High-Speed Railway Alongside Strong Wind Prediction Algorithm Based on Multi-Source Heterogeneous Data Fusion

Abstract

The present invention discloses a high-speed railway along-wind prediction algorithm based on multi-source heterogeneous data fusion. First, wind speed data and meteorological element data are acquired for data preprocessing to construct a multivariate multi-time-span wind speed prediction sample set. Then, a multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion to be trained is constructed. After training the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion, with the detected meteorological element data as the input and the predicted wind speed at the target wind speed monitoring points along the railway as the output for prediction. This method uses multi-time-span data as the input of the model, establishes a multi-branch TCN-BiLstm attention prediction network, uses a temporal convolutional network, a bidirectional long short-term memory network, and an efficient channel attention network as the backbone network of each branch module, and designs a feature fusion module, which can comprehensively capture the patterns and features of wind speed data at different time scales, and improve the robustness and stability of the prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent meteorology, and particularly relates to a high-speed rail along-line gale prediction algorithm based on multi-source heterogeneous data fusion. Background Art

[0002] Gale is the biggest factor affecting the safe operation of high-speed rail. It not only affects the riding comfort, but may also cause the train to be unable to operate normally, seriously affecting the train operation safety. The existing high-speed rail disaster prevention system already has the function of real-time gale monitoring and alarm, but it does not have the function of gale prediction and early warning. Therefore, incidents often occur that the train has entered the gale area when the alarm is issued, or the train enters the gale area at an excessive speed due to insufficient deceleration time window. If the system can predict the wind speed value in the future for a period of time through the wind speed data at the current moment, and then give an early warning of the gale, it can enable the dispatcher and the driver to make emergency disposal preparations in advance, and better ensure the safe operation of high-speed rail.

[0003] In recent years, the research on railway gale early warning has mainly been based on the gale data monitored by a single railway, and methods such as time series have been used to carry out wind speed prediction. At present, scholars have carried out railway wind speed prediction methods such as the ARIMA prediction model based on time series, Kalman filter, machine learning, and artificial neural network around railway wind speed prediction; however, the wind speed time series along the railway shows highly nonlinear and strongly random characteristics. The models based on ARIMA and Kalman filter are relatively simple, but the prediction results have general accuracy. Machine learning and artificial neural network methods are good at summarizing the connections and characteristics between data, and thus predicting based on historical data, which are suitable for short-term wind speed prediction. However, the wind speed along the high-speed rail has the characteristics of multi-factor interaction and complex nonlinearity. A single neural network model cannot deeply mine the effective information between the wind speed along the high-speed rail and various meteorological elements, and obtain more accurate predictions. In the existing research, the influence of meteorological factors is insufficiently considered in the prediction model, and the characteristic correlation between different meteorological factors is ignored, which will lose a large amount of available information and lead to a decrease in prediction accuracy.

[0004] The present invention is based on the following research results: Gale disasters are one of the most serious meteorological disasters endangering the safe operation of high-speed trains. If the instantaneous wind speed exceeds the limit wind speed, only 1-2 seconds is enough to cause the train to capsize, which puts higher requirements on the accuracy and strategy of high-speed rail wind speed prediction.

[0005] At present, the following problems need to be solved in the work of predicting the wind speed along the high-speed rail:

[0006] Problem 1: The current research on wind speed prediction along high-speed rail lines only uses historical wind speed data from a single station, without considering the coupling relationship between multiple factors such as temperature, humidity, and air pressure and wind speed. At the same time, different time intervals contain different information. A single short time interval cannot capture the trend in the data, and a single long time interval cannot accurately grasp the strong volatility and detailed characteristics in a short period. The existing wind speed prediction models along high-speed rail lines only use a single time interval as the input of the model, without considering the impact of multiple time intervals on the prediction results.

[0007] Problem 2: The encoder structure adopted in the current research on wind speed prediction along high-speed rail lines is relatively simple, only using a single backbone network. After introducing multiple meteorological elements and multiple time interval information, the single network's information feature extraction ability is insufficient, and the collaborative information mining is not sufficient. The traditional decoder structure is difficult to make full use of the associations and interactions between different elements, resulting in the model being unable to accurately capture the complex dependence relationships and non-linear associations between meteorological elements, thus affecting the prediction accuracy.

