A high-speed rail line wind speed prediction method based on intermittent prior and cross-scale adaptive fusion

Abstract

The application provides a high-speed rail line wind speed prediction method based on intermittent prior and cross-scale adaptive fusion, comprising the following steps: step 1, collecting and aligning the second-level wind speed and the minute-level average of the monitoring stations along the line, and organizing the model input according to the sliding window; step 2, constructing an intermittent prior module; step 3, extracting the second-level features by using a second-level encoder, gathering the time-varying features of intermittence and fluctuation through intermittent and fluctuation double-path pooling respectively, and forming a second-level gathering representation; step 4, encoding the minute-level historical average by using a minute-level encoder to obtain a minute trend representation; step 5, adaptively fusing the minute trend and the second-level gathering representation through a fusion layer containing intermittent guide cross-scale attention, forming a fusion representation and sending it into a prediction head, outputting a prediction result and performing evaluation. The method can effectively improve the accuracy of the wind speed prediction value, and greatly improve the reliability of the wind speed prediction value.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of time series prediction, in particular to a high-speed rail along the line wind speed prediction method based on intermittent prior and cross-scale adaptive fusion. BACKGROUND

[0002] The current high-speed rail disaster prevention system has a real-time monitoring and alarm function for strong winds, but does not have a warning function, and the disaster warning capability is insufficient. If the system can predict the wind speed value in the future through the current wind speed data, it will help to issue an alarm in time and provide a scientific basis for train scheduling and safety regulations, thereby improving the safety, reliability and risk resistance of the high-speed rail system.

[0003] Although domestic and foreign scholars have conducted a large number of studies on wind speed prediction, there is no research on meeting the prediction of the railway along the line of second-level granularity. And the high-speed rail area is complex, and the strong wind is frequent, so it is difficult to capture, so the multi-step prediction lag problem of the algorithm will inevitably be faced in the wind speed multi-step time series prediction along the railway, and solving this problem has become a bottleneck. At present, most of the researches mainly use the minute average sequence obtained from the second-level wind speed measurement to predict the wind speed. Although this average processing effectively reduces the short-term fluctuations and noise, making the input data more stable, it also eliminates the fine time characteristics inherent in the second-level measurement - these characteristics are essential for capturing fast wind dynamics and ensuring high prediction accuracy.

[0004] Therefore, how to fully tap the time sequence structure characteristics and fluctuation change law of the wind speed signal itself, develop a high-precision prediction method for high-speed railway along the wind speed, and solve the core technical problems such as multi-step prediction lag and mismatch between training samples and predicted samples, has become a key problem that needs to be broken through to improve the disaster warning capability of high-speed rail and ensure the safety of train operation. SUMMARY

[0005] The technical problem to be solved by the present application is that, in view of the shortcomings of the prior art, the present application proposes a high-speed rail along the line wind speed prediction method based on intermittent prior and cross-scale adaptive fusion. This method explicitly quantifies and distinguishes the time structure of "intermittence and fluctuation" without using future information, combines multi-scale time convolution and prior-guided channel attention, double-path pooling and cross-scale attention, realizes the adaptive alignment and fusion of minute-level trend and second-level features, and outputs the final prediction value, which can effectively improve the accuracy of wind speed prediction value.

[0006] The method of the present application comprises the following steps:

[0007] Step 1, collect and align the second-level wind speed and minute-level average of the monitoring site along the line, and organize it into model input according to the sliding window;

[0008] Step 2, Intermittency Prior Module (IPM) construction: Calculate the prior features such as the difference intensity, high-frequency energy, and causal standardized deviation under the causal constraint, and generate the intermittency weight map and soft walker sequence adaptively by a small multi-layer perception network, which are used to quantify the intermittency and duration that remain unchanged at continuous time steps.

[0009] Step 3, extract the second-level features using the second-level encoder, which uses a multi-scale temporal convolutional network (MSTCN) to re-label the channel weights under the guidance of intermittency-guided squeeze-and-excitation (IG-SE), and then aggregate the time-varying features of intermittency and fluctuation through dual-path pooling for intermittency / fluctuation (DP) to form the second-level aggregated representation.

[0010] Step 4, encode the minute-level historical mean using the minute-level encoder to obtain the minute trend representation; the minute-level encoder includes a convolutional neural network and a bidirectional long short-term memory network (CNN-BILSTM).

[0011] Step 5, adaptively fuse the minute trend and the second-level aggregated representation through the fusion layer with intermittency-guided cross-scale attention to form the fused representation and send it to the prediction head, output the prediction result and evaluate the model feasibility.

[0012] Step 1 includes the following steps:

[0013] Step 1-1, let the second-level wind speed observation sequence be where represents the wind speed observation value at the t-th second, t takes values from 1 to n, n represents the total number of wind speed observation points, the sampling interval is 1 second, and the time index is monotonically increasing;

[0014] Step 1-2, to calculate the statistics of a fixed length, take a subsequence of length w from the c-th wind speed observation point, the specific formula is:

[0015] ,

[0016] where represents the sliding window, c satisfies , preferably seconds, corresponding to 3 minutes;

[0017] Step 1-3, calculate the three-minute mean within each window and align the timestamp to the end of the window to ensure that only historical information is used, the specific formula is:

[0018] ,

[0019] ,

[0020] wherein, represents the 3-minute sliding mean value starting from the cth wind speed observation point as the minute-level encoder input; is the mean corresponding time label, so as to avoid information leakage caused by using any future samples; represents the time label of the last observation point of the cth sliding window;

[0021] Step 1-4, the same window , as the input segment of the second encoder, denoted as:

[0022] ,

[0023] wherein, represents the cth sliding window.

