Method for predicting wind speed along high-speed rail based on intermittent prior and cross-scale adaptive fusion

Abstract

The invention provides a method for predicting wind speed along a high-speed rail based on intermittent prior and cross-scale adaptive fusion, which comprises the following steps of: 1, collecting and aligning second-level wind speed and minute-level mean value of monitoring stations along the rail, and organizing as model input according to a sliding window; 2, constructing an intermittent prior module; step 3, extracting second-level features by using a second-level encoder, and converging time-varying features of intermittence and fluctuation through intermittence and fluctuation double-path pooling to form second-level convergence representation; 4, encoding the minute-level historical mean value by adopting a minute-level encoder to obtain minute trend representation; and step 5, adaptively fusing the minute trend and the second-level convergence representation through a fusion layer containing intermittent guidance cross-scale attention, forming a fusion representation, sending the fusion representation to a prediction head, outputting a prediction result, and performing evaluation. According to the method, the accuracy of the wind speed prediction value can be effectively improved, and the reliability of the wind speed prediction value is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of time series forecasting, and in particular to a method for predicting wind speed along high-speed railway lines based on intermittent priors and cross-scale adaptive fusion. Background Technology

[0002] Currently, high-speed rail disaster prevention systems have real-time wind monitoring and alarm functions, but lack early warning capabilities, resulting in insufficient disaster warning capabilities. If the system could predict wind speed values ​​over a future period based on current wind speed data, it would help issue timely warnings and provide a scientific basis for train scheduling and safety procedures, thereby improving the safety, reliability, and resilience of the high-speed rail system.

[0003] Although scholars both domestically and internationally have conducted extensive research on wind speed prediction, there is currently no research specifically addressing second-level granularity prediction along railway lines. Furthermore, the complex terrain and frequent strong winds along high-speed rail routes make it difficult to capture these characteristics. Therefore, the multi-step prediction lag problem in algorithms is an inevitable challenge in multi-step time series prediction of wind speed along railway lines, and solving this problem has become a bottleneck. Most current research utilizes minute-averaged sequences derived from second-level wind speed measurements for wind speed prediction. While this averaging process effectively reduces short-term fluctuations and noise, making the input data more stable, it also eliminates the fine-grained temporal characteristics inherent in second-level measurements—characteristics crucial for capturing rapid wind dynamics and ensuring high prediction accuracy.

[0004] Therefore, how to fully explore the temporal structure characteristics and fluctuation patterns of wind speed signals, develop high-precision prediction methods for wind speed along high-speed railways, and solve core technical problems such as multi-step prediction lag and mismatch between training samples and samples to be predicted has become a key issue that urgently needs to be addressed to improve the disaster prevention and early warning capabilities of high-speed railways and ensure the safety of train operation. Summary of the Invention

[0005] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by proposing a high-speed rail wind speed prediction method based on intermittent prior and cross-scale adaptive fusion. This method explicitly quantifies and distinguishes the temporal structure of "intermittent and fluctuating" elements without using future information. It combines multi-scale temporal convolution with prior-guided channel attention, dual-channel pooling, and cross-scale attention to achieve adaptive alignment and fusion of minute-level trends and second-level features, outputting the final predicted value, thus effectively improving the accuracy of wind speed predictions.

[0006] The method of the present invention includes the following steps: Step 1: Collect and align the second-level and minute-level average wind speeds from monitoring stations along the route, and organize them as model inputs using a sliding window. Step 2: Construct the Intermittency Prior Module (IPM): Under causal constraints, calculate the differential intensity, high-frequency energy, and causal normalization deviation of the second-level wind speed, and use a small multilayer perceptron network to adaptively generate an intermittent weight map and a soft play sequence to quantify the intermittency and duration that remain unchanged in continuous time steps. Step 3: Second-level encoder is used to extract second-level features. The second-level encoder uses a multi-scale temporal convolutional network (MSTCN) to recalibrate the channel weights under intermittency-guided Squeeze-and-Excitation (IG-SE). The time-varying features of intermittency and fluctuation are aggregated by intermittency and fluctuation time-varying features respectively through intermittency and fluctuation dual-path pooling (DP) to form a second-level aggregated representation. Step 4: Encode the historical mean at the minute level using a minute-level encoder to obtain a minute-level trend representation; the minute-level encoder includes a convolutional neural network and a bidirectional long short-term memory network CNN-BILSTM (Convolutional Neural Network + Bidirectional Long Short-Term Memory). Step 5: Adaptively fuse minute-level trends and second-level convergence representations through a fusion layer with intermittently guided cross-scale attention to form a fused representation, which is then fed into the prediction head. The prediction results are output and evaluated to verify the feasibility of the model.

[0007] Step 1 includes the following steps: Step 1-1, set up a second-level wind speed observation sequence ,in This represents the wind speed observation value at second t, where t ranges from 1 to n, and n represents the total number of wind speed observation points. The sampling interval is 1 second, and the time index is monotonically increasing. Steps 1-2 involve calculating statistics for a fixed duration. Starting from the c-th wind speed observation point, a subsequence of length w is taken. The specific formula is as follows: , in, Let c represent a sliding window that satisfies Preferred Seconds correspond to 3 minutes; Steps 1-3 calculate the three-minute average within each window and align the timestamps to the end of the window to ensure that only historical information is used. The specific formula is as follows: , , in, This represents the 3-minute moving average starting from the c-th wind speed observation point, which serves as the input to the minute-level encoder. The mean Corresponding time tags are used to avoid information leakage caused by using arbitrary future samples; The time label represents the last observation point of the c-th sliding window; Steps 1-4, in the same window As the input segment for a second-level encoder, it is denoted as: , in, This represents the c-th sliding window.

