Wind speed prediction method, system, equipment and medium

Abstract

The invention discloses a wind speed prediction method, system and device and a medium, and the method comprises the following steps: carrying out the preprocessing of historical wind speed data, and obtaining a historical time sequence; partitioning the historical time sequence to obtain multivariate time sequence blocks; performing feature conversion on the multivariate time sequence block to obtain a first sequence coding feature; performing space-time attention coding on the first sequence coding feature to obtain a target sequence coding feature; and performing linear prediction on the coding features of the target sequence to obtain a predicted wind speed value time sequence. According to the method, the time dependency relationship of each meteorological acquisition station can be effectively captured, and the prediction accuracy is remarkably improved; the correlation between different meteorological acquisition stations can be modeled, and the prediction performance is further improved; meanwhile, the system can be widely applied to the fields of wind energy production, weather forecast, climate research and the like, and the safety and reliability of a wind energy system can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological prediction, and particularly to a wind speed prediction method, system, device and medium. Background Art

[0002] With the rapid development of the global economy, the excessive consumption of traditional fossil fuels has caused serious environmental pollution problems, such as greenhouse gas emissions, air pollution, etc. These problems pose severe challenges to the human living environment and sustainable development. Therefore, achieving the transformation of the energy structure and vigorously developing clean and renewable new energy has become a global consensus. As a clean and pollution-free energy form, wind energy has become a key area of global energy research and development due to its large reserves, wide distribution and renewable characteristics. Therefore, improving the accuracy of wind speed prediction is crucial for optimizing the operation of wind power systems, enhancing the stability of power grids and improving energy utilization efficiency. However, the development and utilization of wind energy is not easy, and the instability of wind speed is a key factor affecting the efficient utilization of wind energy. The randomness and volatility of wind speed make it difficult to stabilize the output power of wind power generation, posing a huge challenge to the stable operation of the power grid. Due to the strong intermittency, randomness, uncontrollability, etc. of wind speed, the development cost of wind power is relatively high, and it is also not conducive to the stable operation of the wind power grid-connected system. Therefore, accurate wind speed prediction plays a crucial role in improving wind energy utilization efficiency and ensuring the safe and stable operation of the power grid.

[0003] Existing wind speed prediction methods are divided into three categories: physical methods, statistical methods and deep learning-based methods. Physical methods rely on atmospheric observations or numerical weather forecasts, with high costs and are not suitable for short-term predictions. Statistical methods estimate time patterns from historical data through probabilistic methods, such as ARIMA, SVR, etc.; however, the ARIMA model may lose information when dealing with non-stationary time series; the SVR model has a complex optimization process. The Transformer model has achieved success in NLP and CV tasks and has also been used for time series prediction. However, existing Transformer models have limitations: they do not fully consider the complexity and differences of wind speed in different scenarios; it is difficult to capture the complex dependencies between variables when dealing with multivariate time series prediction, and the computational complexity is high, affecting prediction accuracy. Summary of the Invention

[0004] The objective of the present invention is to provide a wind speed prediction method, system, device and medium to solve the problem that the existing wind speed prediction methods have high costs and computational complexities, affecting prediction accuracy.

[0005] To solve the above technical problems, a technical solution adopted by the present invention is to provide a wind speed prediction method, and the method includes the following steps: preprocess the historical wind speed data to obtain a historical time series; partition the historical time series to obtain a multivariate time series block; perform feature transformation on the multivariate time series block to obtain a first sequence encoding feature; perform spatio-temporal attention encoding on the first sequence encoding feature to obtain a target sequence encoding feature; perform linear prediction on the target sequence encoding feature to obtain a predicted wind speed value time series.

[0006] The present invention also provides a wind speed prediction system, and the system includes: a preprocessing unit for preprocessing the historical wind speed data to obtain a discrete time series; a partitioning unit for partitioning the historical time series to obtain a multivariate time series block; a feature transformation unit for performing feature transformation on the multivariate time series block to obtain a first sequence encoding feature; a spatio-temporal attention encoding unit for performing spatio-temporal attention encoding on the first sequence encoding feature to obtain a target sequence encoding feature; a linear prediction unit for performing linear prediction on the target sequence encoding feature to obtain a predicted wind speed value time series.

[0007] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the steps of the above method are implemented.

[0008] The present invention also provides a computer-readable storage medium, and the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0009] The beneficial effects of the present invention are as follows: The present invention discloses a wind speed prediction method, system, device, and medium. The method includes the following steps: preprocess the historical wind speed data to obtain a historical time series; partition the historical time series to obtain a multivariate time series block; perform feature transformation on the multivariate time series block to obtain a first sequence encoding feature; perform spatio-temporal attention encoding on the first sequence encoding feature to obtain a target sequence encoding feature; perform linear prediction on the target sequence encoding feature to obtain a predicted wind speed value time series. This method can effectively capture the time-dependent relationship of each meteorological collection station, significantly improve the prediction accuracy; can also model the correlation between different meteorological collection stations, further improving the prediction performance; at the same time, it can be widely applied to fields such as wind energy production, weather forecasting, and climate research, which helps to improve the safety and reliability of the wind energy system. Description of the Drawings

[0010] Figure 1 It is a schematic flowchart of an embodiment of a wind speed prediction method of the present invention;

[0011] Figure 2 It is a schematic diagram of the distribution of meteorological collection stations in a land scenario in an embodiment of a wind speed prediction method of the present invention;

[0012] Figure 3 It is a schematic diagram of the distribution of meteorological collection stations in an offshore scenario in an embodiment of a wind speed prediction method of the present invention;

[0013] Figure 4 It is a schematic diagram of the distribution of meteorological collection stations in a desert scenario in an embodiment of a wind speed prediction method of the present invention;

[0014] Figure 5 It is a schematic diagram of the composition of a wind speed prediction network in an embodiment of a wind speed prediction method of the present invention;

[0015] Figure 6 It is a schematic diagram of the specific composition of a spatio-temporal attention encoding module in an embodiment of a wind speed prediction method of the present invention;