[0024] Step 2 includes the following steps:

[0025] Step 2-1, to measure the change amplitude of adjacent time points and reduce the influence of transient abnormal values on subsequent discrimination, calculate the first-order difference and robust amplitude:

[0026] ,

[0027] ,

[0028] wherein, is the first-order difference, is the robust amplitude; is a smoothing constant (a );

[0029] Step 2-2, estimate the local motion intensity and the short-time change strength of the relative trend without using future information by using the Exponential Weighted Moving Average (EMA), the specific formula is:

[0030] ,

[0031] ,

[0032] ,

[0033] The exponential moving average EMA is defined as:

[0034] ,

[0035] where, is the smoothing factor; is the local motion intensity at time t; is the trend baseline, used to represent the slowly varying component of the wind speed; is the short-term variation intensity of the relative trend; is the exponential moving average operator for calculating the local motion intensity; is the exponential moving average operator for calculating the short-term trend baseline; is the exponential moving average operator for calculating the short-term variation intensity of the relative trend; is the exponential moving average operator;

[0036] Step 2-3, based on the historical window, use the exponential moving average EMA to calculate the mean and variance and get the normalized deviation, which is used to measure the significance of the current wind speed relative to the historical level, the specific formula is:

[0037] ,

[0038] ,

[0039] ,

[0040] where, is the smoothing factor of the causal mean and variance; is the exponential moving average value at the tth second; is the causal variance estimate; is the normalized deviation; is the exponential moving average operator for calculating the window arithmetic mean; is the wind speed observation at the t-1th second; is the stability term;

[0041] Step 2-4, concatenate the causal features by time step to form the prior feature vector, and concatenate the intermittent prior of the last step to enhance the time continuity, the specific formula is:

[0042] ,

[0043] where, is the prior feature vector, is the last step intermittent weight (optional), Soft run (optional) of the previous step;

[0044] Step 2-5, use small multi-layer perception (MLP) to map the prior feature nonlinearly, output the intermittent discriminant feature quantity, and control the temperature parameter (soft threshold parameter) of the thresholding soft and hard degree. The specific formula is:

[0045] ,

[0046] ,

[0047] Among them, is the intermittent discriminant feature quantity (the larger the value, the more inclined to intermittent); is the unconstrained temperature parameter; is the soft threshold parameter after adding the lower bound; is the minimum value of the soft threshold parameter, , is a positive smoothing activation function, which is used to ensure that the unconstrained temperature parameter is positive; LN is layer normalization;

[0048] Step 2-6, use Sigmoid soft gate to map the intermittent discriminant feature quantity and the soft threshold parameter to the continuous weight in the interval , forming the intermittent weight map, the specific formula is:

[0049] ,

[0050] Among them, is a nonlinear activation function; is the intermittent weight, the larger the value, the closer to "intermittent (multi-step unchanged)", the smaller the value, the closer to "fluctuation";

[0051] Step 2-7, use the differentiable recursive estimation to estimate the duration of the current intermittent state, and improve the stability of the segment boundary, the specific formula is:

[0052] ,

[0053] Among them, the initial value of the soft run is 0, is the discharge coefficient; is the duration estimate of the intermittent state until t-1 second.

[0054] Step 3 includes the following steps:

[0055] Step 3-1, use multi-scale time convolution network (MSTCN) to extract features from the input of the second-level window with a length of w, to cover short, medium and long time receptive fields at the same time. The specific process includes:

[0056] Step 3-1-1, assuming the current second-level window input is S, B parallel one-dimensional convolution branches are constructed, the convolution kernel length of the b-th branch is with the expansion coefficient to form different receptive fields; when , the window starting value or zero is filled, and future samples are not used, and the specific formula is:

[0057] ,

[0058] wherein, ; is the branch channel index; is the output value of the b-th branch at time t, channel ; is the convolution weight of the b-th branch, the j-th output channel at the lag j steps; represents the number of lag steps skipped according to the hole rate; is the input looking back lag steps forward from the current time t; is the bias term of the b-th branch, the j-th output channel;

[0059] Step 3-1-2, the branch outputs are spliced along the channel dimension, and the spliced features are compressed and shaped to obtain unified time sequence features, and the specific formula is:

[0060] ,

[0061] wherein, is the output feature map of the b-th branch; is a channel splicing operator that splices the output feature maps of the B branches; is a channel mixed convolution; H is the intermediate feature map after channel mixing;

[0062] Step 3-1-3, to enhance gradient propagation and numerical stability, the input S is projected to make its channel number match H, and then added to H and normalized, and the specific formula is:

[0063] ,

[0064] wherein is a one-dimensional convolution residual projection on the input S; is the final output feature map;

[0065] Step 3-2, under the intermittent guided channel attention mechanism, the channels of are reweighted, so that the model learns different channel importance on intermittent and fluctuating segments, including the following steps: ​​

[0066] Step 3-2-1, calculate the normalization coefficient of intermittent and fluctuation coverage, which is used as the denominator and avoids the case of zero value, the formula is:

[0067] ,

[0068] ,

[0069] where ; is the normalization denominator of intermittent coverage; is the normalization denominator of fluctuation coverage;

[0070] Step 3-2-2, respectively, in the overall, intermittent and fluctuation three perspectives, do weighted compression (squeeze) for each channel, which is used to extract the statistical characteristics of the channel, the formula is:

[0071] ,

[0072] ,

[0073] ,

[0074] where, is the overall response of the channel, emphasizes the intermittent segment, emphasizes the fluctuation segment; is the output feature map of channel c at time t;