[0008] Problem 3: In the prediction, each meteorological element input has a synergistic effect on the wind speed. The synergistic information and wind speed information obtained from the existing multiple modules and multiple time interval levels cannot be effectively fused, which will lead to the loss of some feature information and the inability to fully explore the potential associations between data, thus limiting the model's comprehensive understanding and prediction ability of the data. Summary of the Invention

[0009] In view of the problems existing in the prior art, the present invention adopts the following technical solutions: The present invention designs a high-speed rail line gale prediction algorithm based on multi-source heterogeneous data fusion, specifically including the following steps

[0010] S1, for the wind speed monitoring point area along the railway to be predicted, at each historical time point within a preset time interval within a preset historical range, obtain the wind speed data of the target wind speed monitoring point in the area, and the meteorological element data of each meteorological station within a preset distance from the target wind speed monitoring point, and perform data preprocessing on the wind speed data and meteorological element data to construct a multi-source multi-time interval wind speed prediction sample set. The multi-source multi-time interval wind speed prediction sample set includes the detected meteorological element data and the detected wind speed of the target wind speed monitoring point along the railway. Among them, the meteorological element data includes wind speed, wind direction, air pressure, temperature, humidity, and rainfall.

[0011] Among them, the data preprocessing of the wind speed data and meteorological element data includes obtaining the average value of the wind speed data of the target wind speed monitoring points along the railway at each historical time point within a preset historical range and a preset time interval, selecting the median to process outliers, selecting the meteorological element data with a large correlation with the railway wind speed monitoring points, and then selecting ridge regression to fill in the missing data. S2, construct a multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion for training; the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion includes an encoder and a decoder connected in sequence. The encoder includes a multi-branch feature extraction network and a feature fusion module connected in sequence; in the encoder, a multi-branch feature extraction network is constructed. The multi-branch feature extraction network includes three feature extraction branches connected in sequence for extracting features at different levels. Among them, each feature extraction branch respectively includes a TCN (Time Convolutional Network), a BiLstm (Bidirectional Long Short-Term Memory Network), and an ECANet (Efficient Channel Attention Network) connected in sequence; the input of the first feature extraction branch is a multi-source multi-time-interval wind speed prediction sample set, the input of each subsequent feature extraction branch is the output of the previous feature extraction branch, and the output of each feature extraction branch is the extracted feature vector;

[0012] The input of the feature fusion module is the feature vectors extracted by the three feature extraction branches. The feature fusion module includes a DO-Conv (Depthwise Parameterized Convolution) with a 1×1 convolution kernel, a batch normalization layer, a GELU (Gaussian Error Linear Unit) non-linear activation function, and an ECANet (Efficient Channel Attention Network) connected in sequence;

[0013] The input of the decoder is the feature vector obtained after feature fusion. The decoder includes a 1×1 convolution layer, an ECANet (Efficient Channel Attention Network), and a BiLstm (Bidirectional Long Short-Term Memory Network), and obtains and outputs the final prediction vector.

[0014] Among them, in the feature fusion module, for the feature vectors extracted by the three input feature extraction branches, the following operations are performed:

[0015] First, the feature vectors obtained by each feature extraction branch are respectively input into a DO-Conv (Depthwise Parameterized Convolution) with a 1×1 convolution kernel to obtain the feature vectors of the high-level feature representations of two branches; secondly, the feature vectors of the high-level feature representations of each branch are respectively input into a batch normalization layer and a GELU (Gaussian Error Linear Unit) non-linear activation function to obtain the feature vectors of the two branches after feature extraction and processing; then, the feature vectors of the two branches after feature extraction and processing are combined to obtain a combined feature vector; the combined feature vector is successively input into an ECANet (Efficient Channel Attention Network) and a DO-Conv (Depthwise Parameterized Convolution) with a 1×1 convolution kernel to obtain the feature vector output by the feature fusion module. The calculation process is as follows:

[0016] X1 = G(BN(DOConv 1×1 (X in1 )))

[0017] X2 = G(BN(DOConv 1×1 (X in2 )))

[0018] X3 = G(BN(DOConv 1×1 (X in3 )))

[0019] W = Concat(X1, X2, X3)

[0020] Y out = G(BN(DOConv 3×1 (Trans(W))))

[0021] Among them, X in1 , X in2 and X in3 are the outputs after the multi-branch feature extraction of the multi-variable multi-time-step wind speed prediction sample set. The input data X1, X2, and X3 are the outputs after passing through the multi-branch feature extraction network and serve as the inputs to the feature fusion module. The network has a data dimension of 128; Y out represents the feature vector output by the feature fusion module. DOConv n×m represents the convolution process using an n×m convolution kernel. Concat(·) represents the concatenation operation based on the channel dimension. Trans(·) represents the process of passing through the efficient channel attention network. BN(·) and G(·) represent the batch normalization and GELU non-linear activation functions respectively.