[0016] Figure 7 It is a partial data flow diagram in an embodiment of a wind speed prediction method of the present invention;

[0017] Figure 8 It is a specific flowchart of step S4 in an embodiment of a wind speed prediction method of the present invention;

[0018] Figure 9 It is a comparison diagram of the prediction results of a wind speed prediction network and multiple existing network models in an embodiment of a wind speed prediction method of the present invention;

[0019] Figure 10 It is a block diagram of the composition of an embodiment of a wind speed prediction system of the present invention;

[0020] Figure 11 It is a schematic diagram of the architecture of an embodiment of an electronic device of the present invention;

[0021] Figure 12 It is a schematic block diagram of an embodiment of a computer-readable storage medium of the present invention. Specific Embodiments

[0022] For ease of understanding of the present invention, the present invention will be described in more detail below with reference to the accompanying drawings and specific embodiments. Preferred embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0023] It should be noted that, unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used in this specification includes any and all combinations of one or more of the related listed items.

[0024] As Figure 1 shown, a wind speed prediction method of the present invention is shown, including the following steps:

[0025] Step S1: Preprocess the historical wind speed data to obtain a historical time series.

[0026] It should be noted that, in one scenario, generally there are multiple meteorological collection stations for collecting meteorological data. The historical wind speed data is a variety of meteorological data related to wind speed collected by each meteorological collection station in one scenario. The historical time series is a multivariate time series obtained by arranging the historical wind speed data in the order of time steps. Each element in the historical time series is a feature matrix, including a variety of meteorological data related to wind speed collected by each meteorological collection station.

[0027] Before step S1, it is also necessary to construct a wind speed data set and a wind speed prediction network I2A-Mamba. The wind speed data set includes historical wind speed data from different scenarios. The wind speed prediction network I2A-Mamba is used to predict the wind speed according to the historical time series to obtain the future predicted wind speed.

[0028] In the present application, a wind speed data set LODWS including three scenarios of land (non-desert area), sea and desert is constructed to evaluate the robustness and adaptability of the wind speed prediction network I2A-Mamba in different natural environments. Among them, the three scenarios of land (non-desert area), sea and desert cover typical terrain, climate and surface conditions. Selecting these scenarios can better reflect different environmental conditions affecting wind speed dynamics.

[0029] Furthermore, the land wind speed data covers typical vegetation, urban and rural landscapes, which helps to understand how surface roughness and obstacles such as buildings and trees affect wind patterns.

[0030] As Figure 2As shown in the figure, for a land scenario (non-desert area), there are 9 meteorological collection stations that can record all relevant meteorological data during certain periods. For this land scenario, 9 meteorological variables are selected from all relevant meteorological variables, namely "surface air pressure measured on an hourly average, air pressure at 100 meters, humidity at 2 meters, surface temperature, temperature at 100 meters, wind speed at 20 meters, wind speed at 120 meters, wind direction at 10 meters, and wind direction at 100 meters". The 9 types of meteorological data corresponding to these 9 meteorological variables are used as the historical wind speed data for this land scenario. Furthermore, these 9 types of meteorological data can be used as input features for the wind speed prediction network I2A-Mamba to predict the wind speed at future times for this land scenario.

[0031] Furthermore, the offshore wind speed data focuses on coastal areas, where the wind speed is affected by factors such as sea surface temperature, atmospheric stability, and offshore distance.

[0032] As Figure 3 shown in the figure, for an offshore scenario, there are 14 meteorological collection stations that can record all relevant meteorological data during certain periods. For this offshore scenario, 9 meteorological variables are selected from all relevant meteorological variables, namely "air temperature, air pressure, dew point, relative humidity, wind direction, and wind speed measured on a 15-minute average, and the maximum wind speed within a 15-minute interval". The 9 types of meteorological data corresponding to these 9 meteorological variables are used as the historical wind speed data for this offshore scenario. Furthermore, these 9 types of meteorological data can be used as input features for the wind speed prediction network I2A-Mamba to predict the wind speed at future times for this offshore scenario.

[0033] Furthermore, the desert wind speed data takes into account the extreme conditions of the arid environment, and dunes, lack of vegetation, and high temperatures have a significant impact on wind speed dynamics.

[0034] As Figure 4 shown in the figure, for a desert scenario, there are 7 meteorological collection stations that can completely measure all relevant meteorological data during certain periods. For this desert scenario, 9 meteorological variables are selected from all relevant meteorological variables, namely "surface air pressure measured on an hourly average, air pressure at 100 meters, humidity at 2 meters, surface temperature, temperature at 100 meters, wind speed at 10 meters, wind speed at 100 meters, wind direction at 10 meters, and wind direction at 100 meters". The 9 types of meteorological data corresponding to these 9 meteorological variables are used as the historical wind speed data for this desert scenario. Furthermore, these 9 types of meteorological data can be used as input features for the wind speed prediction network I2A-Mamba to predict the wind speed at future times for this desert scenario.

[0035] In this application, the wind speed dataset composed of historical wind speed data corresponding to multiple scenarios is crucial for developing accurate wind energy assessment, weather forecasting models, and climate research. Among them, 70% of the data in the wind speed dataset is used for training, and the subsequent 10% and 20% are reserved for validation and testing respectively. If any measurement value at a certain site is missing at a certain time step, linear interpolation is used to fill in the missing value.

[0036] As Figure 5 shown, the wind speed prediction network I2A-Mamba includes a preprocessing module 1, a chunking module 2, a feature transformation module 3, a spatio-temporal attention encoding module 4, and a linear prediction module 5, and the characters on the left side of the figure show the sizes of the output features of each module in this embodiment.

[0037] Furthermore, after obtaining the wind speed dataset and the wind speed prediction network I2A-Mamba, it is necessary to use the preprocessing module 1 in the wind speed prediction network I2A-Mamba to preprocess the corresponding historical wind speed data to obtain a historical time series that can be recognized by the chunking module 2, that is, enter step S1.

[0038] In this embodiment, instance normalization is used for preprocessing. The historical wind speed data input into the network is instance-normalized to obtain the corresponding historical time series.