[0075] Step 3-2-3, splice three kinds of statistics and the global statistics of the intermittent prior into a small multi-layer perception MLP to generate channel weights, which is used to complete the importance estimation of the channel, the formula is:

[0076] ,

[0077] ,

[0078] ,

[0079] where, is the channel weight, , and are the mean and variance of the intermittent weight, respectively, which are used as the prior description of the intermittent proportion and stability of the current window; the small multi-layer perception MLP output is compressed by Sigmoid to the channel weight ;

[0080] Step 3-2-4, multiply the feature map to perform re-scaling, so that the fusion and prediction layers learn different channel importance on intermittent and fluctuation segments, and the formula is:

[0081] ,

[0082] is the feature map after channel re-scaling;

[0083] Step 3-3, adopt intermittent and fluctuation double-path pooling to gather time-varying features of intermittent and fluctuation respectively, and form a second-level aggregation representation for subsequent fusion and prediction.

[0084] In step 3-3, the following calculation formula is adopted:

[0085] ,

[0086] ,

[0087] ,

[0088] wherein, is the channel vector at time step t; C is the number of channels; is the weighted average emphasizing intermittent segments; is the weighted average emphasizing fluctuation segments; is the second-level side output, represents a real vector with length .

[0089] Step 4 includes the following steps:

[0090] Step 4-1, under the premise of aligning at the end of the window, construct a three-minute mean history window and perform causal standardization to provide a stable input sequence, and the specific formula is:

[0091] ,

[0092] ,

[0093] ,

[0094] ,

[0095] wherein, is the history window with as the end, and the window length is K; is the k-th three-minute mean; are respectively the mean and standard deviation calculated only based on ; is the normalized three-minute mean;​

[0096] Step 4-2, local temporal patterns and multi-scale neighborhood information are extracted from the window using one-dimensional convolution, and the convolutional representation is obtained by concatenating in the channel dimension. The specific formula is:

[0097]

[0098]

[0099]

[0100] wherein, is the convolutional output of the rth branch at time step i; is the channel vector after multi-branch concatenation; U is the convolutional representation sequence; is the value of the three-minute mean at index i-q after causal normalization; q is the position index within the convolution kernel; is the convolutional weight of the rth branch at kernel position q; is the bias of the rth branch;

[0101] Step 4-3, the convolutional representation is input into a bidirectional long short-term memory network to model the long and short-term dependencies within the window, and a bidirectional hidden state sequence is obtained. The specific formula is:

[0102]

[0103]

[0104] wherein, are the forward hidden state and backward hidden state of BiLSTM, respectively; is the bidirectional long short-term memory network; is the concatenation matrix;

[0105] Step 4-4, the hidden state sequence is aggregated into a fixed-length minute sequence feature using attention weighting and statistical pooling, so as to be fused and predicted subsequently. The calculation formula is:

[0106]

[0107]

[0108] wherein, is the attention weight, is the minute sequence feature, T represents transposition, is the hyperbolic tangent activation function, used to compress the linearly mapped vector to the range of (-1, 1); e is a learnable weight vector for attention scoring. ​​​​​​​​​

[0109] Step 5 includes the following steps:

[0110] Step 5-1, align the minute-level sequence feature to the same feature space with the second-level side second-level aggregation representation, and adaptively combine through gating, the specific formula is:

[0111] ,

[0112] ,

[0113] ,

[0114] ,

[0115] wherein, is the second-level aggregation representation, y is the fusion vector; , is the learnable parameter of the linear mapping on the minute side; , is the learnable parameter of the linear mapping on the second-level side; , is the learnable parameter of the gating distributor; is the aligned second-level side feature; is the aligned minute side feature; is the element-wise multiplication; g is the gating weight vector;

[0116] Step 5-2, cross-scale attention of intermittent leader: cross-scale retrieval is performed with the minute side vector as the query and the second-level time sequence feature as the key and value, and a prior bias generated by the intermittent weight graph is introduced to highlight the fluctuation segment, which includes the following steps:

[0117] Step 5-2-1, linearly map the minute side feature and the second-level feature to a shared attention space to form the query vector Q, the key vector and the value vector , the specific formula is:

[0118] ,

[0119] ,

[0120] ,

[0121] wherein , is the linear mapping matrix;

[0122] Step 5-2-2, generate a prior bias based on the intermittent weight graph to suppress the intermittent segment and emphasize the fluctuation segment, the specific formula is:

[0123] ;

[0124] wherein, is the prior bias term injected into the attention score at time t;

[0125] Step 5-2-3, calculate the scaled dot-product score and superimpose the prior and causal, length mask to obtain the attention with constraints, the specific formula is:

[0126] ,

[0127] wherein, is the length mask; d is the attention space dimension; is the attention score at time t;

[0128] Step 5-2-4, normalize the score to the attention weight and weight the value vector to obtain the cross-scale weight vector, the specific formula is:

[0129] ,

[0130] ,

[0131] wherein, is the attention weight, z is the cross-scale vector, and exp is the natural exponential function;

[0132] Step 5-3, fusion output and prediction head: vector-level fusion and attention retrieval are deeply combined to output three-minute average wind speed prediction.

[0133] In step 5-3, the following calculation formula is adopted:

[0134] ,

[0135] ,

[0136] wherein is a light regression head, is a fusion mapping parameter; is the next prediction value.

[0137] The application also provides an electronic device comprising a processor and a memory, the memory storing program code which, when executed by the processor, causes the processor to perform the steps of the method.