[0008] Step 2 includes the following steps: Step 2-1: To measure the magnitude of change between adjacent time points and reduce the impact of instantaneous outliers on subsequent discrimination, calculate the first-order difference and robust amplitude: , , in, It is a first-order difference. For robust amplitude; Smoothing constant ( ); Step 2-2: Without using future information, estimate the short-term strength of local movement intensity and relative trend using the Exponential Weighted Moving Average (EMA). The specific formula is as follows: , , , The exponentially weighted moving average (EMA) is defined as follows: , in, For smoothing coefficients; Let be the local motion intensity at time t; It serves as a trend baseline, used to characterize the slowly varying components of wind speed; This represents the strength of short-term changes in relative trends; An exponentially weighted moving average operator for calculating local motion intensity; An exponentially weighted moving average operator for calculating a short-term trend baseline; An exponentially weighted moving average operator for calculating the strength of short-term changes in relative trends; For exponentially weighted moving average operators; Steps 2-3 involve calculating the mean and variance using the exponentially weighted moving average (EMA) based on the historical window, and obtaining the standardized deviation to measure the significance of the current wind speed relative to historical levels. The specific formula is as follows: , , , in, The smoothing coefficient for the causal mean and variance; This is the exponentially weighted moving average at second t; For causal variance estimation; Standardized deviation; An exponentially weighted moving average operator for calculating the arithmetic mean of a window; The wind speed observation value is at the (t-1)th second; It is a stable term; Steps 2-4 involve concatenating the causal features according to time steps to form a prior feature vector, and then incorporating the intermittent prior from the previous step to enhance temporal coherence. The specific formula is as follows: , in, For the prior feature vector, This is the intermittent weight from the previous step (optional). For the previous soft run (optional); Steps 2-5 involve using a small multilayer perceptron (MLP) to perform nonlinear mapping on the prior features, outputting intermittent discriminative features, and controlling the temperature parameter (soft threshold parameter) for thresholding hardness and softness. The specific formula is as follows: , , in, This is a feature for discriminating between intervals (the larger the value, the more likely it is to be intermittent); These are unconstrained temperature parameters; This is the soft threshold parameter after adding a lower bound; This represents the minimum value of the soft threshold parameter. , is a positive-valued smoothing activation function used to ensure that the unconstrained temperature parameter is positive; LN is layer normalization; Steps 2-6: Use a Sigmoid soft gate to map the intermittent discriminative features to the soft threshold parameter. The continuous weights of the intervals are used to form an intermittent weighted graph, and the specific formula is as follows: , in, It is a non-linear activation function; This represents the intermittent weight; a larger value indicates a closer approximation to "intermittent (remaining unchanged for multiple steps)," while a smaller value indicates a closer approximation to "fluctuation." Steps 2-7 use differentiable recursion to estimate the duration of the current intermittent state, improving the stability of the segment boundaries. The specific formula is as follows: , Among them, soft runs The initial value is 0. This is the discharge coefficient; This is an estimate of the duration of the intermittent state up to t-1 seconds.

[0009] Step 3 includes the following steps: Step 3-1: The multi-scale temporal convolutional network MSTCN is used to extract features from a second-level input of length w, so as to simultaneously cover short, medium, and long receptive fields. The specific process includes: Step 3-1-1: Let the current second-level window input be S, construct B parallel one-dimensional convolutional branches, and the b-th branch uses a convolutional kernel length of... With expansion coefficient To form different receptive fields; when The window is filled with the initial value or zero, and future samples are not used. The specific formula is as follows: , in, ; For branch channel index; For the b-th branch at time t, the channel The output value; For the b-th branch, the The convolution weights of each output channel at a lag of j steps; This indicates the number of lag steps skipped based on the hole rate; Looking back from the current time t Input of one lag step; For the b-th branch, the The bias term for each output channel; Step 3-1-2: Concatenate the outputs of each branch along the channel dimension, and compress and shape the concatenated features to obtain unified temporal features. The specific formula is as follows: , in, This is the output feature map of the b-th branch; The channel-dimensional splicing operator splices the output feature maps of B branches; is a channel-mixed convolution; H is the intermediate feature map after channel mixing; Step 3-1-3: To enhance gradient propagation and numerical stability, residual projection is performed on the input S to match its channel number with H. This residual projection is then added to H and normalized at the layer level. The specific formula is as follows: , in This is a projection of the one-dimensional convolutional residual onto the input S; This is the final output feature map; Step 3-2, under the intermittently guided channel attention mechanism, The channels are recalibrated so that the model learns different channel importance for intermittent and fluctuating segments. This includes the following steps: Step 3-2-1: Calculate the normalized coefficients for intermittent and fluctuating coverage, which will be used as the denominator to avoid zero values. The formula is: , , in ; This is the normalized denominator for intermittent coverage; This is the normalized denominator for the volatility coverage. Step 3-2-2 involves performing a weighted compression (squeeze) on each channel from three perspectives: overall, intermittent, and fluctuation, to extract the channel's statistical features. The formula is as follows: , , , in, For the overall response of the channel, Emphasizing intermittent segments, Emphasizing fluctuation segments; The output feature map of channel c at time t; Step 3-2-3: The three types of statistics are concatenated with the global statistics of the intermittent prior, and then fed into a small multilayer perceptron (MLP) to generate channel weights for channel importance estimation. The formula is: , , , in, For channel weights, , and These are the mean and variance of the intermittent weights, respectively, used as prior descriptions of the intermittent proportion and stability of the current window; the output of the small multilayer perceptron (MLP) is compressed into channel weights using a sigmoid function. ; Step 3-2-4 involves multiplicative recalibration of the feature maps, enabling the fusion and prediction layers to learn different channel importances for intermittent and fluctuating segments. The formula is as follows: , in, The feature map after channel recalibration; Step 3-3: Intermittent and fluctuating dual-channel pooling is used to aggregate the time-varying features of intermittent and fluctuating data respectively, and a second-level aggregated representation is formed for subsequent fusion and prediction.

