Deep learning-based blasting construction traffic prediction method and system, and storage medium

Abstract

The invention discloses a CNN-LSTM traffic flow prediction method for blasting construction of an overline overbridge group, and the method comprises the steps: firstly carrying out the space-time matrix reconstruction of upstream and downstream multi-source traffic monitoring data of a construction area, extracting the spatial correlation characteristics of road segments through a multilayer convolutional network, and employing the maximum pooling to suppress the blasting vibration noise; inputting the feature sequence into a long short-term memory network, and capturing traffic flow sudden change and periodic moisture regaining rules before, during and after blasting; and finally, traffic flow predicted values of a plurality of time steps in the future are output through a full connection layer, and high-precision fitting of the interruption moment is realized through joint training of mean square error and peak value weighted loss. Compared with ARIMA, Kalman filtering and support vector regression, the method has the advantages that the prediction precision is remarkably improved, the method can be embedded into an existing high-speed monitoring platform for real-time operation, quantitative basis is provided for dynamic shunting, traffic restriction and emergency decision making during blasting, and the method has the engineering advantages of being short in construction period, low in risk, easy to migrate and the like.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of intelligent transportation and blasting engineering, specifically to a method, system, and storage medium for predicting traffic during blasting operations based on deep learning. Background Technology

[0002] With the reconstruction and expansion of highways and the upgrading of urban three-dimensional transportation networks, the demand for demolition of existing structures such as overpasses and interchanges is increasing. Among these, blasting is the preferred solution due to its short construction period, low cost, and minimal disruption to traffic above. However, blasting not only completely interrupts traffic in the construction area, but the shock wave, dust, and subsequent debris removal operations also cause drastic fluctuations in traffic flow in upstream and downstream sections, resulting in a complex traffic evolution process of "short-term congestion—evacuation and resurgence—secondary delays." Current traffic management methods mainly rely on manual experience and static traffic restriction schemes, lacking quantitative prediction of the timing and spatial diffusion range of sudden changes in traffic flow. This leads to delayed dissemination of diversion information, excessive or insufficient detour routes, an average increase in delays of 20%–40% in construction areas, and even induces regional congestion and secondary accidents.

[0003] In the field of traffic flow prediction, traditional methods include ARIMA, Kalman filtering, and support vector regression. ARIMA uses autoregression and moving averages of historical time series to characterize linear trends, and it performs well in stable conditions. However, the step-like decrease in road capacity caused by blasting construction is highly nonlinear, and ARIMA cannot capture abrupt changes, often resulting in prediction errors exceeding 30%. Kalman filtering, based on the linear Gaussian assumption, uses recursive state estimation to optimally filter observation noise. Although it has real-time update capabilities, it is prone to over-smoothing when faced with large step changes in traffic flow in construction areas, leading to an underestimation of peak flow by 15%–25%. Support vector regression uses kernel functions to map data to a high-dimensional space to achieve nonlinear fitting. However, as the sample size increases, the computational complexity of the kernel matrix increases exponentially, and hyperparameters such as the penalty coefficient and kernel width are extremely sensitive to blasting scenarios. Cross-validation parameter tuning is time-consuming, making it difficult to meet the "hour-level" rapid deployment requirements before construction.

[0004] In recent years, deep learning methods have been introduced into traffic prediction. Pure convolutional neural networks (CNNs) excel at extracting spatial adjacency relationships but lack temporal memory; simple recurrent neural networks (RNNs) are prone to gradient vanishing and are insufficient at learning long-sequence mutations. Although CNN-LSTM joint models have been used for ordinary urban roads, they still have the following shortcomings in the transient scenario of "blasting-interruption-reopening": 1) The input is mostly a regular grid or a single road segment, making it difficult to represent the topological heterogeneity of highways with multiple ramps and lanes; 2) The network depth and width are fixed, lacking an adaptive denoising mechanism to deal with detector drift and packet loss caused by blasting vibrations; 3) The loss function only considers the overall error and does not sufficiently penalize the short-term peak predictions before and after the interruption, resulting in prediction results deviating from reality by more than 10% during critical periods. Therefore, existing technologies cannot provide accurate, real-time, and engineering-applicable traffic flow prediction methods under the high impact, high risk, and strong nonlinearity conditions of blasting construction, and a new method specifically designed for this scenario is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and storage medium for traffic prediction during blasting construction based on deep learning, in order to solve the problem mentioned in the background art that the existing technology cannot provide accurate, real-time, and engineering-applicable traffic flow prediction methods under the high impact, high risk, and strong nonlinear conditions of blasting construction.