[0138] The application also provides a storage medium storing a computer program or instructions which, when executed on a computer, perform the steps of the method.

[0139] Beneficial effects: The high-speed rail line second-level wind speed prediction method based on intermittent prior and cross-scale adaptive fusion (ICAF-Net, Intermittency-guided Cross-scale Adaptive Fusion Network) provided by the application adopts an intermittent prior module (IPM) to explicitly quantify the "intermittency / undulation" structure and generate weights and soft runs; a multi-scale time convolution network (MSTCN) is used to extract short / medium / long receptive field features; an intermittent guide channel attention (IG-SE) is introduced to recalibrate the channel; intermittent / undulation dual-path pooling (DP) is used to respectively gather two types of dynamics; a CNN-BiLSTM is used to encode the minute-level trend; and in the fusion stage, cross-scale attention and gate fusion (CSA+GF) are used for adaptive alignment and key attention of the minute trend and second-level features. The above-mentioned modules work together, which can effectively suppress platform noise, highlight key fluctuations, alleviate multi-step extrapolation lag and error accumulation, and significantly improve the accuracy and reliability of the wind speed prediction value; among them, the core modules such as IPM and CSA are particularly significant for improving the accuracy, and the stability of the wind speed prediction result used for line warning and scheduling decision is greatly improved, meeting the online and low-latency application requirements of the high-speed rail scene. BRIEF DESCRIPTION OF DRAWINGS

[0140] Figure 1 is the overall flowchart of the application.

[0141] Figure 2 is the intermittent prior module and small multi-layer perceptron structure diagram of the application.

[0142] Figure 3 is the time convolution network TCN structure diagram of the application.

[0143] Figure 4 is the second-level encoder MSTCN structure diagram of the application.

[0144] Figure 5 is the intermittent weight guided channel attention (IG-SE) structure diagram of the application.

[0145] Figure 6 is the intermittent / undulation dual-path pooling (DP) structure diagram of the application.

[0146] Figure 7 is the minute encoder CNN-BILSTM structure diagram of the application.

[0147] Figure 8 is the prediction effect diagram of the application on the Dashengguan data set.

[0148] Figure 9 is the prediction effect diagram of the application on the Yangchenghu data set.

[0149] Figure 10 is the prediction effect diagram of the present application on the Zouwanglu data set. DETAILED DESCRIPTION

[0150] The above and / or other aspects of the present application will become apparent and more readily appreciated from the following description, taken in conjunction with the accompanying drawings, in which:

[0151] An intermittent prior and cross-scale adaptive fusion-based high-speed rail line wind speed prediction method is provided in the embodiments of the present application, as shown in the following formula (1): Figure 1 The method comprises the following steps:

[0152] Step 1, collect the second-level wind speed sequence of the line monitoring station, and complete the timestamp unification and station alignment; under the causal constraint, calculate the three-minute mean value in each window based on the second-level sequence using a causal sliding window with a length of w=180 seconds, and align the mean value to the end time point of the window; then organize the data in the window into a model input sample, and do not use the information outside or in the future.

[0153] Step 2, intermittent prior module (IPM) construction, causal local feature extraction is performed on the second-level wind speed sequence in the current sliding window: calculate and collect the prior quantities such as difference intensity, high-frequency energy, standardized deviation, approximate constant segment length / percentage, etc., and splice them into a prior vector in a fixed order; input the prior vector into a small multi-layer perceptron to adaptively map and obtain an intermittent weight map and a soft walker sequence at the time step level. The structure of the intermittent prior module is shown in the following formula (2): Figure 2 The specific steps are as follows:

[0154] Step 2-1, to measure the change amplitude of adjacent time points and reduce the influence of transient outliers on subsequent discrimination, calculate the first-order difference and robust amplitude, and the specific formula is as follows:

[0155] ,

[0156] ,

[0157] Among them, is the first-order difference, is the robust amplitude; is a smoothing constant ( );

[0158] Step 2-2, estimate the local motion intensity and the short-time change strength of the relative trend without using future information using the exponential weighted moving average EMA (Exponential Weighted Moving Average), and the specific formula is as follows:

[0159] ,

[0160] ,

[0161] ,

[0162] The exponential moving average EMA is defined as:

[0163] ,

[0164] wherein, is a smoothing factor; is the local motion intensity at time t; is a trend baseline, used to represent the slowly varying component of the wind speed; is the short-term variation strength of the relative trend; is an exponential moving average operator for calculating the local motion intensity; is an exponential moving average operator for calculating the short-term trend baseline; is an exponential moving average operator for calculating the short-term variation strength of the relative trend; is an exponential moving average operator;

[0165] Step 2-3, based on the historical window, the mean and variance are calculated by the exponential moving average EMA, and the normalized deviation is obtained, which is used to measure the significance of the current wind speed relative to the historical level, and the specific formula is:

[0166] ,

[0167] ,

[0168] ,

[0169] wherein, is a smoothing factor of the causal mean and variance; is the exponential moving average value at the tthsecond; is the causal variance estimate; is the normalized deviation; is an exponential moving average operator for calculating the window arithmetic mean; is the wind speed observation value at the t-1thsecond; is a stability term;

[0170] Step 2-4, the causal features are spliced according to the time step to form the prior feature vector, and the intermittent prior of the last step is spliced to enhance the time continuity, and the specific formula is:

[0171] ,

[0172] wherein, is a prior feature vector, The intermittent weight of the previous step, The soft run of the previous step;

[0173] Step 2-5, use small multi-layer perception MLP to map the prior feature nonlinearly, output the intermittent discriminant feature quantity, and control the temperature parameter of the thresholding soft and hard degree. The specific formula is:

[0174] ,

[0175] ,

[0176] Among them, The intermittent discriminant feature quantity; The unconstrained temperature parameter; The soft threshold parameter after adding the lower bound; The minimum value of the soft threshold parameter, , The positive smoothing activation function; LN is the layer normalization;

[0177] Step 2-6, use Sigmoid soft gate to map the intermittent discriminant feature quantity and the soft threshold parameter to the continuous weight in the interval , Form the intermittent weight map, the specific formula is:

[0178] ,

[0179] Among them, The nonlinear activation function; The intermittent weight;

[0180] Step 2-7, use the differentiable recursive estimation to estimate the duration of the current intermittent state, and improve the stability of the segment boundary. The specific formula is:

[0181] ,

[0182] Among them, the initial value of the soft run Is 0, The discharge coefficient; The duration estimate of the intermittent state until t-1 second.

[0183] 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:

[0184] 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.

[0185] 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:

[0186] Step 3-2-1: Calculate the normalized coefficients for intermittent and fluctuating coverage. The formula is:

[0187] ,

[0188] ,

[0189] in ; This is the normalized denominator for intermittent coverage. This is the normalized denominator for the fluctuation coverage.

[0190] Step 3-2-2: Perform weighted compression on each channel from three perspectives: overall, intermittent, and fluctuation. The formula is as follows:

[0191] ,

[0192] ,

[0193] ,

[0194] in, for the overall response of the channel, highlighting the pause segment, highlighting the fluctuation segment; for the output feature map of channel c at time t;

[0195] Step 3-2-3, splice the three types of statistics with the global statistics of the pause prior into a small multi-layer perception MLP to generate channel weights for completing the importance estimation of the channel, the formula is:

[0196] ,

[0197] ,

[0198] ,

[0199] wherein, is the channel weight, , and are the mean and variance of the pause weight respectively, the output of the small multi-layer perception MLP is compressed by Sigmoid to the channel weight ;

[0200] Step 3-2-4, multiply the feature map to re-scale it, so that the fusion and prediction layers learn different channel importance on the pause and fluctuation segments, the formula is:

[0201] ,

[0202] wherein, is the feature map after channel re-scaling;

[0203] Step 3-3, adopt pause and fluctuation dual-path pooling to gather time-varying features of pause and fluctuation respectively, and form a second-level aggregation representation for subsequent fusion and prediction. The structure diagram of the DP module is as Figure 6 shown, and the specific formula is:

[0204] ,

[0205] ,

[0206] ,

[0207] wherein, is the channel vector at time step t; C is the number of channels; is the weighted average of the highlighted pause segment; is the weighted average of the highlighted fluctuation segment; is the second-level side output, denotes a real vector with a length of .

[0208] Step 4, encode the minute-level historical mean using a combination of convolution and bidirectional long short-term memory network (CNN-BiLSTM, Convolutional Neural Network + Bidirectional Long Short-Term Memory), where convolution captures local patterns and neighborhood smoothing, and BiLSTM models slow trends and context dependencies, to obtain a minute trend representation. The specific steps are as follows:

[0209] Step 4-1, construct a three-minute mean historical window under the premise of alignment at the window end and perform causal standardization, with the specific formula being:

[0210] ,

[0211] ,

[0212] ,

[0213] ,

[0214] wherein, is a historical window with as the end, and the window length is K; is the k-K+1 three-minute mean; are the mean and standard deviation calculated based on respectively; is the normalized three-minute mean;

[0215] Steps 4-2~4-4, minute encoder (CNN-BiLSTM, see Figure 7 ). First, set several one-dimensional convolution branches (small / medium / large kernel, etc.) on the minute mean sequence within the window, so as to extract local time patterns and multi-scale neighborhood information; the convolution outputs of each branch are concatenated in the channel dimension to obtain a continuous convolution representation sequence. Then, the sequence is fed into a bidirectional LSTM to jointly model the short-term smooth changes and slow trends within the window through forward and backward links, to obtain a bidirectional hidden state sequence. Finally, attention weighting (or equivalent statistical pooling, such as average / maximum) is used at the output end to aggregate the hidden states to obtain a fixed-length "minute sequence feature" vector, which is used for subsequent fusion and prediction with the second-level aggregation representation.

[0216] Step 5 includes: in the fusion layer, adaptively combining the minute trend representation and the second-level aggregation representation, using gated fusion (GF, Gated Fusion) and introducing cross-scale attention driven by an intermittent weight map to preferentially focus on fluctuation segments and suppress intermittent segments; outputting the final wind speed prediction value through the prediction head, and calculating the evaluation indexes MAE, RMSE and MAPE to verify the feasibility of the model. The specific steps are as follows:

[0217] Step 5-1, align the minute sequence features on the minute side to the same feature space as the second-level aggregation representation on the second-level side, and adaptively combine them through gating. The specific formula is:

[0218] ,

[0219] ,

[0220] ,

[0221] ,

[0222] wherein, is the second-level aggregation representation, and y is the fusion vector; , is the learnable parameter of the linear mapping on the minute side; , is the learnable parameter of the linear mapping on the second-level side; , is the learnable parameter of the gating distributor; is the aligned second-level side feature; is the aligned minute side feature; is the element-wise multiplication; g is the gating weight vector;

[0223] Step 5-2, cross-scale attention of intermittent leader: cross-scale retrieval is performed with the minute side vector as the query and the second-level time series feature as the key and value, and a prior bias generated by the intermittent weight map is introduced to highlight the fluctuation segment. The method includes the following steps:

[0224] Step 5-2-1, linearly map the minute side feature and the second-level feature to a shared attention space to form query, key and value vectors. The specific formula is:

[0225] ,

[0226] ,

[0227] ,

[0228] wherein, , is a linear mapping matrix.