[0039] It should be noted that instance normalization is a data preprocessing technique. It calculates the mean and standard deviation separately for each channel of each input sample and uses these statistics to normalize the elements within each channel of the sample. This normalization method is independent of batches and channels, and each sample has its own normalization parameters. During the network training phase, the preprocessing module 1 is used to instance-normalize a given sample and convert it into an instance-normalized form (reflected by multivariate time series in this application) to reduce distribution shift. Among them, converting to the instance-normalized form refers to a preprocessing operation performed on the data in a given sample during the training phase. The purpose of this operation is to reduce distribution shift, make the data have a more consistent distribution characteristic between different samples, help the network better learn data features, and improve the training effect and generalization ability of the network.

[0040] In testing or actual applications, new input samples also need to be instance-normalized to ensure that the data distribution is consistent with that during training, thereby improving the generalization ability and prediction accuracy of the wind speed prediction network I2A-Mamba. Specifically, when using the wind speed prediction network I2A-Mamba trained with instance normalization for prediction, the newly input samples should be preprocessed according to the same normalization parameters (i.e., the mean and standard deviation used during training). This can ensure that the wind speed prediction network I2A-Mamba can maintain the learned feature representations and decision boundaries when processing new data, thus making accurate predictions.

[0041] Specifically, in a scenario, historical wind speed data corresponding to multiple time steps are respectively represented as a feature matrix, where a feature matrix includes various meteorological data measured by multiple meteorological collection stations in a scene at a time step, that is, each time step corresponds to a feature matrix. Then, multiple feature matrices corresponding to multiple time steps form a historical time series, and by inputting this historical time series into the chunking module 2 in the wind speed prediction network I2A-Mamba, the specific operations of steps S2 - S5 can be sequentially performed by the wind speed prediction network I2A-Mamba to obtain the predicted wind speeds corresponding to the locations of each meteorological collection station at multiple future time steps, that is, to obtain a time series of predicted wind speed values. Therefore, the wind speed prediction problem can be expressed as:

[0042]

[0043] where J represents the number of future time steps to be predicted; represents the predicted wind speed at the locations of n meteorological collection stations at time step (t + j), and is a vector containing n elements, and each element corresponds to the predicted wind speed value at the location of a station at time step (t + j); F represents the wind speed prediction network I2A-Mamba, which is an algorithm or function that realizes the conversion from input data to prediction results; (f (t) , f (t-1) ,..., f (t-(I-1)) ) represents the historical time series; f (t-i) ∈R n×9 represents the feature matrix corresponding to time step (t - i), including 9 recorded measurement values respectively measured by n meteorological collection stations, I represents the retrospective window, that is, the number of historical time steps used for prediction.

[0044] Furthermore, I = 5 and J = 3. That is, the wind speed prediction network I2A-Mamba will refer to the current time step t and the historical wind speed data corresponding to the past 5 time steps (t-1), (t-2), (t-3), (t-4) to predict the wind speed at the 3 future time steps (t+1), (t+2), (t+3).

[0045] Taking Figure 4 the desert scene shown as an example, at 09:00:00+00:00 on September 13, 2020, the 9 meteorological data collected by the meteorological collection site OakCreek RD were (90860, 89840, 5.93, 8.16, 35.37, 5.62, 62.38, 66.26, 25.79) respectively, and the 9 meteorological data collected by the meteorological collection site Mojave were (92020, 90980, 4.9, 8.02, 35.66, 4.6, 53.91, 57.61, 26.41) respectively. The 9 meteorological data collected by the meteorological collection site Chaffee were (91820, 90780, 7.94, 35.81, 26.28, 5.02, 5.42, 50.23, 47.35) respectively. The 9 meteorological data collected by the meteorological collection site Bakersfield Tehachapi Hwy were (89210, 88200, 7.56, 8.07, 34.07, 7.1, 61.08, 65.75, 25.49) respectively. The 9 meteorological data collected by the meteorological collection site Barstow-Bakersfield Hwy were (92230, 91190, 4.3, 36.36, 26.18, 4.07, 46.33, 48.47, 8.11) respectively. The 9 meteorological data collected by the meteorological collection site Mojave-Barstow Hwy were (92740, 91690, 3.09, 8.94, 35.46, 2.9, 46.98, 46.09, 25.43) respectively. The 9 meteorological data collected by the meteorological collection site Californla City Blvd were (92500, 91460, 2.91, 8.27, 26.15, 2.86, 36.82, 37.55, 36.85) respectively.

[0046] Therefore, the feature matrix corresponding to this scene at this time is:

[0047]

[0048] Furthermore, multiple feature matrices corresponding to multiple time steps in this scene are integrated to form a historical time series T = {T1, T2, T3,...}.

[0049] In this application, the historical wind speed data for each time step is encapsulated in a matrix, facilitating subsequent data processing and analysis. When dealing with meteorological time series data, it has the advantages of clear structure, time series continuity, scalability, convenience for machine learning applications, and improved analysis efficiency.

[0050] Step S2: Divide the historical time series into chunks to obtain multivariate time series chunks.

[0051] Combine Figure 5 and Figure 7 , and use the chunking module 2 to chunk the historical time series 10 to obtain the multivariate time series chunk 20.

[0052] Furthermore, given a historical time series 10 as X ∈ R L×V , L is the lookback window, which refers to the number of historical time steps that the wind speed prediction network I2A - Mamba uses for wind speed prediction; V is the number of channels, representing the number of different meteorological variables observed at each time step. Among them, X = x1,..., x L , x1 to x L represent the observations at different time steps, and x l ∈ R V , l ∈ (1, 2,..., L), that is, the observation at time step l is a vector containing V elements, corresponding to V different channels (meteorological variables).

[0053] It should be noted that each channel corresponds to an index. Therefore, X i,v ∈ R L can be represented as the univariate historical time series corresponding to the channel with index v, and v ∈ (1, 2,..., V). For each univariate historical time series X :,v , it can be divided into multiple univariate historical time series chunks using a sliding window with length P and step size S:

[0054]

[0055] Among them, Patching represents the chunking operation; represents multiple univariate historical time series chunks, and is the number of univariate historical time series chunks.