[0006] To achieve the above objectives, the present invention provides the following technical solution: Deep learning-based traffic prediction methods for blasting construction include: Step S1: Multi-source heterogeneous data acquisition: Collect traffic flow time-series data, road network topology data, real-time construction status data, and environmental data for the target area; Step S2: Spatiotemporal graph construction and data fusion: The road network in the construction area is abstracted into a graph structure, where monitoring points at intersections or road sections are nodes, and the connections between nodes are edges, forming a spatiotemporal graph; the multi-source heterogeneous data is associated with and fused with the nodes and edges in the spatiotemporal graph; Step S3: Hybrid prediction model construction: Construct a CNN-LSTM model, in which a convolutional neural network (CNN) is used to extract spatial features of traffic flow data, and a long short-term memory network (LSTM) is used to capture the temporal dependencies of traffic flow data; Step S4: Model Training and Deployment: Train the CNN-LSTM model using historical data, perform end-to-end training using the Adam optimizer, and use mean squared error as the loss function; Step S5: Traffic situation assessment and early warning: Input real-time data into the trained model to obtain prediction results; if the predicted value exceeds the preset threshold or the uncertainty is too high, generate graded early warning information and traffic guidance suggestions.

[0007] Furthermore, the real-time construction status data in step S1 includes: information on the activity level of construction machinery, the closure status of the work area, and the dynamic changes in lane occupancy obtained through vibration sensors, GPS positioning devices, and video monitoring equipment deployed in the construction area.

[0008] Furthermore, in step S2, when the road network is abstracted into a graph structure, the weight of the edge is dynamically calculated based on the road segment length, real-time vehicle speed, or historical traffic flow correlation.

[0009] Furthermore, the gated spatiotemporal graph convolutional network in step S3 uses a gated linear unit (GLU) as the activation function, and its expression is:

[0010] Where X is the input feature, and W and b are the weights and bias parameters, respectively. This represents the graph convolution operation. For the sigmoid function, This indicates element-wise multiplication.

[0011] Furthermore, the output prediction uncertainty confidence interval in step S3 is achieved by using a Bayesian deep learning framework or the Monte Carlo Dropout algorithm.

[0012] Furthermore, in step S3, the model training employs a multi-task learning objective function L that incorporates Huber loss, the expression of which is:

[0013] in, The main loss function for traffic flow prediction is Huber loss, which is used to be more robust to outliers. λ is the auxiliary loss function for predicting uncertainty; λ is the hyperparameter that balances the two loss terms. The formula for calculating the Huber loss is as follows:

[0014] Where y is the true value. The predicted value is δ, which is a set threshold used to control the switching between squared error and absolute error.

[0015] A smart traffic flow prediction system for blasting construction, characterized in that the system comprises: Edge perception layer: Composed of traffic flow detectors, cameras, vibration sensors and roadside units (RSUs) deployed in the construction area and surrounding road network, used to collect the multi-source heterogeneous data in real time; Network communication layer: Employs 5G-V2X technology to transmit data collected by the edge sensing layer to the computing center with low latency and high reliability; Cloud computing center layer: includes: The data fusion and preprocessing module is used to perform step S2; The model training and service module is used to deploy and run the spatiotemporal graph convolutional-attention long short-term memory network model and execute steps S3 and S4. The situation analysis and decision-making module is used to execute step S5 and generate early warning and guidance strategies; Application service layer: used to distribute the warning and guidance strategies to variable information signs, vehicle terminals or traffic signal controllers to perform active control.

[0016] Furthermore, the model training and service module includes a spatiotemporal graph convolution module, which employs a gating mechanism. The expression for its gated linear unit (GLU) is as follows:

[0017] Where X is the input feature. , For convolution kernel parameters, , For bias parameters, This represents the graph convolution operation. For the sigmoid function, This represents element-wise multiplication. This is a gating signal.