[0229] Step 5-2-2, generate prior bias for suppressing intermittent segments and emphasizing fluctuation segments based on the intermittent weight map, and the specific formula is:

[0230] ,

[0231] Wherein, is the prior bias term injected into the attention score at time t;

[0232] Step 5-2-3, calculate the scaled dot product score and superimpose the prior and causal, length mask to obtain the constrained attention, and the specific formula is:

[0233] ,

[0234] Wherein: is the length mask; d is the attention space dimension; is the attention score at time t;

[0235] Step 5-2-4, normalize the score to the attention weight and weight the value vector to obtain the cross-scale weight vector, and the specific formula is:

[0236] ,

[0237] ,

[0238] Wherein: is the attention weight, is the cross-scale vector, and exp is the natural exponential function;

[0239] Step 5-3, fusion output and prediction head: vector-level fusion and deep combination of attention retrieval are carried out, and three-minute average wind speed prediction is output, and the specific formula is:

[0240] ,

[0241] ,

[0242] Wherein is a light regression head, is a fusion mapping parameter; LN is layer normalization; is the next prediction value.

[0243] In this embodiment, the data is derived from the measured data of the Jiangsu section along the Beijing-Shanghai high-speed rail. In order to test the robustness of the model proposed by the method and the effectiveness of each module, the following experiments are carried out, and the specific examples are verified as follows:

[0244] I. In order to verify the universality of the IPM-MSTCN-IGSE-DP / CNN-BILSTM wind speed prediction model proposed in the application in different environments, three high-speed railway stations with different wind conditions are selected in the experiment, namely Dashengguan Bridge Station, Yangcheng West Lake Tunnel Station and Zouwang Road Station in Jurong City. Dashengguan Bridge Station belongs to open terrain of bridge, the average wind speed is large, the wind direction is stable, and the low frequency trend is dominant, and the gust is moderate. The tunnel channel effect of Yangcheng West Lake Tunnel Station is significant, the high frequency fluctuation is strong, and the wind direction changes quickly. Zouwang Road Station in Jurong City is suburban along the road, and the surface roughness is large, the average wind speed is low but the fluctuation is significant, the local gust occurs frequently, and the wind direction is dispersed.

[0245] As shown in Figure 8 , Figure 9 , Figure 10 Under different wind conditions, the prediction curve of the method proposed in the application 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 cannot accurately reflect the sharp change and long-term trend of the wind speed, and there is obvious fluctuation distortion and delay. Especially at the moment of wind speed mutation, the predicted value and the actual wind speed have large deviation.

[0246] II. Verify the effectiveness of each module in the IPM-MSTCN-IGSE-DP / CNN-BILSTM wind speed prediction model proposed in the application:

[0247] In order to verify the effectiveness of each module of the proposed model, the prediction error statistics of MSTCN-SE (multi-scale time convolution network-channel attention), IPM-MSTCN-SE (intermittent prior module-multi-scale time convolution network-channel attention), IPM-MSTCN-IGSE (intermittent prior module-multi-scale time convolution network-intermittent weight guided channel attention), IPM-MSTCN-IGSE-DP (intermittent prior module-multi-scale time convolution network-intermittent weight guided channel attention-intermittent / fluctuation double path pooling) are compared, and the commonly used evaluation indexes: mean absolute error (Mean Absolute Error, MAE), root mean squared error (Root Mean Squared Error, RMSE) and mean absolute percentage error (Mean Absolute Percentage Error, MAPE) are used to evaluate the prediction accuracy of each model. The specific results are shown in Table 1.

[0248] Table 1 Evaluation results table

[0249]

[0250] From the results in the table, it can be seen that the IPM-MSTCN-IGSE-DP model proposed in the application has the minimum values of the MAE, RMSE and MAPE error evaluation indexes. The addition of the IPM, IG-SE and DP modules significantly improves the accuracy compared with the benchmark model, and the method proposed in the application has better prediction performance than the traditional model.

[0251] The embodiment of the application also provides an electronic device, including a memory, a processor and a computer program stored in the memory and capable of running on the processor. It should be noted that each module in the above system corresponds to the specific steps of the method provided by the embodiment of the application, has the corresponding function module and beneficial effects of executing the method. Technical details not described in detail in the embodiment can be referred to the method provided by the embodiment of the application.

[0252] The embodiment of the application also provides a computer-readable storage medium, which stores a computer program. It should be noted that each module in the above system corresponds to the specific steps of the method provided by the embodiment of the application, has the corresponding function module and beneficial effects of executing the method. Technical details not described in detail in the embodiment can be referred to the method provided by the embodiment of the application.

[0253] The application provides a high-speed rail line wind speed prediction method based on intermittent priori and cross-scale adaptive fusion, and there are many methods and approaches to specifically implement the technical scheme, and the above description is only the preferred embodiment of the application. It should be noted that for ordinary skilled persons in the art, without departing from the principle of the application, some improvements and refinements can be made, which should be regarded as the protection scope of the application. The components not explicitly described in the embodiment can be realized by the prior art.

