components and the pre-calculated cross-site spatiotemporal autocorrelation function, positive and negative sample pairs for self-supervised contrastive learning are constructed, and the parameters of the spatiotemporal graph neural network encoder and the temporal decomposition module are optimized through the cross-site spatiotemporal contrastive loss function; The long-term forecasting module is configured to construct a decompositional long-term forecasting architecture. The decompositional long-term forecasting architecture takes trend components, periodic components, and residual components as inputs and outputs the future wind speed forecast results for each wind measurement station in the target area. The spatiotemporal graph neural network encoder, the time series decomposition module, and the decompositional long-term forecasting architecture are combined to form a forecasting model. The forecasting model is trained using a two-stage training strategy of self-supervised pre-training and supervised fine-tuning. Based on the trained forecasting model, long-term wind speed forecasting for multiple stations is realized.

Images