[0010] In step 3-3, the following calculation formula is used: , , , in, Let C be the channel vector at time step t; C is the number of channels. To emphasize the weighted average of intermittent segments; To emphasize the weighted average of fluctuating segments; For second-level side output, Indicates length is A real vector.

[0011] Step 4 includes the following steps: Step 4-1: Construct a historical window of three-minute averages while aligning the ends of the window, and perform causal standardization to provide a stable input sequence. The specific formula is as follows: , , , , in, For This is the historical window at the end, with a window length of K; This is the (k-K+1)th three-minute average; Based on only Calculate the mean and standard deviation; This is the standardized three-minute average; Step 4-2: Extract local temporal patterns and multi-scale neighborhood information from the window using one-dimensional convolution, and concatenate them in the channel dimension to obtain the convolutional representation. The specific formula is as follows: , , , in, , , This is the convolution output of the r-th branch at time step i; The channel vector is the result of concatenating multiple branches; U represents the convolutional sequence. is the value of the three-minute mean after causal standardization at index iq; q is the position index within the convolution kernel; Let be the convolution weight of the r-th branch at kernel position q; Let be the bias of the r-th branch; Step 4-3: Feed the convolutional representation into a bidirectional long short-term memory network to model the long short-term dependencies within the window and obtain the bidirectional hidden state sequence. The specific formula is as follows: , , in, These are the forward hidden state and backward hidden state of BiLSTM, respectively. It is a bidirectional long short-term memory network; To splice the matrix; Step 4-4: Attention weighting and statistical pooling are used to aggregate the hidden state sequence into fixed-length minute sequence features for subsequent fusion and prediction. The calculation formula is as follows: , , in, For attention weights, To measure sequence features in minutes, T represents transpose. is the hyperbolic tangent activation function, used to compress the linearly mapped vector to the range (-1, 1); e is the learnable weight vector for attention scoring.

[0012] Step 5 includes the following steps: Step 5-1: Align the minute-level minute sequence features with the second-level second-level converged representation to the same feature space, and adaptively combine them using gating. The specific formula is as follows: , , , , in, The convergence is represented in seconds, and y is the fusion vector; , Learnable parameters for the minute-side linear mapping; , Learnable parameters for second-level linear mapping; , Learnable parameters for the gating distributor; The aligned second-level side features; For the aligned minute-side features; This is element-wise multiplication; g is the gated weight vector; Step 5-2, Cross-scale attention to intermittent leaders: Cross-scale retrieval is performed using minute-level lateral vectors as queries and second-level time series features as keys and values. A prior bias generated by the intermittent weight map is introduced to highlight fluctuating segments. This specifically includes the following steps: Step 5-2-1: Linearly map the minute-level features and second-level features to the shared attention space to form the query vector Q and the key vector. AND value vector The specific formula is as follows: , , , in , It is a linear mapping matrix; Step 5-2-2: Generate prior biases for suppressing intermittent segments and emphasizing fluctuating segments based on the intermittent weight map. The specific formula is as follows: ; in, This refers to the prior bias term injected into the attention scoring at time t; Step 5-2-3: Calculate the scaled dot product score and superimpose the prior, causal, and length masks to obtain constrained attention. The specific formula is as follows: , in, d is the length mask; d is the dimension of the attention space; Score the attention at time t; Step 5-2-4: Normalize the scores into attention weights and then weight and converge the value vectors to obtain a cross-scale weight vector. The specific formula is as follows: , , in, Here, z represents the attention weights, z is the cross-scale vector, and exp is the natural exponential function. Step 5-3, Fusion Output and Prediction Head: Deeply combine vector-level fusion and attention retrieval to output a three-minute average wind speed prediction.

[0013] In step 5-3, the following calculation formula is used: , , in For lightweight return head, These are the parameters for fusion mapping; This is the next predicted value.