Claims

1. A high-speed rail along the line wind speed prediction method based on intermittent prior and cross-scale adaptive fusion, characterized in that, The method comprises the following steps: Step 1, collecting and aligning the second-level wind speed and minute-level average of the monitoring station along the line, and organizing the model input in a sliding window; Step 2, constructing an intermittent prior module: calculating the prior characteristics of the second-level wind speed under the causal constraint, and adaptively generating an intermittent weight map and a soft walker sequence by a small multi-layer perception network, for quantifying the intermittency and duration that remain unchanged in continuous time steps; Step 3, extracting second-level features by using a second-level encoder, wherein the second-level encoder adopts a multi-scale time convolution network MSTCN, re-labels the channel weight under the guidance of intermittent channel attention, and aggregates intermittent and fluctuation time-varying features through intermittent and fluctuation double-path pooling respectively, to form a second-level aggregation representation; Step 4, encoding the minute-level historical average by using a minute-level encoder to obtain a minute trend representation; the minute-level encoder comprises a convolutional neural network and a bidirectional long short-term memory network CNN-BILSTM; Step 5, adaptively fusing the minute trend and the second-level aggregation representation through a fusion layer containing intermittent guidance cross-scale attention to form a fusion representation and send it to a prediction head, output a prediction result and evaluate it to verify the feasibility of the model.

2. The method of claim 1, wherein, Step 1 comprises the following steps: Step 1-1, set the first-level wind speed observation sequence wherein denotes the wind speed observation value at the t-th second, t takes the value of 1~n, and n denotes the total number of wind speed observation points; Step 1-2, taking a subsequence with a length of w from the cth wind speed observation point, and the specific formula is: ， wherein represents a sliding window, c satisfies ; Step 1-3, calculating the three-minute average in each window, and aligning the timestamp to the end of the window, and the specific formula is: ， ， wherein, denotes the 3-minute sliding mean with the c-th wind speed observation point as the starting point as the minute-level encoder input; is the mean the corresponding time label; denotes the time label of the last observation point of the c-th sliding window; Step 1-4, same window as input segment to the second level encoder, denoted as: ， wherein denotes the c-th sliding window.

3. The method of claim 2, wherein, Step 2 comprises the following steps: Step 2-1, calculating the first-order difference and the robust amplitude: ， ， wherein, is a first order difference, is a robust amplitude; is a smoothing constant; Step 2-2, estimating the local motion intensity and the strength of the short-time change of the relative trend by using the exponential weighted moving average EMA without using future information, and the specific formula is: ， ， ， The exponential weighted moving average EMA is defined as: ， wherein, is a smoothing coefficient; is a local motion intensity at time t; is a trend baseline; is a short-term change intensity of the relative trend; is an exponentially weighted moving average operator for calculating the local motion intensity; is an exponentially weighted moving average operator for calculating the short-term trend baseline; is an exponentially weighted moving average operator for calculating the short-term change intensity of the relative trend; is an exponentially weighted moving average operator; Step 2-3, calculating the mean and variance by using the exponential weighted moving average EMA based on the historical window to obtain the normalized deviation, which is used to measure the significance of the current wind speed relative to the historical level, and the specific formula is: ， ， ， wherein, is a smoothing coefficient for the causal mean and variance; is an exponentially weighted moving average of the tth second; is a causal variance estimate; is a normalized deviation; is an exponentially weighted moving average operator of the arithmetic mean of the computation window; is the t-1th second wind speed observation; is a stabilizing term; Step 2-4, concatenating the causal features by time steps to form a prior feature vector, and concatenating the intermittent prior of the previous step to enhance the time continuity, and the specific formula is: ， wherein, is the prior feature vector, is the previous step gap weight, is the previous step soft run; Step 2-5, using a small multi-layer perception network MLP to perform nonlinear mapping on the prior feature to output intermittent discriminant feature quantities, and controlling the temperature parameter of the thresholding softness and hardness, and the specific formula is: ， ， wherein, is an intermittent discriminant feature; is an unconstrained temperature parameter; is a soft threshold parameter with lower bound added; is a minimum value of the soft threshold parameter, , is a positive normalized smooth activation function; LN is a layer normalization. Step 2-6, map the intermittent discriminant feature quantity and soft threshold parameter to The continuous weight of the interval forms an intermittent weight map, and the specific formula is: ， wherein, is a non-linear activation function; is an intermittent weight; Step 2-7, using a differentiable recursive estimation to estimate the duration of the current intermittent state, and improving the stability of the segment boundary, and the specific formula is: ， wherein the soft run length has an initial value of 0, is a bleed-off coefficient; is an estimate of the duration of time that has been in the intermittent state up to t-1 seconds.