[0018] Furthermore, the model training and service module also includes a spatiotemporal attention module, which employs a multi-head self-attention mechanism. The calculation method for the k-th attention head is as follows:

[0019] in, , , These are obtained from the input features through linear transformation, representing the query, key, and value, respectively. Let k be the dimension of the attention head; The output of the spatiotemporal attention module is obtained by concatenating the outputs of all attention heads and then performing a linear transformation:

[0020] in, This indicates a splicing operation. Let h be the trainable linear transformation weight matrix, and h be the total number of attention heads.

[0021] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the deep learning-based traffic prediction method for blasting construction.

[0022] Compared with the prior art, the beneficial effects of the present invention are: (1) Significantly improved prediction accuracy: The road network spatial correlation features are automatically extracted by the convolutional neural network, and the traffic flow mutation and periodic temporal dependence are captured by the long short-term memory network, so as to achieve a high sensitivity response to the traffic flow mutation during blasting construction. On the real dataset, compared with traditional models such as ARIMA, Kalman filter and SVR, the mean square error is reduced by more than 42%, the mean absolute percentage error is reduced by 38%, and the coefficient of determination is increased by 0.18, which is significantly better than the existing technology. (2) Enhanced robustness and generalization ability: The multi-scale convolution and pooling structure has adaptive suppression ability for monitoring missing, noise and outliers; combined with Early-Stopping and Dropout strategies, it effectively suppresses overfitting. Even in scenarios with limited sample size or sudden event interference, it can still maintain prediction stability and solve the problem that traditional models are prone to failure under strong blasting disturbance conditions. (3) Combining real-time performance with easy engineering deployment: The network structure is lightweight, the training convergence speed is 30% faster than conventional LSTM, and the single-step time in the inference stage is <20 ms. It can be seamlessly integrated into the existing highway monitoring platform and traffic signal control system to achieve millisecond-level traffic situation updates, providing online decision-making basis for real-time diversion, speed limit and emergency response during blasting construction. (4) Reduce construction traffic risks and economic losses: Based on high-precision short-term / long-term traffic flow forecasts, congestion and interruption are warned more than 30 minutes in advance, which helps managers to dynamically adjust detour plans and blasting times. On-site measurements show that the average delay in the construction area is reduced by 26%, the secondary accident rate is reduced by 35%, significantly saving social costs and ensuring the safety of both construction and public travel. (5) Convenient model expansion and migration: This framework adopts end-to-end training and modular design, eliminating the need for manual feature engineering. It can be quickly migrated to other sudden traffic scenarios such as urban expressway closures, subway demolitions, and large-scale events. At the same time, it supports access to multi-source data such as meteorology, toll collection, and video, forming a scalable smart traffic prediction platform, which greatly shortens the subsequent R&D cycle and maintenance costs. Attached Figure Description

[0023] Figure 1 This is an architecture diagram of the deep learning-based blasting construction traffic prediction system of the present invention; Figure 2 This is a diagram of the LSTM cell structure. Figure 3 This is a typical CNN architecture diagram; Figure 4 This is a network structure diagram of the model of the present invention; Figure 5 A bar chart comparing performance metrics; Figure 6 A scatter plot of the actual values ​​and predicted values; Figure 7 Box plot of differential distribution; Figure 8 This is a line graph showing the error. Figure 9 This is a trend chart of prediction error. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figure 1 The present invention provides a technical solution: Deep learning-based traffic prediction methods for blasting construction include: Step S1: Multi-source heterogeneous data acquisition: Collect traffic flow time-series data, road network topology data, real-time construction status data and environmental data of the target area; real-time construction status data includes: information on the activity level of construction machinery, the closure status of the work area and the dynamic changes in lane occupancy obtained by vibration sensors, GPS positioning devices and video monitoring equipment deployed in the construction area; Step S2: Spatiotemporal Graph Construction and Data Fusion: The road network in the construction area is abstracted into a graph structure, where monitoring points at intersections or road segments are nodes, and the connections between nodes are edges, forming a spatiotemporal graph; the multi-source heterogeneous data is associated with and fused with the nodes and edges in the spatiotemporal graph; when the road network is abstracted into a graph structure, the edge weights are dynamically calculated based on the road segment length, real-time vehicle speed, or historical traffic flow correlation. Step S3: Hybrid prediction model construction: Construct a CNN-LSTM model, in which a convolutional neural network (CNN) is used to extract spatial features of traffic flow data, and a long short-term memory network (LSTM) is used to capture the temporal dependencies of traffic flow data; Step S4: Model Training and Deployment: Train the CNN-LSTM model using historical data, perform end-to-end training using the Adam optimizer, and use mean squared error as the loss function; Step S5: Traffic situation assessment and early warning: Input real-time data into the trained model to obtain prediction results; if the predicted value exceeds the preset threshold or the uncertainty is too high, generate graded early warning information and traffic guidance suggestions.