[0014] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0015] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0016] Beneficial Effects: The second-level wind speed prediction method along high-speed railways proposed in this invention (ICAF-Net, Intermittency-guided Cross-scale Adaptive Fusion Network) uses an Intermittency Prior Module (IPM) to explicitly quantify the "intermittency / fluctuation" structure and generate weights and soft runs; it uses a Multi-Scale Temporal Convolutional Network (MSTCN) to extract short / medium / long receptive field features; it introduces Intermittent Guided Channel Attention (IG-SE) to recalibrate channels; it uses intermittent / fluctuation dual-path pooling (DP) to converge the two types of dynamics respectively; it encodes minute-level trends with CNN-BiLSTM; and in the fusion stage, it uses cross-scale attention and gated fusion (CSA+GF) to adaptively align and focus on minute trends and second-level features. The above modules work together to effectively suppress platform noise, highlight key fluctuations, alleviate the lag and error accumulation of multi-step extrapolation, and significantly improve the accuracy and reliability of wind speed predictions. Among them, core modules such as IPM and CSA are particularly significant in improving accuracy. The stability of wind speed prediction results used for early warning and scheduling decisions along the line is greatly improved, meeting the online and low-latency application requirements of high-speed rail scenarios. Attached Figure Description

[0017] Figure 1 This is the overall flowchart of the present invention.

[0018] Figure 2 This is a structural diagram of the intermittent prior module and the small multilayer perceptron of the present invention.

[0019] Figure 3 This is a structural diagram of the Temporal Convolutional Network (TCN) of this invention.

[0020] Figure 4 This is a structural diagram of the second-level encoder MSTCN of the present invention.

[0021] Figure 5 This is a structural diagram of the Intermittent Weighted Channel Attention (IG-SE) of this invention.

[0022] Figure 6 This is a diagram of the intermittent / fluctuating dual-channel pool (DP) structure of the present invention.

[0023] Figure 7 This is a diagram of the CNN-BILSTM structure of the minute encoder of this invention.

[0024] Figure 8 This is a graph showing the prediction results of the present invention on the Dashengguan dataset.

[0025] Figure 9 This is a graph showing the prediction results of this invention on the Yangcheng Lake dataset.

[0026] Figure 10 This is a graph showing the prediction results of this invention on the Zouwanglu dataset. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0028] This invention proposes a method for predicting wind speed along high-speed railway lines based on intermittent priors and cross-scale adaptive fusion, such as... Figure 1 As shown, it includes the following steps: Step 1: Collect second-level wind speed sequences from monitoring stations along the route and align the timestamps with the stations. Under causal constraints, calculate the three-minute mean within each window using a causal sliding window of length w=180 seconds based on the second-level sequence, and align the mean to the end time point of the window. Then, organize the data within the window into model input samples, without using information from outside the window or future times.

[0029] Step 2: Construction of the Intermittent Prior Module (IPM). Within the current sliding window, causal local feature extraction is performed on the second-level wind speed sequence: prior quantities such as differential intensity, high-frequency energy, standardized deviation, and the length / proportion of approximately constant segments are calculated and aggregated, and concatenated in a fixed order to form a prior vector. This prior vector is then input into a small multilayer perceptron for adaptive mapping, resulting in a time-step-level intermittent weight map and a soft-play sequence. The structure of the intermittent prior module is as follows: Figure 2 As shown, the specific steps are as follows: Step 2-1: To measure the magnitude of change between adjacent time points and reduce the impact of instantaneous outliers on subsequent discrimination, calculate the first-order difference and robust amplitude. The specific formula is as follows: , , in, It is a first-order difference. For robust amplitude; Smoothing constant ( ); Step 2-2: Without using future information, estimate the short-term strength of local movement intensity and relative trend using the Exponential Weighted Moving Average (EMA). The specific formula is as follows: , , , The exponentially weighted moving average (EMA) is defined as follows: , in, For smoothing coefficients; Let be the local motion intensity at time t; It serves as a trend baseline, used to characterize the slowly varying components of wind speed; This represents the strength of short-term changes in relative trends; An exponentially weighted moving average operator for calculating local motion intensity; An exponentially weighted moving average operator for calculating a short-term trend baseline; An exponentially weighted moving average operator for calculating the strength of short-term changes in relative trends; For exponentially weighted moving average operators; Steps 2-3 involve calculating the mean and variance using the exponentially weighted moving average (EMA) based on the historical window, and obtaining the standardized deviation to measure the significance of the current wind speed relative to historical levels. The specific formula is as follows: , , , in, The smoothing coefficient for the causal mean and variance; This is the exponentially weighted moving average at second t; For causal variance estimation; Standardized deviation; An exponentially weighted moving average operator for calculating the arithmetic mean of a window; The wind speed observation value is at the (t-1)th second; It is a stable term; Steps 2-4 involve concatenating the causal features according to time steps to form a prior feature vector, and then incorporating the intermittent prior from the previous step to enhance temporal coherence. The specific formula is as follows: , in, For the prior feature vector, For the intermittent weights in the previous step, This is the previous soft journey; Steps 2-5 involve using a small multilayer perceptron (MLP) to perform nonlinear mapping on the prior features, outputting intermittent discriminative features, and controlling the temperature parameters for thresholding hardness and softness. The specific formula is as follows: , , in, For intermittent discrimination features; These are unconstrained temperature parameters; This is the soft threshold parameter after adding a lower bound; This represents the minimum value of the soft threshold parameter. , LN is a positive-valued smooth activation function; LN is a layer normalization function. Steps 2-6: Use a Sigmoid soft gate to map the intermittent discriminative features to the soft threshold parameter. The continuous weights of the intervals are used to form an intermittent weighted graph, and the specific formula is as follows: , in, It is a non-linear activation function; Intermittent weighting; Steps 2-7 use differentiable recursion to estimate the duration of the current intermittent state, improving the stability of the segment boundaries. The specific formula is as follows: , Among them, soft runs The initial value is 0. This is the discharge coefficient; This is an estimate of the duration of the intermittent state up to t-1 seconds.