4. The method of claim 3, wherein, Step 3 comprises the following steps: Step 3-1, using a multi-scale time convolution network MSTCN to extract features from the second-level window input with a length of w to simultaneously cover short, medium and long time receptive fields, and the specific process comprises: Step 3-1-1, assuming the current second-level window input is S, construct B parallel one-dimensional convolution branches, the b-th branch uses a convolution kernel with length and the expansion coefficient to form different receptive fields; when fill with the window starting value or zero, without using future samples, the specific formula is: ， wherein, ; is an index of a branch channel; is an output value of the channel at time t for the b-th branch; is a convolution weight of the b-th branch for the j-th output channel at a lag j steps; denotes a number of lag steps skipped by a hole rate; is an input looking back lag steps forward from the current time t; is a bias term for the b-th branch for the j-th output channel. Step 3-1-2, concatenating the outputs of each branch along the channel dimension, and compressing and shaping the concatenated features to obtain unified time sequence features, and the specific formula is: ， wherein, is the output feature map of the b-th branch; is a channel-wise concatenation operator, concatenating the output feature maps of the B branches; is a channel-mixed convolution; H is the intermediate feature map after channel mixing. Step 3-1-3, projecting the input S to make its channel number match H, and adding H to do layer normalization, and the specific formula is: ， wherein is a one-dimensional convolutional residual projection on input S; 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, calculating the normalization coefficient of the intermittent and fluctuation coverage, and the formula is: ， ， wherein ; is the normalization denominator for the intermittent coverage; is the normalization denominator for the fluctuating coverage; Step 3-2-2, weighted compression is performed on each channel in the overall, intermittent and fluctuation three perspectives respectively, and the formula is: ， ， ， wherein, is the overall response of the channel, emphasizes intermittent segments, emphasizes fluctuating segments; is the output feature map of channel c at time t; Step 3-2-3, the three types of statistics and the global statistics of the intermittent prior are spliced, and the channel weight is generated by a small multi-layer perception MLP to complete the importance estimation of the channel, and the formula is: ， ， ， wherein, is the channel weight, , and are the mean and variance of the intermittent weight, respectively, small multi-layer perceptron (MLP) output compressed by Sigmoid to the channel weight ; Step 3-2-4, the feature map is multiplied by the re-labeling, so that the fusion and prediction layer learn different channel importance on the intermittent and fluctuation segments, and the formula is: ， wherein, is the feature map after channel re-scaling; Step 3-3, the intermittent and fluctuation time-varying features are aggregated by the intermittent and fluctuation two-way pooling respectively, and the second-level aggregation representation is formed for subsequent fusion and prediction.

5. The method of claim 4, wherein, In step 3-3, the following calculation formula is used: ， ， ， wherein, is the channel vector for time step t; C is the number of channels; is the weighted average over the emphasized pause segments; is the weighted average over the emphasized fluctuation segments; is the second-level side output, denotes a real-valued vector of length .

6. The method of claim 5, wherein, Step 4 includes the following steps: Step 4-1, under the premise of aligning at the end of the window, a three-minute mean history window is constructed and causally standardized, and the specific formula is: ， ， ， ， wherein, is a historical window of length K ending at is a historical window of length K ending at is the k-th three-minute average; are the mean and standard deviation, respectively, of the computed averages; is the normalized three-minute average; Step 4-2, one-dimensional convolution is used to extract local time patterns and multi-scale neighborhood information from the window, and the convolution representation is obtained by splicing in the channel dimension, and the specific formula is: ， ， ， wherein, , , is the convolution output of the rth branch at time step i; is the concatenated channel vector; U is the convolution representation sequence; is the value of the three-minute mean at index i-q after causal normalization; q is the position index within the kernel; is the convolution weight of the rth branch at kernel position q; is the bias of the rth branch; Step 4-3, the convolution representation is input into the bidirectional long short-term memory network to model the long and short term dependencies within the window, and the bidirectional hidden state sequence is obtained, and the specific formula is: ， ， wherein, are the forward and backward hidden states of the BiLSTM, respectively; is a bidirectional long short-term memory network; is a concatenation matrix; Step 4-4, the hidden state sequence is aggregated into a fixed-length minute sequence feature by using attention weighting and statistical pooling, so as to be used for subsequent fusion and prediction, and the calculation formula is: ， ， where: is the attention weight, is the minute sequence feature, T denotes transpose, is the hyperbolic tangent activation function; e is a learnable weight vector for attention scoring.

7. The method of claim 6, wherein, Step 5 includes the following steps: Step 5-1, the minute sequence feature on the minute side and the second-level aggregation representation on the second side are aligned to the same feature space, and are adaptively combined through gating, and the specific formula is: ， ， ， ， wherein, is a second-level aggregation representation, y is a fusion vector; , is a learnable parameter for the linear mapping on the minute side; , is a learnable parameter for the linear mapping on the second side; , is a learnable parameter for the gating distributor; is the aligned second-side feature; is the aligned minute-side feature; is an element-wise multiplication; g is a gating weight vector; Step 5-2, cross-scale attention of the intermittent guide: taking the minute side vector as the query, the second-level time sequence feature as the key and value for cross-scale retrieval, and introducing the prior bias generated by the intermittent weight map to highlight the fluctuation segment, which includes the following steps: Step 5-2-1, linearly map the minute-level features and the second-level features to a shared attention space to form a query vector Q, a key vector and a value vector , specifically as follows: ， ， ， wherein , is a linear mapping matrix; Step 5-2-2, based on the intermittent weight map, the prior bias that suppresses the intermittent segment and emphasizes the fluctuation segment is generated, and the specific formula is: ； wherein, is a prior bias term injected into the attention score at time t; Step 5-2-3, the scaled dot product score is calculated and the prior and causal, length mask are superimposed to obtain the attention with constraints, and the specific formula is: ， wherein is a length mask; d is the attention spatial dimension; is the attention score at time t; Step 5-2-4, the score is normalized to the attention weight, and the value vector is weighted and aggregated to obtain the cross-scale weight vector, and the specific formula is: ， ， where: are attention weights, z is a cross-scale vector, and 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 the three-minute mean wind speed prediction.

8. The method of claim 7, wherein, In step 5-3, the following calculation formula is used: ， ， wherein is a light weight regression head, is a fused mapping parameter; is a next step prediction value.

9. An electronic device, comprising: The processor and the memory are included, and the memory stores program code, when the program code is executed by the processor, the processor executes the steps of the method in any one of claims 1 to 8.

10. A storage medium, characterized by The computer program or instructions are stored, when the computer program or instructions are run on the computer, the steps of the method in any one of claims 1 to 8 are executed.

Images