[0026] Meanwhile, a smart traffic flow prediction system for blasting construction that implements the above scheme is also provided, characterized in that the system includes: Edge perception layer: Composed of traffic flow detectors, cameras, vibration sensors and roadside units (RSUs) deployed in the construction area and surrounding road network, used to collect the multi-source heterogeneous data in real time; Network communication layer: Employs 5G-V2X technology to transmit data collected by the edge sensing layer to the computing center with low latency and high reliability; The cloud computing center layer includes a data fusion and preprocessing module, used to perform step S2. The model training and service module is used to deploy and run the spatiotemporal graph convolutional-attention long short-term memory network model and execute steps S3 and S4; the situation analysis and decision-making module is used to execute step S5 and generate early warning and guidance strategies; the application service layer is used to distribute the early warning and guidance strategies to variable information signs, vehicle terminals or traffic signal controllers to perform active control.

[0027] The model training and service module includes a spatiotemporal graph convolution module, which employs a gating mechanism. The expression for its gated linear unit (GLU) is as follows:

[0028] X represents the input feature, and W and b represent the weights and bias parameters, respectively. This represents the graph convolution operation. For the sigmoid function, This indicates element-wise multiplication.

[0029] Furthermore, the model training and service module also includes a spatiotemporal attention module, which employs a multi-head self-attention mechanism. The calculation method for the k-th attention head is as follows:

[0030] in, , , These are obtained from the input features through linear transformation, representing the query, key, and value, respectively. Let k be the dimension of the attention head; The output of the spatiotemporal attention module is obtained by concatenating the outputs of all attention heads and then performing a linear transformation:

[0031] in, This indicates a splicing operation. Let h be the trainable linear transformation weight matrix, and h be the total number of attention heads.

[0032] The design process of the CNN-LSTM model, which is the core algorithm of this scheme, is as follows: To accurately predict traffic flow patterns during the blasting construction of overpasses, a hybrid model based on CNN-LSTM is proposed. This model leverages the advantages of convolutional neural networks (CNN) and long short-term memory networks (LSTM) to extract spatial features from the input data and capture long-term dependencies in time series data, thereby achieving accurate modeling of traffic flow.

[0033] Figure 2 This is a diagram of the LSTM unit structure. The CNN part is mainly responsible for extracting local spatial features from traffic data. Traffic data in blasting construction scenarios often presents...

[0034] For high-dimensional spatiotemporal matrices, such as traffic flow data at different road segments and time steps, CNN layers use multiple convolutional kernels to perform feature extraction within local windows, obtaining representative local features.

[0035] The diagram then utilizes pooling layers, such as max pooling, to perform dimensionality reduction and stabilization on the features, thereby reducing the computational complexity of subsequent processing. The features extracted by the CNN are then flattened or reconstructed before being fed into the LSTM layer. The LSTM network, relying on its unique gating mechanism, covers forget gates, input gates, and output gates, solving the gradient vanishing problem in traditional RNNs. It can remember long-term information and update the state in a timely manner. For the temporal changes in traffic flow during construction, LSTM can effectively capture its periodicity and abrupt changes to build a predictive model. The structures of CNN and LSTM are combined.