[0030] Step 3: A multi-scale temporal convolutional network (MSTCN) is used to extract second-level features across short / medium / long receptive fields. Channels are recalibrated using intermittency-guided squeeze-and-excitation (IG-SE), and the two types of dynamics are pooled using dual-path pooling for intermittency / fluctuation (DP) to form an interpretable second-level converged representation. The specific steps are as follows: Step 3-1: A multi-scale temporal convolutional network (MSTCN) is used to extract features from a 180-second window of input to simultaneously cover short, medium, and long receptive fields, expanding the causal convolution as follows: Figure 3 As shown in (a), the TCN residual block is as follows Figure 3 As shown in (b), the MSTCN structure is as follows Figure 4 As shown.

[0031] Step 3-2, under the intermittently guided channel attention mechanism, The channels are recalibrated so that the model learns different channel importance for intermittent and fluctuating segments. The IG-SE structure is as follows: Figure 5 As shown, the specific steps are as follows: Step 3-2-1: Calculate the normalized coefficients for intermittent and fluctuating coverage. The formula is: , , in ; This is the normalized denominator for intermittent coverage; This is the normalized denominator for the volatility coverage. Step 3-2-2: Perform weighted compression on each channel from three perspectives: overall, intermittent, and fluctuation. The formula is as follows: , , , in, For the overall response of the channel, Emphasizing intermittent segments, Emphasizing fluctuation segments; The output feature map of channel c at time t; Step 3-2-3: The three types of statistics are concatenated with the global statistics of the intermittent prior, and then fed into a small multilayer perceptron (MLP) to generate channel weights for channel importance estimation. The formula is: , , , in, For channel weights, , and The mean and variance of the intermittent weights are respectively used. The output of the small multilayer perceptron (MLP) is compressed into channel weights using a sigmoid function. ; Step 3-2-4 involves multiplicative recalibration of the feature maps, enabling the fusion and prediction layers to learn different channel importances for intermittent and fluctuating segments. The formula is as follows: , in, The feature map after channel recalibration; Step 3-3 employs intermittent and fluctuating dual-channel pooling to aggregate the time-varying features of intermittent and fluctuating data respectively, forming a second-level aggregated representation for subsequent fusion and prediction. The DP module structure diagram is shown below. Figure 6 As shown, the specific formula is: , , , in, Let C be the channel vector at time step t; C is the number of channels. To emphasize the weighted average of intermittent segments; To emphasize the weighted average of fluctuating segments; For second-level side output, Indicates length is A real vector.

[0032] Step 4: A combination of convolutional neural network and bidirectional long short-term memory (CNN-BiLSTM) is used to encode the minute-level historical mean. Convolution captures local patterns and neighborhood smoothness, while BiLSTM models slowly changing trends and contextual dependencies to obtain a minute-level trend representation. The specific steps are as follows: Step 4-1: Construct a historical window of the three-minute mean while aligning the ends of the window, and then perform causal standardization. The specific formula is as follows: , , , , in, For This is the historical window at the end, with a window length of K; This is the (k-K+1)th three-minute average; Based on only Calculate the mean and standard deviation; This is the standardized three-minute average; Steps 4-2 to 4-4, minute encoder (CNN-BiLSTM, see...) Figure 7 First, several one-dimensional convolutional branches (small / medium / large kernels, etc.) are set on the minute-average sequence within the window to extract local temporal patterns and multi-scale neighborhood information. The convolutional outputs of each branch are concatenated along the channel dimension to obtain a continuous convolutional representation sequence. This sequence is then fed into a bidirectional LSTM, where the forward and backward links jointly model the short-term smooth changes and slow trends within the window, resulting in a bidirectional hidden state sequence. Finally, attention weighting (or equivalent statistical pooling, such as average / maximum) is applied at the output to converge the hidden states, resulting in a fixed-length "minute sequence feature" vector, which is used for subsequent fusion and prediction with the second-level converged representation.

[0033] Step 5 includes: adaptively combining the minute-level trend representation and the second-level convergent representation in the fusion layer, employing gated fusion (GF) and introducing cross-scale attention driven by intermittent weight maps to prioritize fluctuating segments and suppress intermittent segments; outputting the final wind speed prediction value through the prediction head, and calculating the evaluation metrics MAE, RMSE, and MAPE to verify the model's feasibility. The specific steps are as follows: Step 5-1: Align the minute-level minute sequence features with the second-level second-level converged representation to the same feature space, and adaptively combine them using gating. The specific formula is as follows: , , , , in, The convergence is represented in seconds, and y is the fusion vector; , Learnable parameters for the minute-side linear mapping; , Learnable parameters for second-level linear mapping; , Learnable parameters for the gating distributor; The aligned second-level side features; For the aligned minute-side features; This is element-wise multiplication; g is the gated weight vector; Step 5-2, Cross-scale attention of intermittent leaders: Cross-scale retrieval is performed using minute-level lateral vectors as queries and second-level time series features as keys and values, and a prior bias generated by the intermittent weight map is introduced to highlight fluctuating segments. The method includes the following steps: Step 5-2-1: Linearly map the minute-level features and second-level features to the shared attention space to form query, key, and value vectors. The specific formula is as follows: , , , in, , It is a linear mapping matrix.