[0022] In the decoder, the input to the decoder is the feature vector Y out output by the feature fusion module. First, a 1×1 convolutional layer is used to modify the number of input channels. Then, the ECANet efficient channel attention network is used to establish long-term feature relationships in the feature data, and important feature representations are extracted to enhance the input features. The enhanced input features are used as the input to the BiLSTM bidirectional long short-term memory network. The calculation process is as follows:

[0023] W = Conv 1×1 (Y out )

[0024] T = Trans(W)

[0025] L = BiLSTM(T)

[0026]

[0027] Among them, Y out is the output of the feature fusion module, and Y out is used as the input of the decoder. Conv 1×1 (·) represents the convolution process using a 1×1 convolution kernel. Trans(·) and BiLSTM(·) respectively represent the processes of passing through the ECANet efficient channel attention network and the BiLstm bidirectional long short-term memory network. Concat(·) represents the splicing operation based on the channel dimension. is the final prediction vector.

[0028] S3. Based on the multi-source multi-time-span wind speed prediction sample set, using the detected wind speed at the target wind speed monitoring point along the railway and the corresponding detected meteorological element data as the input, and the detected wind speed at the target wind speed monitoring point with a preset time interval as the output, train the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion to obtain the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion.

[0029] The model learning and training process includes the following steps:

[0030] S301. Use the multi-source multi-time-span wind speed prediction sample set obtained in S1 as the input of the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion. Among them, 80% of the multi-source multi-time-span wind speed prediction sample set is used for training, and 20% is used for testing.

[0031] S302. Select the cross-entropy as the maximum loss function, select the Adam optimizer to update the learning rate of the model. The learning rate is [1e-4, 1e-3], the exponential decay rate coefficient selected by the model is 0.95, and the number of iterations is set to 200 times for single-step prediction.

[0032] The evaluation indexes of the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion include MAE (Mean Absolute Error), RMSE (Root Mean Square Error), and MAPE (Mean Absolute Percentage Error), to evaluate the prediction accuracy of the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion.

[0033] S4. Use the detected meteorological element data as the input and the predicted wind speed at the target wind speed monitoring point along the railway as the output, and use the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion for prediction.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. Use the historical wind speed of the target site and multi - meteorological element variables of surrounding meteorological stations as common input data, comprehensively considering the influence of multiple factors on wind speed. At the same time, use multi - time - interval data as the input of the model to comprehensively capture the patterns and characteristics of wind speed data at different time scales, improving the robustness and stability of the prediction.

[0036] 2. Establish a multi - branch TCN - BiLstm attention prediction network, using the Temporal Convolutional Network (TCN), Bidirectional Long Short - Term Memory Network (BiLstm), and Efficient Channel Attention Network (ECANet) based on the Transformer variant as the backbone network's shunt modules. The Temporal Convolutional Network (TCN) can effectively analyze the feature associations existing between data; the Bidirectional Long Short - Term Memory Network (BiLstm) can capture long - term dependencies in the sequence; the Efficient Channel Attention Network (ECANet) can enhance the ability of feature representation and selectively focus on the features most important for the wind speed prediction task, enabling the network to better distinguish and utilize the feature information of different channels.