[0056] In this embodiment, first, the univariate historical time series of each channel are all converted into univariate historical time series chunks with the same chunk length P and chunk number N; then, the univariate historical time series chunks of different channels are combined to obtain the multivariate time series chunk 20. This process can be expressed as:

[0057]

[0058] Among them, represents the multivariate time series block 20, and

[0059] In this application, the above operations can ensure that all channel time series blocks have the same length P and the same number N, achieving data standardization, which is convenient for subsequent analysis and model training. At the same time, the chunking operation can be regarded as feature extraction. Each univariate historical time series block is used as a summary of the time window data to capture local features and patterns; then, the univariate historical time series blocks of different channels are combined to form a multivariate time series block (or called a window, segment), which can capture the interactions between channels.

[0060] Step S3: Perform feature transformation on the multivariate time series block to obtain the first sequence encoding feature.

[0061] As Figure 5 and Figure 7 shown, use the feature transformation module 3 to perform feature transformation on the multivariate time series block 20 to obtain the first sequence encoding feature 30.

[0062] Specifically, first mark each block of the multivariate time series block 20 and linearly project it into a vector of size D, and then perform learnable positional encoding to obtain the first sequence encoding feature 30. This process can be expressed as:

[0063]

[0064] Among them, Z0 represents the first sequence encoding feature 30, and Z0 ∈ R V×N×D ; W p ∈ R P×D ; W pos ∈ R N×D .

[0065] In this embodiment, by splitting the multivariate time series data into multivariate time series blocks 20 and processing each block, the local features in the time series can be captured more carefully. Then, the multivariate time series block 20 is mapped to a vector space of a fixed size D by the linear projection step, which helps to unify the feature dimensions between different blocks and is convenient for subsequent processing and analysis. Then, learnable positional encoding is introduced, enabling the network to capture the relative position information between elements in the time series, which helps to understand the dynamic changes of the time series data. The learnable positional encoding can be adaptively adjusted according to the data distribution and task requirements, thereby improving the generalization ability of the network model.

[0066] Through the above steps, the method can effectively extract the features corresponding to historical time series data, which contain both local information (through block-by-block marking and linear projection) and global position information (through learnable position encoding). These features can be used as the input of subsequent machine learning or deep learning models for various time series analysis tasks such as prediction, classification, and clustering.

[0067] In summary, through block-by-block marking, linear projection, and learnable position encoding, the method realizes the effective feature extraction of multivariate time series data, providing strong support for subsequent time series analysis tasks.

[0068] Step S4: Perform spatio-temporal attention encoding on the encoded features of the first sequence to obtain the encoded features of the target sequence.

[0069] Combined with Figure 5 and Figure 6 , the spatio-temporal attention encoding module 4 includes K I2A-Mamba modules 41. The I2A-Mamba module 41 includes an Intra-Station Mamba module 42 and an Inter-station Attention module 43, which are used to capture cross-time dependencies and cross-channel dependencies respectively.

[0070] It should be noted that Mamba shows great potential in the fields of natural language processing (NLP), computer vision (CV), and stock prediction. In these fields, the consistency of semantic information enables words, image patches, or stock metrics to be regarded as tokens. However, in multivariate time series, different channels may represent completely different physical phenomena, so it is inappropriate to regard different channels at the same time point as a single token. Although a single time step of each channel lacks intrinsic semantic meaning, by aggregating multiple time points into multivariate time series blocks at the subsequence level through the chunking module 2, the semantic content can be enriched and the local receptive field of the tokens can be extended. Therefore, this application retains the structure of the original Mamba module, and at the same time uses the chunking module to divide the input historical time series 10 into multivariate time series blocks 20, and then uses these multivariate time series blocks 20 as the input of the Intra-Station Mamba module 42.

[0071] Such as Figure 6As shown in the figure, the in-station Mamba module 42 includes a selective state space model (SSM), a convolutional layer (Conv1d), a linear layer (Linear), an activation function (Silu), and an RMS normalization layer (RMS Norm), etc. The in-station Mamba module 42 can perform selective state space processing on the input multivariate time series block 20, capture the complex dynamic features and dependencies within the univariate historical time series, and learn the patterns of the time series within a local range. The inter-station attention module 43 consists of an average pooling layer (AvgPool), a max pooling layer (MaxPool), a linear layer (Linear), activation functions (Gelu, Sigmoid), etc. Based on the output of the in-station Mamba module 42, the inter-station attention module 43 extracts the channel features through a pooling operation, and then through linear transformation and activation function processing, calculates the attention weights between different stations (channels), captures the interdependencies between different stations in the multivariate time series, synthesizes the information of each station, and improves the understanding and prediction ability of the overall time series.

[0072] Specifically, as Figure 8 shown, step S4 includes the following sub-steps:

[0073] Step S41: Perform selective state space processing on the first sequence encoded feature to obtain time series embedding features.

[0074] Combined with Figure 5 and Figure 7 , the first sequence encoded feature 30 is input into the spatio-temporal attention encoding module 4, and first undergoes block-by-block selective state space processing by the first in-station Mamba module 42 to obtain time series embedding features 40.

[0075] It should be noted that the selective state space model (SSM) adopts a selective state space mechanism, including a continuous state space mechanism and a discrete state space mechanism, and the corresponding first sequence encoded feature includes continuous input features and discrete input features.

[0076] In this application, for single-channel continuous input features, the continuous state space mechanism generates response features based on the observation of the hidden state and continuous input features, and this mechanism can be expressed as the basic formula of the selective state space model (State SpaceModel, SSM). SSM is a mathematical framework for modeling sequence data, and its first continuous formula in continuous form is usually expressed as:

[0077] h'(t) = Ah(t) + Bx(t);

[0078] y(t) = Ch(t);

[0079] where \(x(t)\in\mathbb{R}\) represents the single-channel continuous input feature; \(h(t)\in\mathbb{R}\) N represents the hidden state; \(h'(t)\) represents the derivative of the hidden state \(h(t)\), i.e., the rate of change of the state over time; \(A\in\mathbb{R}\) N×N represents the continuous state transition matrix, which describes how the state changes over time; \(B\in\mathbb{R}\) N×1 represents the first continuous projection matrix, indicating the influence of the single-channel continuous input feature \(x(t)\) on the hidden state \(h(t)\); \(y(t)\in\mathbb{R}\) represents the continuous response feature; \(C\in\mathbb{R}\) 1×N represents the second continuous projection matrix, indicating the influence of the hidden state \(h(t)\) on the continuous response feature \(y(t)\).