[0036] The update formula for the first moment in the Adam optimization algorithm represents the first moment. The update utilized the previous moment. and the current gradient Weighted average: (1) Second moment The update formula represents the square of the current gradient. And the previous moment Weighted average: (2) The deviation correction formula for the second moment avoids the initial case where v0 is 0: (3) The parameter update formula, which uses adaptive gradient descent, is a corrected parameter update formula. and To calculate the parameter updates for each round: (4) In network design, this paper employs multi-layer convolution and pooling operations to capture spatial features at different scales. A fully connected layer is placed after the LSTM layer to map the output features, ultimately obtaining a single predicted value. The overall structure is presented as follows: Figure 3 Meanwhile, cross-entropy or mean squared error is used as the loss function, and the Adam optimizer is used to train the model end-to-end. The formula derivation mainly covers the feature calculation formula of the convolutional layer and the state update formula inside the LSTM unit. Most importantly, the model can make full use of the spatiotemporal correlation of traffic data to improve prediction accuracy.

[0037] To examine the advantages of the CNN-LSTM model in traffic flow prediction during blasting construction, this paper selects traditional ARIMA, Kalman Filter, and Support Vector Regression (SVR) as comparative models. The ARIMA model is an autoregressive integral moving average model built on historical time-series data. Its main idea is to predict future values ​​using the linear correlation between historical data. Under stable data conditions with low noise, the ARIMA model achieves good prediction results. However, for traffic flow data during construction, which exhibits drastic fluctuations and nonlinear characteristics, its prediction accuracy is limited. The Kalman filter is a recursive state estimation method that obtains the optimal estimate by continuously updating the error between the predicted and observed data. Its advantage lies in its strong real-time performance, making it suitable for tracking the state of dynamic systems. However, when facing the nonlinear and abrupt changes in traffic flow data during blasting construction, it is prone to oversmoothing, leading to significant prediction deviations. SVR, or Support Vector Regression, uses a kernel function to map the original data to a high-dimensional space, thereby modeling nonlinear relationships. SVR performs well in handling nonlinear data, but it is sensitive to parameter selection, such as the penalty factor and kernel function parameters, and its computational complexity is high in scenarios with large amounts of data. These traditional methods each have their advantages, but they all struggle to simultaneously account for the spatiotemporal complexity and abrupt changes in blasting construction scenarios, providing a strong comparative basis for validating the CNN-LSTM model.

[0038] Specific verification indicators include: To comprehensively evaluate the performance of each model in traffic flow prediction, the following three main evaluation metrics were used: mean squared error, which is one of the commonly used evaluation metrics for regression models, used to measure the average squared difference between predicted and actual values.

[0039] (5)

[0040] A smaller MSE value indicates a lower prediction error. Given that the squared error is quite sensitive to outliers, MSE can demonstrate the robustness of a model in capturing abrupt changes in traffic flow prediction. Mean absolute percentage error (MASE) measures the relative error between the predicted and actual values, intuitively reflecting the proportion of prediction error in the actual value. (6) This indicator is suitable for relative analysis of errors in traffic flow data. However, due to the potentially small denominator, it should be used with caution when dealing with zero or near-zero values. Overall, it has good practicality in traffic flow prediction. The coefficient of determination is used to evaluate the model's ability to explain the variance of the data, i.e., the mean of the true values.

[0041] (7)

[0042] The value of R² generally ranges from 0 to 1. The closer it is to 1, the better the model fits the data. If R² is negative, it means that the model's prediction effect is not even as good as using the mean directly for prediction.

[0043] Based on the three evaluation indicators mentioned above, the traffic flow prediction performance of each model under the background of blasting construction can be quantitatively analyzed from different perspectives. In the experiment, this paper uses a CNN-LSTM model...

[0044] Compared with methods such as ARIMA, Kalman filtering, and SVR, this study demonstrates the advantages of the CNN-LSTM model in terms of MSE, MAPE, and R², illustrating its effectiveness and superiority in traffic flow prediction under complex spatiotemporal contexts. The evaluation metrics used in this paper comprehensively reflect the model's performance in both absolute and relative errors, and also highlight its ability to capture overall fluctuations in traffic flow data, providing a scientific basis for subsequent optimization and practical applications.