[0037] 3. Design a feature fusion module that can explore the potential associations and interactions between different features, weight and select different features, make adaptive adjustments according to their contribution degrees to the prediction task, improve the attention to important features, and reduce the influence of unimportant features, thereby improving the accuracy and robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] To more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required to be used in the embodiments of the present application. Obviously, the following described drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0039] Figure 1 It is the specific implementation flowchart of the present invention;

[0040] Figure 2 It is the multi - branch convolutional attention prediction network of the present invention;

[0041] Figure 3 It is the fusion module designed by the present invention;

[0042] Figure 4 It is the decoder module designed by the present invention;

[0043] Figure 5 It is the prediction model designed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] The features and exemplary embodiments of various aspects of the present application will be described in detail below. In the following detailed description, many specific details are set forth in order to provide a comprehensive understanding of the present application. However, it will be apparent to those skilled in the art that some of these specific details may not be required for the implementation of the embodiments of the present application. The following description of the embodiments is only intended to provide a better understanding of the present application by showing examples of the present application.

[0045] In the description of the present application, it should be understood that with regard to the orientation description, such as the orientation or positional relationship indicated by up, down, front, back, left, right, etc., is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the embodiments of the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be understood as a limitation on the embodiments of the present application.

[0046] In the description of the embodiments of the present application, the meaning of several is one or more, the meaning of multiple is two or more, greater than, less than, exceeding, etc. are understood as not including the number itself, and above, below, within, etc. are understood as including the number itself. If the first and second are described only for the purpose of distinguishing technical features, they cannot be understood as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features or implicitly indicating the sequence relationship of the indicated technical features.

[0047] In the description of the embodiments of the present application, unless otherwise clearly defined, words such as setting, installing, connecting, etc. should be understood in a broad sense, and those skilled in the art can reasonably determine the specific meanings of the above words in the present application in combination with the specific content of the technical solution.

[0048] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The embodiments will be described in detail below with reference to the drawings.

[0049] As Figure 1 shown, the present invention designs a high-speed rail along-line strong wind prediction algorithm based on multi-source heterogeneous data fusion, which specifically includes the following steps

[0050] S1. For the area of the wind speed monitoring points along the railway to be predicted, at each historical time point within a preset time interval within a preset historical range, obtain the wind speed data of the target wind speed monitoring points in the area, as well as the meteorological element data of each meteorological station within a preset distance from the target wind speed monitoring points, and perform data preprocessing on the wind speed data and the meteorological element data to construct a multi-source multi-time-interval wind speed prediction sample set. The multi-source multi-time-interval wind speed prediction sample set includes the detected meteorological element data and the detected wind speed of the target wind speed monitoring points along the railway corresponding thereto. Among them, the meteorological element data includes wind speed, wind direction, air pressure, temperature, humidity, and rainfall.

[0051] In one of the embodiments, the wind speed monitoring data measured by the anemometer at a height of 4 m above the ground at the K1066 wind speed monitoring station along the high-speed railway in Jiangsu Province from January 1, 2018 to December 31, 2018 is selected. The time resolution is at the second level and includes the strong wind monitoring data. To meet the task requirement of predicting the wind speed at the 5-min moment in the present invention, the second-level wind speed data of the K1066 wind speed monitoring station is aggregated into an average value every 5 min. The purpose of doing this is to obtain more stable and smooth wind speed data, and at the same time reduce the data noise and fluctuations brought by the time resolution. According to the analysis, the wind monitoring data of the meteorological stations within 20 km along the railway is representative and correlated with the railway monitoring data. Therefore, the present invention selects the monitoring data of the meteorological station about 9.31 km away from the K1066 wind speed monitoring station. These data include the information of meteorological elements such as wind speed, wind direction, air pressure, temperature, humidity, rainfall, etc. during the period from January 1, 2018 to December 31, 2018, and the time resolution is 5 min. At the same time, in order to ensure the integrity and authenticity of the data, the present invention selects the median to process the outliers. The median can better reflect the overall trend and changes of the data set and avoid the information loss that may be caused by directly deleting the outliers. At the same time, ridge regression is selected to fill in the missing data.

[0052] The influence degrees of different meteorological information on the wind speed prediction results are not the same. Therefore, by constructing variables highly correlated with the target variable as the data set, the prediction results can be made more accurate and reliable. In the present invention, the Spearman correlation coefficient method is used to calculate the correlation coefficients between the actual wind speed and each meteorological element of the meteorological station. The historical wind speed of the K1066 wind speed monitoring station, the wind speed of the meteorological station, the wind direction of the meteorological station, the air pressure of the meteorological station, the temperature of the meteorological station, and the humidity of the meteorological station are selected to construct a multi-input data set, with a total of 105,120 pieces of data. At the same time, the time-distance correlation curve is used as the basis for selecting the time distance. The peak height of the correlation curve reflects the correlation strength between the corresponding time distance and the target variable. A higher peak indicates a stronger correlation between the time distance and the target variable, while a lower peak indicates a weaker correlation. The present invention selects two time distances of 45 min and 225 min as the multi-time-distance input of the model.