[0034] Step 5-2-2: Generate prior biases for suppressing intermittent segments and emphasizing fluctuating segments based on the intermittent weight map. The specific formula is as follows: , in, This refers to the prior bias term injected into the attention scoring at time t; Step 5-2-3: Calculate the scaled dot product score and superimpose the prior, causal, and length masks to obtain constrained attention. The specific formula is as follows: , in: d is the length mask; d is the dimension of the attention space; Score the attention at time t; Step 5-2-4: Normalize the scores into attention weights and then weight and converge the value vectors to obtain a cross-scale weight vector. The specific formula is as follows: , , in: For attention weights, For a cross-scale vector, exp is the natural exponential function; Step 5-3, Fusion Output and Prediction Head: The vector-level fusion and attention retrieval are deeply combined to output a three-minute average wind speed prediction. The specific formula is as follows: , , in For lightweight return head, Here are the fusion mapping parameters; LN is the layer normalization. This is the next predicted value.

[0035] In this embodiment, the data comes from measured data along the Jiangsu section of the Beijing-Shanghai High-Speed ​​Railway. To verify the robustness of the proposed model and the effectiveness of each module, the following experiments were conducted, with specific examples provided:

[0036] I. To verify the versatility of the proposed IPM-MSTCN-IGSE-DP / CNN-BILSTM wind speed prediction model under different environments, this experiment selected three high-speed rail stations with different wind conditions: Dashengguan Bridge Station, Yangcheng West Lake Tunnel Station, and Zouwang Road Station in Jurong City. Dashengguan Bridge Station is located in an open bridge area, with relatively high average wind speed and stable wind direction, mainly exhibiting a low-frequency trend with moderate gusts. Yangcheng West Lake Tunnel Station shows a significant tunnel channel effect, strong high-frequency fluctuations, and rapid changes in wind direction. Zouwang Road Station in Jurong City is located along a suburban road with high surface roughness, resulting in low average wind speed but significant fluctuations, frequent local gusts, and dispersed wind directions.

[0037] like Figure 8 , Figure 9 , Figure 10 As shown, under different wind conditions, the prediction curve of the method proposed in this invention is more consistent with the actual wind speed change, and the error is lower than that of other models. The prediction curve of the traditional model fails to accurately reflect the rapid changes and long-term trends of wind speed, and obvious fluctuations, distortions and delays occur. Especially at the moment of sudden change in wind speed, the predicted value deviates significantly from the actual wind speed.

[0038] II. Verifying the effectiveness of each module in the IPM-MSTCN-IGSE-DP / CNN-BILSTM wind speed prediction model proposed in this invention: To verify the effectiveness of each module of the proposed model, the prediction errors of MSTCN-SE (multi-scale temporal convolutional network - channel attention), IPM-MSTCN-SE (intermittent prior module - multi-scale temporal convolutional network - channel attention), IPM-MSTCN-IGSE (intermittent prior module - multi-scale temporal convolutional network - intermittent weight-guided channel attention), and IPM-MSTCN-IGSE-DP intermittent prior module - multi-scale temporal convolutional network - intermittent weight-guided channel attention - intermittent / fluctuating dual-channel pooling) were statistically compared. Common evaluation metrics, including Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE), were used to evaluate the prediction accuracy of each model. The specific results are shown in Table 1.

[0039] Table 1 Evaluation Results

[0040]

[0041] The results in the table show that the IPM-MSTCN-IGSE-DP model proposed in this invention has the lowest values ​​in all three error evaluation indicators: MAE, RMSE, and MAPE. Furthermore, the addition of the IPM, IG-SE, and DP modules significantly improves accuracy compared to the baseline model. Therefore, the method proposed in this invention exhibits superior predictive performance compared to traditional models.

[0042] This invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. It should be noted that each module in the above system corresponds to a specific step of the method provided in this invention, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention.

[0043] This invention also proposes a computer-readable storage medium storing a computer program. It should be noted that each module in the above system corresponds to a specific step of the method provided in this invention, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention.

[0044] This invention provides a method for predicting wind speed along high-speed railway lines based on intermittent prior and cross-scale adaptive fusion. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for predicting wind speed along high-speed railway lines based on intermittent prior and cross-scale adaptive fusion, characterized in that, Includes the following steps: Step 1: Collect and align the second-level and minute-level average wind speeds from monitoring stations along the route, and organize them as model inputs using a sliding window. Step 2, constructing an intermittent prior module: under causal constraints, calculate prior features for second-level wind speed, and use a small multilayer perceptron network to adaptively generate intermittent weight maps and soft play sequences to quantify the intermittency and duration that remain unchanged in continuous time steps; Step 3: Second-level encoder is used to extract second-level features. The second-level encoder uses a multi-scale temporal convolutional network (MSTCN) to recalibrate the channel weights under intermittent channel attention. Intermittent and fluctuating time-varying features are aggregated through intermittent and fluctuating dual-channel pooling to form a second-level aggregated representation. Step 4: Encode the historical mean at the minute level using a minute-level encoder to obtain a minute-level trend representation; the minute-level encoder includes a convolutional neural network and a bidirectional long short-term memory network CNN-BILSTM. Step 5: Adaptively fuse minute-level trends and second-level convergence representations through a fusion layer with intermittently guided cross-scale attention to form a fused representation, which is then fed into the prediction head. The prediction results are output and evaluated to verify the feasibility of the model.