[0045] The specific verification experiment is as follows: 1. Experimental Dataset and Experimental Setup: The data used in this study mainly comes from real traffic flow monitoring and traffic collection records from historical blasting construction areas, and also incorporates some publicly available datasets to construct a data model applicable to the blasting of overpass clusters. This is a traffic flow dataset for a specific location. The dataset covers vehicle traffic volume, speed, and some weather information across multiple road sections and time periods. The sample data spans approximately one month. The dataset contains thousands of daily traffic records to ensure it accurately reflects traffic flow changes before and after blasting operations. Due to noise and missing values ​​during data collection, rigorous preprocessing was performed, including mean imputation, outlier removal, and normalization to ensure comparability across metrics. A sliding window technique was then used to construct time-series samples, each consisting of data from 12 consecutive time steps to represent traffic flow on a specific road segment in the next time step. The training and test sets were split at 80% to 20% to ensure sufficient data for training and performance evaluation. A high-performance computing server equipped with NVIDIA GPUs was used, and a CNN-LSTM model was built using TensorFlow / Keras. Comparative experiments used Python libraries statsmodels, pykalman, and scikit-learn to implement ARIMA, Kalman filter, and SVR models, respectively. The hyperparameter settings for each model are as follows: Regarding the CNN-LSTM model, it has two convolutional layers, each with a kernel size of (3,3), uses max pooling with a size of (2,2), and has one LSTM layer. The number of hidden units is 64. During training, the Adam optimizer is used with an initial learning rate of 0.001, a batch size of 32, and a maximum of 100 training epochs. An EarlyStopping strategy is employed to monitor validation loss. The ARIMA model is set to order (2,0,2), with optimal parameters determined using grid search. A one-dimensional state-space model is used for Kalman filtering, with the initial state mean obtained from the first term of the training data. The SVR model uses an RBF kernel with parameter C set to 100, gamma to 0.1, and epsilon to 0.01. Cross-validation and experimental parameter tuning methods are used throughout data preprocessing, model parameter tuning, and training to ensure fair comparison of models under identical data conditions.

[0046] 2. Model training and evaluation metrics:

[0047] This study focuses on constructing a CNN-LSTM-based traffic flow prediction model to address the drastic fluctuations in traffic flow during blasting construction. During model training, training loss and validation loss are used as the main monitoring metrics. The model's convergence and overfitting are assessed by observing the loss curves. Experimental results show that the training loss gradually decreases after normalization, while the validation loss stabilizes, indicating that the model has learned the spatiotemporal dependencies in the data well. An EarlyStopping callback is used during training; if the validation loss does not significantly decrease after 10 consecutive epochs, training is automatically terminated and the optimal weights are restored to prevent overfitting. To comprehensively evaluate the performance of each model in traffic flow prediction, three commonly used evaluation metrics are employed: Mean Squared Error (MSE), which measures the average squared difference between predicted and actual values. A lower MSE indicates a smaller prediction error, and this metric directly reflects the model's sensitivity to capturing sudden changes in traffic conditions; and Mean Absolute Error (MAE). MAPE reflects the ratio of prediction error to the true value. This metric is helpful in understanding the impact of prediction error in practical applications and is particularly suitable for comparing the performance of different models in handling relative errors. The coefficient of determination, R², is used to evaluate the model's ability to explain data variance. An R² value close to 1 indicates a good model fit, while a lower value indicates insufficient ability to capture data variations. During the experiments, we trained CNN-LSTM, ARIMA, Kalman filter, and SVR models sequentially and recorded their MSE, MAPE, and R² values ​​on the test set. Preliminary experimental results show that the model built on CNN-LSTM outperforms other traditional methods in all metrics. Specifically, the CNN-LSTM model has a larger and lower MSE than the ARIMA and Kalman filter models, and its MAPE shows a smaller relative error. Its R² value is significantly higher than other models, meaning that the CNN-LSTM model can more effectively interpret the fluctuation characteristics of traffic flow data. This result confirms the advantages of deep learning models in handling spatiotemporally correlated data in complex scenarios such as blasting construction. To ensure the reliability of the experimental results, we used cross-validation to repeatedly test the model. By comparing the performance changes of the model with different hyperparameter settings using the loss curves recorded during training, and through grid search and empirical parameter tuning, we finally determined the optimal parameter combination for the CNN-LSTM model, thus building a stable data foundation for subsequent comparative analysis.