[0053] S2. Construct a multi-branch TCN-BiLstm attention prediction network model for training based on multi-source heterogeneous data fusion; the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion includes an encoder and a decoder connected in sequence, and the encoder includes a multi-branch feature extraction network and a feature fusion module connected in sequence; in the encoder, construct a multi-branch feature extraction network, and the multi-branch feature extraction network includes three feature extraction branches connected in sequence for extracting features at different levels. Among them, each feature extraction branch respectively includes a TCN (Temporal Convolutional Network), a BiLstm (Bidirectional Long Short-Term Memory Network), and an ECANet (Efficient Channel Attention Network) connected in sequence; the input of the first feature extraction branch is a multi-source multi-time-span wind speed prediction sample set, the input of each subsequent feature extraction branch is the output of the previous feature extraction branch, and the output of each feature extraction branch is the extracted feature vector;

[0054] The input of the feature fusion module is the feature vectors extracted by the three feature extraction branches. The feature fusion module includes a DO-Conv (Depthwise Parameterized Convolution) with a 1×1 convolutional kernel, a batch normalization layer, a GELU (Gaussian Error Linear Unit) non-linear activation function, and an ECANet (Efficient Channel Attention Network) connected in sequence;

[0055] The input of the decoder is the feature vector obtained after feature fusion. The decoder includes a 1×1 convolutional layer, an ECANet (Efficient Channel Attention Network), and a BiLstm (Bidirectional Long Short-Term Memory Network), to obtain the final prediction vector and output it.

[0056] In one of the embodiments, the entire network adopts an encoder-decoder structure, and the final prediction accuracy depends on the accuracy of feature information extraction. Existing research has shown that convolutional neural networks are good at extracting local feature information, but lack accuracy in grasping global information; recurrent neural networks establish long-term dependencies in time series data through recurrent connections and are suitable for processing sequence tasks. However, in long sequence prediction tasks, recurrent neural networks may be interfered with in terms of information transmission and gradients; the Transformer multi-head attention mechanism can focus on global information and also closely focus on important regions, capable of making up for the shortcomings of convolution and handling long sequence tasks well. Therefore, the prediction network encoder of the present invention adopts a multi-branch mode. Wind speed data often contains features at multiple levels. Therefore, three stacked convolutional layers are constructed to extract features at different levels, and the input of the subsequent convolutional layer is the output of the previous convolutional layer. By stacking multiple convolutional layers, the model can gradually learn more abstract and high-level non-linear feature representations, thus better capturing the complex relationships and patterns in wind speed data. At the same time, each convolutional layer can also gradually capture a larger range of context information. A recurrent neural network and a Transformer are connected in series to each branch. The recurrent neural network is used to capture the long-term dependencies of the sequence, and then the Transformer module is used to extract the global features of the sequence, selectively focusing on the features that are most important for the wind speed prediction task. Then, the features of multiple branches are put into a fusion module to obtain an effective fusion.

[0057] Among them, the convolutional module uses a temporal convolutional network (TCN). Temporal convolution (TCN) is an improvement on traditional convolution and is more efficient in dealing with time series problems. TCN can fully consider the time domain features of wind speed and meteorological element time series, effectively analyze the correlations existing in the data, and has more stable gradients and higher computational efficiency.

[0058] Among them, the recurrent neural network selects a bidirectional long short-term memory network (BiLstm). Wind speed data usually has obvious temporal dependence relationships. Compared with the traditional long short-term memory network (Lstm), the bidirectional long short-term memory network (BiLstm) can extract the non-linear features of wind speed from different directions. Through the combination of forward and backward Lstms, it can capture the long-term dependencies in the sequence and has better feature extraction ability.

[0059] Among them, the Transformer module selects the Efficient Channel Attention Network (ECANet) based on the Transformer variant. ECANet avoids dimensionality reduction and realizes cross-channel interaction on the basis of the SENet network, which can enhance the feature representation ability. At the same time, it can learn the importance of each channel, enabling the network to better distinguish and utilize the feature information of different channels, selectively focus on the features more important for wind speed, suppress unimportant channels, and reduce the response to noise and redundant information.