[0080] It should be noted that the hidden state \(h(t)\) can also be understood as the current state quantity, which represents the potential state of the system at any given time step \(t\). \(N\) represents the dimension of the hidden state, that is, the hidden state \(h(t)\) is an \(N -\)dimensional vector, and its elements may represent different information in different application scenarios. In wind speed prediction, it may be related to the feature representation of different meteorological variables.

[0081] where, when the continuous input feature and the continuous response feature are composed of \(V\) channels, i.e., \(x(t)\in\mathbb{R}\) V and \(y(t)\in\mathbb{R}\) V , the selective state space model (SSM) is independently applied to each channel. This means that \(A\in\mathbb{R}\) V×N×N , \(B\in\mathbb{R}\) V×N , \(C\in\mathbb{R}\) V×N . To optimize memory usage, \(A\) can be compressed to a size of \(V\times N\). Thereafter, unless otherwise specified, this application will specifically focus on multi-channel systems and the compressed form of \(A\).

[0082] However, for the discrete state space mechanism, in practical applications, especially when dealing with discrete time steps, it is necessary to discretize the above first continuous formula to adapt to data processing and analysis at discrete time points.

[0083] Furthermore, through a certain discretization method (such as the zero - order hold method ZOH), using the continuous state transition matrix \(A\) and the first continuous projection matrix \(B\), the discrete state transition matrix and the first discrete projection matrix are derived. Among them, for the discretized SSM formula, that is, the first discrete formula is usually expressed as:

[0084]

[0085] y t \(=\) \(Ch\) t ;

[0086] where,[[]] Represents the discrete state transition matrix, which is used to describe the transition relationship of states at discrete time points, reflecting the conversion from continuous to discrete, and preparing for the subsequent state update under discrete time; Represents the first discrete projection matrix, which reflects the change in the influence mode of the input on the state after discretization at discrete time points; exp represents matrix exponential operation; Δ represents the sampling time interval, Δ ∈ R V , and ΔA is invertible; I represents the identity matrix; h t-1 Represents the hidden state at the previous discrete time step (t - 1); h t and x t Represent the hidden state and discrete input feature at discrete time step t respectively; y t Represents the discrete response feature at discrete time step t; C ∈ R 1×N Represents the second continuous projection matrix.

[0087] Furthermore, the above discrete response operation can be efficiently computed using global convolution, expressed as:

[0088]

[0089] where L represents the length of the sequence, x represents the discrete input feature, and y represents the discrete response feature.

[0090] It should be noted that in the in-station Mamba module 42, this discretized form of SSM is used to construct the core part of the model, generating responses by observing the hidden state and input features. The in-station Mamba module 42 achieves better results in long sequence modeling and reduces computational complexity by optimizing the calculation method and parameters of SSM. At the same time, the in-station Mamba module 42 also combines other technologies (such as selective SSM calculation and linear transformation, etc.) to further improve the performance and efficiency of the module.

[0091] It should be pointed out that the dimensions of the continuous state transition matrix A, the first continuous projection matrix B, and the sampling time interval Δ will change with the input sequence length L, the number of channels V, etc., and these parameters will be dynamically adjusted according to the input features to capture the context information of the input features, selectively perform state transitions, so as to better adapt to different input patterns and better capture the complex features and dependencies of the wind speed time series in wind speed prediction.

[0092] In this application, the in-station Mamba module 42 processes the encoded features 30 of the first sequence of the input block by block, which can capture the complex dynamic features and dependencies within a single site time series and learn the patterns of the time series within a local range.

[0093] Step S42: Perform inter-station channel attention processing on the time series embedding features to obtain inter-station channel attention features.

[0094] As Figure 6 shown, the inter-station attention module 43 receives the output of the intra-station Mamba module 42. Using the inter-station channel attention mechanism, based on the output of the intra-station Mamba module 42, it extracts the features of each channel through a pooling operation, and then through linear transformation and activation function processing, calculates the attention weights between different stations (channels), captures the interdependencies between different stations in the multivariate time series, synthesizes the information of each station, and improves the understanding and prediction ability of the overall time series.

[0095] In this application, for the time series embedding features obtained after passing through the k-th intra-station Mamba module 42, the process of performing inter-station channel attention operation on the corresponding k-th inter-station attention module 43 to obtain intermediate features can be expressed as:

[0096] Attl(H k ) = sigmoid(MLP(MaxPool(H k )) + MLP(AvgPool(H k )));

[0097] Where, H k represents the time series embedding features output by the k-th intra-station Mamba module 42, and H k ∈R V×N×D ; Sigmoid represents the Sigmoid activation function; MLP represents the operation corresponding to the multi-layer perceptron MLP; MaxPool represents max pooling; AvgPool represents average pooling.

[0098] It should be noted that applying max pooling MaxPool and average pooling AvgPool to the last two dimensions can generate the maximum descriptor and average descriptor reflecting the overall features of each channel. At the same time, the multi-layer perceptron MLP performs linear transformation and non-linear activation (such as the Gelu function) on the input data through trainable weights, and then extracts features and captures the complex relationships in the data. In the inter-station attention module 42, the trainable weights act on the maximum descriptor and average descriptor generated by max pooling MaxPool and average pooling AvgPool, perform transformation processing on them, thereby affecting the calculation result of channel attention, and can help the network better capture the dependencies between different channels and improve the accuracy of wind speed prediction. Therefore, the process of the above-mentioned k-th inter-station attention module 43 performing inter-station channel attention operation to obtain intermediate features can be further expressed as:

[0099]

[0100] Wherein, W1 and W0 respectively represent the first trainable weight and the second trainable weight; and respectively represent the maximum descriptor and the average descriptor; Gelu represents the Gelu activation function.