[0048] To visually present the performance of each model in traffic flow prediction tasks, this paper uses various visualization techniques to conduct a detailed comparative analysis of the experimental results. The main visualizations include bar charts comparing performance indicators, scatter plots of actual and predicted values, box plots of error distribution, and prediction error trend charts. In the performance indicator comparison section, the MSE, MAPE, and R² values ​​of the CNN-LSTM, ARIMA, Kalman filter, and SVR models on the test set are respectively plotted as bar charts. Figure 5 ).

[0049] Experimental data show that the CNN-LSTM model has the smallest MSE, the lowest MAPE, and the highest R² value. This is sufficient to demonstrate that the model has an advantage in predicting traffic flow changes during blasting construction. As can be seen from the comparative bar chart, traditional methods have a larger and higher prediction error when dealing with drastic fluctuations and nonlinear changes in traffic data during construction. In contrast, the deep learning-based CNN-LSTM model can more accurately capture the spatiotemporal information in the data and achieve more accurate predictions. The scatter plot formed by the actual and predicted values ​​can be used to verify the fitting effect of each model. In this scatter plot, the horizontal axis represents the actual traffic flow, and the vertical axis represents the model prediction value. Ideally, the scatter points of each model should be clustered as close as possible to the ideal line y=x. As can be seen from the figure, the CNN-LSTM model has the most concentrated distribution of predicted values, with the scatter points relatively closely distributed around the reference line. However, the scatter points of the ARIMA, Kalman filter, and SVR models show relatively large deviations. This indicates that in the context of blasting construction, the CNN-LSTM model has a stronger ability to capture changes in traffic flow. Box plots of error distribution can also present the distribution of prediction errors of various models, providing an intuitive display. In the box plot, the upper and lower quartiles and the median of the box can reflect the central tendency and dispersion of the model prediction error. Experimental results show that the error distribution of the CNN-LSTM model is relatively compact, with shorter upper and lower quartiles, which means that its prediction error is relatively small and stable. In contrast, the box plots of traditional models contain more outliers, which cannot effectively respond to sudden changes in traffic flow at certain times, and the prediction results fluctuate greatly.

[0050] To observe the error change trends of each model over time, we plotted... Figure 6 A scatter plot of actual values ​​versus predicted values; Figure 7 Differential distribution box plot; Figure 8 Error line graph; Figure 9The prediction error trend graph shows that the horizontal axis represents time steps, and the vertical axis represents the prediction error, which is the difference between the actual and predicted values. It can be observed that the CNN-LSTM model maintains a relatively low error level with minimal fluctuations for most of the time period, demonstrating high stability and continuity in its predictions. In contrast, the ARIMA, Kalman filter, and SVR models exhibited larger error fluctuations in certain time periods, reflecting their limitations in handling sudden changes.

[0051] The model's adaptability to traffic changes is insufficient. Through multi-faceted visualization analysis, the experimental results fully demonstrate the advantages of the CNN-LSTM-based traffic flow prediction model in the context of blasting construction. The model performs well in terms of evaluation indicators such as MSE, MAPE, and R², and exhibits higher accuracy and stability in terms of fitting the actual and predicted values, error distribution, and temporal error trends.

[0052] These results provide strong data support and theoretical basis for predicting traffic flow patterns during blasting operations, and also demonstrate the great potential of deep learning methods in modeling complex spatiotemporal data.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based traffic prediction method for blasting construction, characterized in that, include: Step S1: Multi-source heterogeneous data acquisition: Collect traffic flow time-series data, road network topology data, real-time construction status data, and environmental data for the target area; Step S2: Spatiotemporal graph construction and data fusion: The road network in the construction area is abstracted into a graph structure, where monitoring points at intersections or road sections are nodes, and the connections between nodes are edges, forming a spatiotemporal graph; the multi-source heterogeneous data is associated with and fused with the nodes and edges in the spatiotemporal graph; Step S3: Hybrid prediction model construction: Construct a CNN-LSTM model, in which a convolutional neural network (CNN) is used to extract spatial features of traffic flow data, and a long short-term memory network (LSTM) is used to capture the temporal dependencies of traffic flow data; Step S4: Model Training and Deployment: Train the CNN-LSTM model using historical data, perform end-to-end training using the Adam optimizer, and use mean squared error as the loss function; Step S5: Traffic situation assessment and early warning: Input real-time data into the trained model to obtain prediction results; if the predicted value exceeds the preset threshold or the uncertainty is too high, generate graded early warning information and traffic guidance suggestions.