[0060] Such as Figure 3As shown, in one of the embodiments, in the feature fusion module, due to the strong volatility and randomness of the wind speed itself, feature extraction is a difficult task. Since there is a large amount of useless information in the collaborative information and wind speed information obtained from multiple modules and multi-time interval levels of the model, it is particularly important to filter out this information. If the obtained feature information cannot be fully integrated, the noise contained therein will have a great impact on the final prediction result. Therefore, the present invention innovatively adds a fusion module to the prediction network to fuse information at different levels. In the backbone network, the feature information extracted by each branch needs to establish a complementary relationship, and the information between features is guided through the fusion module to filter out the feature information that is more meaningful for wind speed. Feature fusion can effectively fuse the collaborative information and wind speed information obtained from multiple modules and multi-time interval levels, and fully explore the potential associations between data. In the fusion module, depthwise separable convolution (DO-Conv) is used instead of traditional convolution. Traditional convolution uses fixed weight parameters in the convolution kernel, while depthwise separable convolution introduces additional parameters, so that each input channel and output channel has its own weight parameters; and parameter sharing can be performed, but different channels can have different weight parameters, which has a stronger feature learning ability than traditional convolution. The fusion module is divided into two parts. First, the feature vectors obtained by each feature extraction branch are passed through a DO-Conv depthwise separable convolution with a 1×1 convolution kernel to filter the feature information and enhance the feature extraction ability; secondly, it passes through a batch normalization layer and a GELU activation function. The GELU activation function is used instead of the traditional ReLU. Since the idea of random regularization is added in GELU, the accuracy of the network is improved. Then, the information extracted by the two branches is combined. After combination, it passes through an efficient channel attention network (ECANet) and a DO-Conv depthwise separable convolution with a 1×1 convolution kernel in sequence. Among them, the ECANet network can adaptively learn the correlation between features and weight the features according to the importance of the correlation, which helps to capture the long-range dependence relationship between the input features and extract more accurate and useful feature representations; after passing through the depthwise separable convolution with a 1×1 convolution kernel, the feature dimension can be changed to extract more useful feature representations. The calculation process is as follows:

[0061] X1 = G(BN(DOConv 1×1 (X in1 )))

[0062] X2 = G(BN(DOConv 1×1 (X in2 )))

[0063] X3 = G(BN(DOConv 1×1 (X in3 )))

[0064] W = Concat(X1, X2, X3)

[0065] Y out = G(BN(DOConv 3×1 (Trans(W))))

[0066] Among them, X in1 , X in2 and X in3 are the outputs after the multi-branch feature extraction of the multi-source multi-time-span wind speed prediction sample set. The input data X1, X2, and X3 are the outputs after passing through the multi-branch feature extraction network and serve as the inputs to the feature fusion module. The data dimension of the network is 128; Y out represents the feature vector output by the feature fusion module. DOConv n×m represents the convolution process using an n×m convolution kernel. Concat(·) represents the concatenation operation based on the channel dimension. Trans(·) represents the process of passing through the efficient channel attention network. BN(·) and G(·) represent the batch normalization and GELU non-linear activation functions respectively.

[0067] As Figure 4 shown, most of the existing decoders use a single convolutional neural network or recurrent neural network to decode the feature data and form the prediction data, which causes the model to lose some important information and makes it challenging to recover the data details. In the present invention, a new decoder module is proposed to solve the above problems. In the decoding module, first, a 1×1 convolutional layer is used to modify the number of input channels, and then the Transformer module is used to establish long-term feature relationships in the feature data; the efficient channel attention network (ECANet) can extract important feature representations and enhance the input features, and use them as the input to the bidirectional long short-term memory network (BiLSTM). BiLSTM can consider the previous prediction results and context information to adjust and correct the prediction, and gradually generate the wind speed prediction sequence for future time steps, thereby improving the overall prediction accuracy. The calculation process is as follows:

[0068] W = Conv 1×1 (Y out )

[0069] T = Trans(W)

[0070] L = BiLSTM(T)

[0071]

[0072] In one embodiment, Y out is the output of the feature fusion module, and Y outAs the input to the decoder, Conv 1×1 (·) represents the convolution process using a 1×1 convolution kernel. Trans(·) and BiLSTM(·) represent the processes through the ECANet efficient channel attention network and the BiLstm bidirectional long short-term memory network respectively. Concat(·) represents the concatenation operation based on the channel dimension. is the final prediction vector.