[0101] Further, the intermediate feature Attl(H k ) obtained by the k-th inter-station attention module 43 and the time series embedding feature H k output by the k-th in-station Mamba module 42 are multiplied in matrix to obtain the inter-station channel attention feature output by the k-th inter-station attention module 43 as:

[0102] CA k = Attl(H k ) ⊙ H k ;

[0103] Wherein, CA k represents the inter-station channel attention feature output by the k-th inter-station attention module 43; ⊙ represents matrix multiplication.

[0104] Step S43: Superimpose the inter-station channel attention feature and the first sequence encoding feature to obtain the second sequence encoding feature.

[0105] As Figure 6 and Figure 7 shown, after obtaining the inter-station channel attention feature 50 output by the first inter-station attention module 43, superimpose it with the first sequence encoding feature 30 to obtain the second sequence encoding feature. Then use the second sequence encoding feature to enter the next iteration.

[0106] Step S44: Replace the first sequence encoding feature with the second sequence encoding feature, return to the first step, and repeat steps S41 to S43 for the first time. By analogy, when repeating for the k-th time, k = 2, 3,..., correspondingly replace the k-th sequence encoding feature with the (k + 1)-th sequence encoding feature and return to step S41 until the target sequence encoding feature is obtained.

[0107] As Figures 5 - 7 shown, the first sequence encoding feature 30 is input into the spatio-temporal attention encoding module 4 composed of K I2A-Mamba modules 41 for spatio-temporal attention encoding. The specific operations in the spatio-temporal attention encoding module 4 can be expressed as:

[0108] H k = PatchMamba(Z k-1 );

[0109] Z k = Attl(H k)☉H k +Z k-1 ;

[0110] where PatchMamba represents a per-block selective state space processing operation, H k represents the response feature output by the k-th in-station Mamba module 42, that is, the time series embedding feature input to the k-th inter-station attention module 43; Z k-1 represents the input feature input to the (k - 1)-th in-station Mamba module 42 or the first sequence encoding feature 30; Z k represents the inter-station channel attention feature output by the k-th inter-station attention module 43; Attl represents the inter-station channel attention operation; ⊙ represents matrix multiplication; k = 1, 2,..., K, and when k = K, Z K represents the target sequence encoding feature.

[0111] In this application, by implementing the in-station Mamba module 42 at the patch level, the temporal dependencies between different time patches can be captured. For the inter-station dependencies, this application proposes a lightweight inter-station channel attention mechanism to model various cross-channel relationships, including weighted summation and proportional relationships. The core idea of the I2A-Mamba module 41 is to capture the temporal dependence and spatial correlation in wind speed data by combining the in-station and inter-station attention mechanisms. Therefore, through the combined action of the in-station Mamba module 42 and the inter-station attention module 43, the wind speed prediction network I2A-Mamba can effectively handle the wind speed prediction task in multiple natural scenarios.

[0112] Step S5: Perform linear prediction on the target sequence encoding feature to obtain the predicted wind speed value time series.

[0113] Combined with Figure 5 , the linear prediction module 5 includes an RMS normalization activation layer (RMS Norm&Silu), a flattening layer (Flatten), a linear projection layer (Linear Projection), and a multivariate prediction layer (Multivariate Prediction).

[0114] In this embodiment, the linear prediction module 5 receives the target sequence encoding feature output from the spatio-temporal attention encoding module 4, and performs RMS normalization operation, Silu activation function activation, and flattening operation on the target sequence encoding feature in sequence by the RMS normalization activation layer (RMS Norm&Silu), the flattening layer (Flatten), and the linear projection layer (Linear Projection), and performs linear projection using the projection weight matrix to obtain the corresponding multivariate prediction time series. This process can be expressed as:

[0115]

[0116] Among them, Z K represents the target sequence coding feature; represents the multivariate prediction time series; W proj represents the projection weight matrix; Flatten represents the flattening operation; Silu represents the Silu activation function; RMS represents the RMS normalization operation.

[0117] In this embodiment, if the goal is to predict the wind speed values at T future time steps, then W proj ∈R (N *D)×T . Among them, x L+1 to x L+T represent the predicted values at T different time steps, and x l' ∈R n×V , l' ∈ (L + 1,..., L + T), that is, the predicted value at time step l' is an n×V-dimensional vector, corresponding to V kinds of meteorological data respectively corresponding to the locations of n meteorological collection stations in a scenario containing n meteorological collection stations.

[0118] Furthermore, the Multivariate Prediction layer performs wind speed prediction based on the multivariate prediction time series to obtain a predicted wind speed value sequence, and the predicted wind speed value sequence includes the predicted wind speed values at T future time steps respectively corresponding to the locations of n meteorological collection stations.

[0119] Among them, for a meteorological collection station, wind speed prediction is performed based on the V kinds of meteorological data corresponding to the T future time steps respectively, and then the T predicted wind speed values respectively corresponding to the location of the meteorological collection station at the T future time steps can be inferred.

[0120] For example Figure 9As shown, the wind speed prediction network I2A-Mamba of the present invention and multiple other existing network models respectively perform wind speed prediction based on the historical wind speed data of the offshore scenario, and correspondingly generate predicted wind speed values. The first column shows the predicted wind speed values for 10 time steps generated by the wind speed prediction network I2A-Mamba; the second to fifth columns respectively show the predicted wind speed values for 10 time steps generated by the Autoformer, Informer, LogSparse, and Transformer network models; the last column shows the true value GT actually collected at 10 time steps, that is, the actual wind speed magnitude. Generally speaking, the predicted wind speed values obtained by the wind speed prediction network I2A-Mamba of the present invention are closer to the actual wind speed magnitude, and the prediction performance of the wind speed prediction network I2A-Mamba is more stable.