2. The method according to claim 1, characterized in that, Step 1 includes the following steps: Step 1-1, set up a second-level wind speed observation sequence ,in This represents the wind speed observation value at second t, where t ranges from 1 to n, and n represents the total number of wind speed observation points. Step 1-2: Starting from the c-th wind speed observation point, take a subsequence of length w, using the following formula: ， in, Let c represent a sliding window that satisfies ; Steps 1-3: Calculate the three-minute average within each window and align the timestamps to the end of the window. The specific formula is as follows: ， ， in, This represents the 3-minute moving average starting from the c-th wind speed observation point, which serves as the input to the minute-level encoder. The mean Corresponding time tags; The time label represents the last observation point of the c-th sliding window; Steps 1-4, in the same window As the input segment for a second-level encoder, it is denoted as: ， in, This represents the c-th sliding window.

3. The method according to claim 2, characterized in that, Step 2 includes the following steps: Step 2-1, calculate the first-order difference and robust amplitude: ， ， in, It is a first-order difference. For robust amplitude; It is a smoothing constant; Step 2-2: Without using future information, estimate the short-term strength of local movement intensity and relative trend using the exponentially weighted moving average (EMA). The specific formula is as follows: ， ， ， The exponentially weighted moving average (EMA) is defined as follows: ， in, For smoothing coefficients; Let be the local motion intensity at time t; As the trend baseline; This represents the strength of short-term changes in relative trends; An exponentially weighted moving average operator for calculating local motion intensity; An exponentially weighted moving average operator for calculating a short-term trend baseline; An exponentially weighted moving average operator for calculating the strength of short-term changes in relative trends; For exponentially weighted moving average operators; Steps 2-3 involve calculating the mean and variance using the exponentially weighted moving average (EMA) based on the historical window, and obtaining the standardized deviation to measure the significance of the current wind speed relative to historical levels. The specific formula is as follows: ， ， ， in, The smoothing coefficient for the causal mean and variance; This is the exponentially weighted moving average at second t; For causal variance estimation; Standardized deviation; An exponentially weighted moving average operator for calculating the arithmetic mean of a window; The wind speed observation value is at the (t-1)th second; It is a stable term; Steps 2-4 involve concatenating the causal features according to time steps to form a prior feature vector, and then incorporating the intermittent prior from the previous step to enhance temporal coherence. The specific formula is as follows: ， in, For the prior feature vector, For the intermittent weights in the previous step, This is the previous soft journey; Steps 2-5 involve using a small multilayer perceptron (MLP) to perform nonlinear mapping on the prior features, outputting intermittent discriminative features, and controlling the temperature parameters for thresholding hardness and softness. The specific formula is as follows: ， ， in, For intermittent discrimination features; These are unconstrained temperature parameters; This is the soft threshold parameter after adding a lower bound; This represents the minimum value of the soft threshold parameter. , LN is a positive-valued smooth activation function; LN is a layer normalization function. Steps 2-6: Use a Sigmoid soft gate to map the intermittent discriminative features to the soft threshold parameter. The continuous weights of the intervals are used to form an intermittent weighted graph, and the specific formula is as follows: ， in, It is a non-linear activation function; Intermittent weighting; Steps 2-7 use differentiable recursion to estimate the duration of the current intermittent state, improving the stability of the segment boundaries. The specific formula is as follows: ， Among them, soft runs The initial value is 0. This is the discharge coefficient; This is an estimate of the duration of the intermittent state up to t-1 seconds.

4. The method according to claim 3, characterized in that, Step 3 includes the following steps: Step 3-1: The multi-scale temporal convolutional network MSTCN is used to extract features from a second-level input of length w, so as to simultaneously cover short, medium, and long receptive fields. The specific process includes: Step 3-1-1: Let the current second-level window input be S, construct B parallel one-dimensional convolutional branches, and the b-th branch uses a convolutional kernel length of... With expansion coefficient To form different receptive fields; when The window is filled with the initial value or zero, and future samples are not used. The specific formula is as follows: ， in, ; For branch channel index; For the b-th branch at time t, the channel The output value; For the b-th branch, the The convolution weights of each output channel at a lag of j steps; This indicates the number of lag steps skipped based on the hole rate; Looking back from the current time t Input of one lag step; For the b-th branch, the The bias term for each output channel; Step 3-1-2: Concatenate the outputs of each branch along the channel dimension, and compress and shape the concatenated features to obtain unified temporal features. The specific formula is as follows: ， in, This is the output feature map of the b-th branch; The channel-dimensional splicing operator splices the output feature maps of B branches; is a channel-mixed convolution; H is the intermediate feature map after channel mixing; Step 3-1-3: Perform residual projection on the input S to match its channel number with H, then add it to H and perform layer normalization. The specific formula is as follows: ， in This is a projection of the one-dimensional convolutional residual onto the input S; This is the final output feature map; Step 3-2, under the intermittently guided channel attention mechanism, The channels are recalibrated so that the model learns different channel importance for intermittent and fluctuating segments. This includes the following steps: Step 3-2-1: Calculate the normalized coefficients for intermittent and fluctuating coverage. The formula is: ， ， in ; This is the normalized denominator for intermittent coverage; This is the normalized denominator for the volatility coverage. Step 3-2-2: Perform weighted compression on each channel from three perspectives: overall, intermittent, and fluctuation. The formula is as follows: ， ， ， in, For the overall response of the channel, Emphasizing intermittent segments, Emphasizing fluctuation segments; The output feature map of channel c at time t; Step 3-2-3: The three types of statistics are concatenated with the global statistics of the intermittent prior, and then fed into a small multilayer perceptron (MLP) to generate channel weights for channel importance estimation. The formula is: ， ， ， in, For channel weights, , and The mean and variance of the intermittent weights are respectively used. The output of the small multilayer perceptron (MLP) is compressed into channel weights using a sigmoid function. ; Step 3-2-4 involves multiplicative recalibration of the feature maps, enabling the fusion and prediction layers to learn different channel importances for intermittent and fluctuating segments. The formula is as follows: ， in, The feature map after channel recalibration; Step 3-3: Intermittent and fluctuating dual-channel pooling is used to aggregate the time-varying features of intermittent and fluctuating data respectively, and a second-level aggregated representation is formed for subsequent fusion and prediction.