[0073] S3. Based on the multi-source multi-time-span wind speed prediction sample set, with the detected wind speed at the target wind speed monitoring point along the railway and the corresponding detected meteorological element data as the input, and the detected wind speed at the target wind speed monitoring point with a preset time interval as the output, train the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion to obtain the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion.

[0074] The model learning and training process includes the following steps:

[0075] S301. Use the multi-source multi-time-span wind speed prediction sample set obtained in S1 as the input to the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion. Among them, 80% of the multi-source multi-time-span wind speed prediction sample set is used for training, and 20% is used for testing.

[0076] S302. Select cross-entropy as the maximum loss function, select the Adam optimizer to update the learning rate of the model. The learning rate is [1e-4, 1e-3], the exponential decay rate coefficient selected by the model is 0.95, and the number of iterations is set to 200 times for single-step prediction.

[0077] The evaluation indicators of the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion include MAE (Mean Absolute Error), RMSE (Root Mean Square Error), and MAPE (Mean Absolute Percentage Error) to evaluate the prediction accuracy of the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion.

[0078] S4. Use the detected meteorological element data as the input and the predicted wind speed at the target wind speed monitoring point along the railway as the output, and use the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion for prediction.

Claims

1. A gale prediction algorithm for railway lines based on multi-source heterogeneous data fusion, characterized in that It includes the following steps: S1. For the wind speed monitoring point area along the railway to be predicted, at each historical time point with a preset time interval within a preset historical range, obtain the wind speed data of the target wind speed monitoring point in the area, as well as the meteorological element data of each meteorological station within a preset distance from the target wind speed monitoring point, and perform data preprocessing on the wind speed data and meteorological element data to construct a multi-source multi-time-interval wind speed prediction sample set. The multi-source multi-time-interval wind speed prediction sample set includes the detected meteorological element data and the corresponding detected wind speed of the target wind speed monitoring point along the railway. Among them, the meteorological element data includes wind speed, wind direction, air pressure, temperature, humidity, and rainfall; S2. Construct a multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion for training; The multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion includes an encoder and a decoder connected in sequence. The encoder includes a multi-branch feature extraction network and a feature fusion module connected in sequence; In the encoder, construct a multi-branch feature extraction network. The multi-branch feature extraction network includes three feature extraction branches connected in sequence for extracting features at different levels. Among them, each feature extraction branch respectively includes a TCN (Time Convolutional Network), a BiLstm (Bidirectional Long Short-Term Memory Network), and an ECANet (Efficient Channel Attention Network) connected in sequence. The input of the first feature extraction branch is the multi-source multi-time-interval wind speed prediction sample set, the input of each subsequent feature extraction branch is the output of the previous feature extraction branch, and the output of each feature extraction branch is the extracted feature vector; The input of the feature fusion module is the feature vectors extracted by the three feature extraction branches. The feature fusion module includes a DO-Conv (Depthwise Parameterized Convolution) with a 1×1 convolution kernel, a batch normalization layer, a GELU (Gaussian Error Linear Unit) non-linear activation function, and an ECANet (Efficient Channel Attention Network) connected in sequence; The input of the decoder is the feature vector obtained after feature fusion. The decoder includes a 1×1 convolutional layer, an ECANet (Efficient Channel Attention Network), and a BiLstm (Bidirectional Long Short-Term Memory Network) to obtain the final prediction vector and output it; S3. Based on the multi-source multi-time-interval wind speed prediction sample set, with the detected wind speed of the target wind speed monitoring point along the railway and the corresponding detected meteorological element data as the input, and the detected wind speed at a preset time interval from the target wind speed monitoring point as the output, train the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion for training to obtain the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion; S4. With the detected meteorological element data as the input and the predicted wind speed of the target wind speed monitoring point along the railway as the output, use the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion for prediction.