[0121] Based on the same inventive concept, as Figure 10 shown, the present invention also provides a wind speed prediction system, the system includes:

[0122] A preprocessing unit 101, configured to preprocess the historical wind speed data to obtain a historical time series.

[0123] A chunking unit 102, configured to chunk the historical time series to obtain a multivariate time series chunk.

[0124] A feature transformation unit 103, configured to perform feature transformation on the multivariate time series chunk to obtain a first sequence encoded feature.

[0125] A spatio-temporal attention encoding unit 104, configured to perform spatio-temporal attention encoding on the first sequence encoded feature to obtain a target sequence encoded feature.

[0126] A linear prediction unit 105, configured to perform linear prediction on the target sequence encoded feature to obtain a predicted wind speed value time series.

[0127] In this application, the other technical features in the above wind speed prediction system are the same as the features disclosed in the above method embodiment, and will not be elaborated here.

[0128] Based on the same inventive concept, the present application also provides an electronic device, the electronic device includes a processor, a memory, and a communication circuit, and the processor is respectively connected to the memory and the communication circuit; wherein, the communication circuit is used for communication connection, the memory is used to store a computer program, and the processor is used to execute the computer program to implement the above method.

[0129] Please refer to Figure 11 , the electronic device described in the embodiment of the present application may specifically include a processor 210 and a memory 220. The memory 220 is coupled to the processor 210.

[0130] The processor 210 is used to control the operation of the electronic device. The processor 210 may also be referred to as a CPU (Central Processing Unit). The processor 210 may be an integrated circuit chip with the ability to process signals. The processor 210 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The general-purpose processor may be a microprocessor or the processor 210 may also be any conventional processor, etc.

[0131] The memory 220 is used to store computer programs and may be a RAM, a ROM, or other types of storage terminals. Specifically, the memory 220 may include one or more computer-readable storage media, which may be non-transitory or transitory. The memory 220 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage terminals, flash storage terminals. In some embodiments, the non-transitory computer-readable storage media in the memory 220 is used to store at least one program code.

[0132] The processor 210 is used to execute the computer programs stored in the memory 220 to implement the methods described in the method embodiments of the present application.

[0133] In some embodiments, the electronic device may further include: a peripheral terminal interface 230 and at least one peripheral terminal. The processor 210, the memory 220, and the peripheral terminal interface 230 may be connected through a bus or signal lines. Each peripheral terminal may be connected to the peripheral terminal interface 230 through a bus, signal lines, or a circuit board. Specifically, the peripheral terminal includes at least one of a radio frequency circuit 240, a display screen 250, an audio circuit 260, and a power supply 270.

[0134] The peripheral terminal interface 230 may be used to connect at least one I / O (Input / Output) related peripheral terminal to the processor 210 and the memory 220. In some embodiments, the processor 210, the memory 220, and the peripheral terminal interface 230 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 210, the memory 220, and the peripheral terminal interface 230 may be implemented on a separate chip or circuit board, and this embodiment does not limit this.

[0135] The radio frequency circuit 240 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The radio frequency circuit 240 communicates with the communication network and other Internet of Things devices through electromagnetic signals, and the radio frequency circuit 240 is the communication circuit of the electronic device. The radio frequency circuit 240 converts electrical signals into electromagnetic signals for transmission, or converts the received electromagnetic signals into electrical signals. Optionally, the radio frequency circuit 240 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a subscriber identity module card, and so on. The radio frequency circuit 240 can communicate with other terminals through at least one wireless communication protocol. The wireless communication protocol includes but is not limited to: the World Wide Web, metropolitan area network, intranet, generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area network, and / or WiFi (Wireless Fidelity) network. In some embodiments, the radio frequency circuit 240 may further include a circuit related to NFC (Near Field Communication), which is not limited in this application.

[0136] The display screen 250 is used to display the UI (User Interface). The UI may include graphics, text, icons, videos, and any combination thereof. When the display screen 250 is a touch display screen, the display screen 250 also has the ability to collect touch signals on or above the surface of the display screen 250. The touch signals can be input to the processor 210 for processing as control signals. At this time, the display screen 250 can also be used to provide virtual buttons and / or virtual keyboards, also known as soft buttons and / or soft keyboards. In some embodiments, there may be one display screen 250, which is provided on the front panel of the electronic device; in other embodiments, there may be at least two display screens 250, which are respectively provided on different surfaces of the electronic device or in a foldable design; in other embodiments, the display screen 250 may be a flexible display screen, which is provided on the curved surface or folding surface of the electronic device. Even, the display screen 250 can be set to an irregular non-rectangular shape, that is, a special-shaped screen. The display screen 250 can be prepared using materials such as LCD (Liquid Crystal Display) and OLED (Organic Light-Emitting Diode).

[0137] The audio circuit 260 may include a microphone and a speaker. The microphone is used to collect sound waves of the operator and the environment, and convert the sound waves into electrical signals for input to the processor 210 for processing, or input to the radio frequency circuit 240 to achieve voice communication. For the purpose of stereo collection or noise reduction, there may be multiple microphones, which are respectively arranged at different parts of the electronic device. The microphone may also be an array microphone or an omnidirectional collection microphone. The speaker is used to convert the electrical signal from the processor 210 or the radio frequency circuit 240 into sound waves. The speaker may be a traditional thin film speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can not only convert the electrical signal into sound waves audible to humans, but also convert the electrical signal into sound waves inaudible to humans for uses such as ranging. In some embodiments, the audio circuit 260 may further include a headphone jack.

[0138] The power supply 270 is used to supply power to each component in the electronic device. The power supply 270 may be alternating current, direct current, a primary battery or a rechargeable battery. When the power supply 270 includes a rechargeable battery, the rechargeable battery may be a wired rechargeable battery or a wireless rechargeable battery. A wired rechargeable battery is a battery charged through a wired line, and a wireless rechargeable battery is a battery charged through a wireless coil. The rechargeable battery may also be used to support fast charging technology.

[0139] For a detailed description of the functions and execution processes of each functional module or component in the electronic device embodiments of this application, reference may be made to the descriptions in the method embodiments of this application above, and details are not repeated here.