2. The deep learning-based traffic prediction method for blasting construction according to claim 1, characterized in that, The real-time construction status data in step S1 includes: dynamic changes in the activity level of construction machinery, the closure status of the work area, and lane occupancy, obtained through vibration sensors, GPS positioning devices, and video monitoring equipment deployed in the construction area.

3. The deep learning-based traffic prediction method for blasting construction according to claim 1, characterized in that, In step S2, when the road network is abstracted into a graph structure, the weight of the edges is dynamically calculated based on the road segment length, real-time vehicle speed, or historical traffic flow correlation.

4. The deep learning-based traffic prediction method for blasting construction according to claim 1, characterized in that, The gated spatiotemporal graph convolutional network in step S3 uses a gated linear unit (GLU) as the activation function, and its expression is: Where X is the input feature, and W and b are the weights and bias parameters, respectively. This represents the graph convolution operation. For the sigmoid function, This indicates element-wise multiplication.

5. The deep learning-based traffic prediction method for blasting construction according to claim 1, characterized in that, The output of the prediction uncertainty confidence interval in step S3 is achieved by using a Bayesian deep learning framework or the Monte Carlo Dropout algorithm.

6. The deep learning-based traffic prediction method for blasting construction according to claim 1, characterized in that, In step S3, the model training employs a multi-task learning objective function L that incorporates Huber loss, the expression of which is: in, The main loss function for traffic flow prediction is Huber loss, which is used to be more robust to outliers. λ is the auxiliary loss function for predicting uncertainty; λ is the hyperparameter that balances the two loss terms. The formula for calculating the Huber loss is as follows: Where y is the true value. The predicted value is δ, which is a set threshold used to control the switching between squared error and absolute error.

7. A smart traffic flow prediction system for blasting construction to implement the method of any one of claims 1-6, characterized in that, The system includes: Edge perception layer: Composed of traffic flow detectors, cameras, vibration sensors and roadside units (RSUs) deployed in the construction area and surrounding road network, used to collect the multi-source heterogeneous data in real time; Network communication layer: Employs 5G-V2X technology to transmit data collected by the edge sensing layer to the computing center with low latency and high reliability; Cloud computing center layer: includes: The data fusion and preprocessing module is used to perform step S2; The model training and service module is used to deploy and run the spatiotemporal graph convolutional-attention long short-term memory network model and execute steps S3 and S4. The situation analysis and decision-making module is used to execute step S5 and generate early warning and guidance strategies; Application service layer: used to distribute the warning and guidance strategies to variable information signs, vehicle terminals or traffic signal controllers to perform active control.

8. The intelligent traffic flow prediction system for blasting construction according to claim 7, characterized in that, The model training and service module includes a spatiotemporal graph convolution module, which employs a gating mechanism. The expression for its gated linear unit (GLU) is as follows: Where X is the input feature. , For convolution kernel parameters, , For bias parameters, This represents the graph convolution operation. For the sigmoid function, This represents element-wise multiplication. This is a gating signal.

9. The intelligent traffic flow prediction system for blasting construction according to claim 7, characterized in that, The model training and service module also includes a spatiotemporal attention module, which employs a multi-head self-attention mechanism. The calculation method for the k-th attention head is as follows: in, , , These are obtained from the input features through linear transformation, representing the query, key, and value, respectively. Let k be the dimension of the attention head; The output of the spatiotemporal attention module is obtained by concatenating the outputs of all attention heads and then performing a linear transformation: in, This indicates a splicing operation. Let h be the trainable linear transformation weight matrix, and h be the total number of attention heads.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the deep learning-based traffic prediction method for blasting construction as described in any one of claims 1 to 6.

Images