5. The method according to claim 4, characterized in that, In step 3-3, the following calculation formula is used: ， ， ， in, Let C be the channel vector at time step t; C is the number of channels. To emphasize the weighted average of intermittent segments; To emphasize the weighted average of fluctuating segments; For second-level side output, Indicates length is A real vector.

6. The method according to claim 5, characterized in that, Step 4 includes the following steps: Step 4-1: Construct a historical window of the three-minute mean while aligning the ends of the window, and then perform causal standardization. The specific formula is as follows: ， ， ， ， in, For This is the historical window at the end, with a window length of K; This is the (k-K+1)th three-minute average; Based on only Calculate the mean and standard deviation; This is the standardized three-minute average; Step 4-2: Extract local temporal patterns and multi-scale neighborhood information from the window using one-dimensional convolution, and concatenate them in the channel dimension to obtain the convolutional representation. The specific formula is as follows: ， ， ， in, , , This is the convolution output of the r-th branch at time step i; The channel vector is the result of concatenating multiple branches; U represents the convolutional sequence. is the value of the three-minute mean after causal standardization at index iq; q is the position index within the convolution kernel; Let be the convolution weight of the r-th branch at kernel position q; Let be the bias of the r-th branch; Step 4-3: Feed the convolutional representation into a bidirectional long short-term memory network to model the long short-term dependencies within the window and obtain the bidirectional hidden state sequence. The specific formula is as follows: ， ， in, These are the forward hidden state and backward hidden state of BiLSTM, respectively. It is a bidirectional long short-term memory network; To splice the matrix; Step 4-4: Attention weighting and statistical pooling are used to aggregate the hidden state sequence into fixed-length minute sequence features for subsequent fusion and prediction. The calculation formula is as follows: ， ， in: For attention weights, To measure sequence features in minutes, T represents transpose. is the hyperbolic tangent activation function; e is the learnable weight vector for attention scoring.

7. The method according to claim 6, characterized in that, Step 5 includes the following steps: Step 5-1: Align the minute-level minute sequence features with the second-level second-level converged representation to the same feature space, and adaptively combine them using gating. The specific formula is as follows: ， ， ， ， in, The convergence is represented in seconds, and y is the fusion vector; , Learnable parameters for the minute-side linear mapping; , Learnable parameters for second-level linear mapping; , Learnable parameters for the gating distributor; The aligned second-level side features; For the aligned minute-side features; This is element-wise multiplication; g is the gated weight vector; Step 5-2, Cross-scale attention to intermittent leaders: Cross-scale retrieval is performed using minute-level lateral vectors as queries and second-level time series features as keys and values. A prior bias generated by the intermittent weight map is introduced to highlight fluctuating segments. This specifically includes the following steps: Step 5-2-1: Linearly map the minute-level features and second-level features to the shared attention space to form the query vector Q and the key vector. AND value vector The specific formula is as follows: ， ， ， in , It is a linear mapping matrix; Step 5-2-2: Generate prior biases for suppressing intermittent segments and emphasizing fluctuating segments based on the intermittent weight map. The specific formula is as follows: ； in, This refers to the prior bias term injected into the attention scoring at time t; Step 5-2-3: Calculate the scaled dot product score and superimpose the prior, causal, and length masks to obtain constrained attention. The specific formula is as follows: ， in d is the length mask; d is the dimension of the attention space; Score the attention at time t; Step 5-2-4: Normalize the scores into attention weights and then weight and converge the value vectors to obtain a cross-scale weight vector. The specific formula is as follows: ， ， in: Here, z represents the attention weights, z is the cross-scale vector, and exp is the natural exponential function. Step 5-3, Fusion Output and Prediction Head: Deeply combine vector-level fusion and attention retrieval to output a three-minute average wind speed prediction.

8. The method according to claim 7, characterized in that, In step 5-3, the following calculation formula is used: ， ， in For lightweight return head, These are the parameters for fusion mapping; This is the next predicted value.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 8.

Images