[0140] In several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the electronic device embodiments described above are merely illustrative. For example, the division of modules or units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of devices or units can be in electrical, mechanical or other forms.

[0141] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0142] In addition, in each embodiment of the present application, each functional unit can be integrated into a processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0143] Based on the same inventive concept, the present application also provides a computer-readable storage medium storing a computer program that can be executed by a processor to implement the above method.

[0144] Please refer to Figure 12 , when the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in the computer-readable storage medium 300. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions / computer programs to enable an Internet of Things device (which can be a personal computer, a server, or a network terminal, etc.) or a processor to execute all or part of the steps of the methods in various embodiments of the present application. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs, as well as electronic terminals such as computers, mobile phones, laptop computers, tablet computers, cameras, etc. having the above storage media.

[0145] The description of the execution process of the program data in the computer-readable storage medium can refer to the description in the above method embodiments of the present application, and will not be repeated here.

[0146] It can be seen that the present invention discloses a wind speed prediction method, system, device, and medium. The method includes the following steps: preprocessing historical wind speed data to obtain a historical time series; partitioning the historical time series to obtain a multivariate time series block; performing feature transformation on the multivariate time series block to obtain a first sequence coding feature; performing spatio-temporal attention coding on the first sequence coding feature to obtain a target sequence coding feature; and performing linear prediction on the target sequence coding feature to obtain a predicted wind speed value time series. This method can effectively capture the time-dependent relationship of each meteorological collection station, significantly improve the prediction accuracy; can also model the correlation between different meteorological collection stations, further improving the prediction performance; at the same time, it can be widely applied to fields such as wind energy production, weather forecasting, and climate research, helping to improve the safety and reliability of the wind energy system.

[0147] The above are only embodiments of the present invention, and thus do not limit the protection scope of the present invention. Any equivalent structural transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, is included in the protection scope of the present invention.

Claims

1. A wind speed prediction method, characterized in that: The method comprises the following steps: Preprocess the historical wind speed data to obtain the historical time series; Dividing the historical time series into blocks to obtain multivariate time series blocks; Performing feature conversion on the multivariate time series block to obtain a first sequence coding feature; Performing spatiotemporal attention encoding on the first sequence encoding features to obtain target sequence encoding features; Linear prediction is performed on the target sequence encoding features to obtain a predicted wind speed value time series.

2. The wind speed prediction method according to claim 1, characterized in that: The method further comprises: A wind speed data set is constructed, wherein the wind speed data set includes the historical wind speed data from a plurality of different scenarios.

3. The wind speed prediction method according to claim 1, characterized in that: The step of dividing the historical time series into blocks to obtain multivariate time series blocks includes: Using a sliding window, the historical time series is divided into multiple multivariate time series blocks with the same length and step size.

4. The wind speed prediction method according to claim 1, characterized in that: The step of performing spatiotemporal attention encoding on the first sequence encoding feature to obtain a target sequence encoding feature includes: The first step is to perform selective state space processing on the first sequence encoding feature to obtain a time series embedding feature; The second step is to perform inter-station channel attention processing on the time series embedding features to obtain inter-station channel attention features; The third step is to superimpose the inter-station channel attention feature with the first sequence coding feature to obtain the second sequence coding feature; The fourth step is to replace the first sequence coding feature with the second sequence coding feature, return to the first step, repeat the first to third steps for the first time, and so on. When repeating for the kth time, k=2,3,..., correspondingly use the (k+1)th sequence coding feature to replace the kth sequence coding feature and return to the first step until the target sequence coding feature is obtained.

5. The wind speed prediction method according to claim 4, characterized in that: When the first sequence encoding feature is a continuous input feature, the selective state space processing includes continuous state space processing; and performing the selective state space processing on the first sequence encoding feature to obtain the time series embedding feature includes: The continuous input feature is processed in the continuous state space by using a first continuous formula to obtain a continuous response feature; wherein the first continuous formula is: h'(t)=Ah(t)+Bx(t); y(t) = Ch(t); Among them, x(t) represents the continuous input feature; h(t) represents the hidden state; h'(t) represents the derivative of the hidden state h(t); A represents the continuous state transfer matrix; B represents the first continuous projection matrix; y(t) represents the continuous response feature; C represents the second continuous projection matrix.

6. The wind speed prediction method according to claim 4, characterized in that: When the first sequence coding feature is a discrete input feature, the selective state space processing includes discrete state space processing; and performing the selective state space processing on the first sequence coding feature to obtain the time series embedding feature includes: The discrete input feature is processed in the discrete state space by using a first discrete formula to obtain a discrete response feature; wherein the first discrete formula is: the t =Ch t ; Wherein, A represents the continuous state transfer matrix; B represents the first continuous projection matrix; represents the discrete state transfer matrix; represents the first discrete projection matrix; exp represents the matrix exponential operation; Δ represents the sampling time interval, and ΔA is reversible; I represents the identity matrix; h t-1 represents the hidden state at the last discrete time step (t-1); h t and x t Respectively represent the hidden state at discrete time step t and the discrete input feature; y t represents the discrete response feature at discrete time step t; C represents the second continuous projection matrix.

7. The wind speed prediction method according to claim 4, characterized in that: The performing inter-station channel attention processing on the time series embedding feature to obtain the inter-station channel attention feature includes: Performing an inter-station channel attention operation on the time series embedding features to obtain intermediate features; The intermediate features and the time series embedding features are matrix multiplied to obtain the inter-station channel attention features.

8. A wind speed prediction system, characterized in that: The system comprises: A preprocessing unit, used for preprocessing the historical wind speed data to obtain a discrete time series; A block division unit, used for dividing the historical time series into blocks to obtain multivariate time series blocks; A feature conversion unit, used for performing feature conversion on the multivariate time series block to obtain a first sequence coding feature; a spatiotemporal attention encoding unit, configured to perform spatiotemporal attention encoding on the first sequence encoding feature to obtain a target sequence encoding feature; The linear prediction unit is used to perform linear prediction on the target sequence encoding features to obtain a predicted wind speed value time series.

9. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images