2. The railway along - line strong - wind prediction algorithm based on multi - source heterogeneous data fusion according to claim 1, characterized in that, The data preprocessing of the wind speed data and meteorological element data described in S1 includes obtaining the average value of the wind speed data of the target wind speed monitoring points along the railway at each historical time point within a preset historical range and a preset time interval, selecting the median to process outliers, selecting the meteorological element data with a large correlation with the railway wind speed monitoring points, and then selecting ridge regression to fill in the missing data.

3. The railway along - line strong - wind prediction algorithm based on multi - source heterogeneous data fusion according to claim 1, characterized in that, In the feature fusion module described in S2, for the feature vectors extracted by the three input feature extraction branches, the following operations are performed: First, the feature vectors obtained by each feature extraction branch are respectively input into a DO-Conv depth parametric convolution with a 1×1 convolution kernel to obtain the feature vectors of the high-level feature representations of the two branches; second, the feature vectors of the high-level feature representations of each branch are respectively input into a batch normalization layer and a GELU non-linear activation function to obtain the feature vectors of the two branches after feature extraction and processing; then, the feature vectors of the two branches after feature extraction and processing are combined to obtain a combined feature vector; the combined feature vector is successively input into an ECANet efficient channel attention network and a DO-Conv depth parametric convolution with a 1×1 convolution kernel to obtain the feature vector output by the feature fusion module. The calculation process is as follows: X1 = G(BN(DOConv 1×1 (X in1 ))) X2 = G(BN(DOConv 1×1 (X in2 ))) X3 = G(BN(DOConv 1×1 (X in3 ))) W = Concat(X1, X2, X3) Y out = G(BN(DOConv 3×1 (Trans(W)))) Among them, X in1 , X in2 and X in3 are the outputs after the multi-branch feature extraction of the multi-source multi-time-span wind speed prediction sample set. The input data X1, X2, and X3 are the outputs after passing through the multi-branch feature extraction network and are used as the inputs of the feature fusion module. The network has a data dimension of 128; Y out represents the feature vector output by the feature fusion module. DOConv n×m represents the convolution process using an n×m convolution kernel. Concat(·) represents the concatenation operation based on the channel dimension, Trans(·) represents the process of passing through the efficient channel attention network, and BN(·) and G(·) represent batch normalization and the GELU non-linear activation function, respectively.

4. The railway along - line strong - wind prediction algorithm based on multi - source heterogeneous data fusion according to claim 3, characterized in that, In the decoder described in S2, the input of the decoder is the feature vector Y output by the feature fusion module out , first use a 1×1 convolutional layer to modify the number of input channels, then use the ECANet efficient channel attention network to establish long-term feature relationships in the feature data, and enhance the input features by extracting important feature representations. Use the enhanced input features as the input of the BiLSTM bidirectional long short-term memory network. The calculation process is as follows: W = Conv 1×1 (Y out ) T = Trans(W) L = BiLSTM(T) Among them, Y out is the output of the feature fusion module, and Y out is used as the input of the decoder. Conv 1×1 (·) represents the convolution process using a 1×1 convolution kernel. Trans(·) and BiLSTM(·) respectively represent the processes of passing through the ECANet efficient channel attention network and the BiLstm bidirectional long short-term memory network. Concat(·) represents the splicing operation based on the channel dimension. is the final prediction vector.

5. The railway along - line strong - wind prediction algorithm based on multi - source heterogeneous data fusion according to claim 1, characterized in that, The model learning and training process described in S3 includes the following steps: S301, using the multi-source multi-time interval wind speed prediction sample set obtained in S1 as the input of the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion. Among them, 80% of the multi-source multi-time interval wind speed prediction sample set is used for training, and 20% is used for testing; S302, selecting cross-entropy as the maximum loss function, selecting the Adam optimizer to update the learning rate of the model, the learning rate is [1e-4, 1e-3], the exponential decay rate coefficient selected by the model is 0.95, and the number of iterations is set to 200 times for single-step prediction.

6. The algorithm for predicting strong winds along railway lines based on multi-source heterogeneous data fusion according to claim 1, wherein The model evaluation metrics described in S3 include MAE (Mean Absolute Error), RMSE (Root Mean Square Error), and MAPE (Mean Absolute Percentage Error), which are used to evaluate the prediction accuracy of the multi-branch TCN-BiLstm attention prediction network model based on multi-source heterogeneous data fusion.

Images