Winter asphalt pavement surface temperature prediction method based on integrated deep learning

Abstract

The present application relates to a kind of winter asphalt pavement surface temperature prediction method based on integrated deep learning, belong to highway traffic meteorological monitoring technical field.The method includes: according to correlation coefficient method and expert knowledge, the meteorological factor that is accumulated to the road surface temperature Influence is selected;Sliding window method is used to cut time series data, and supervised learning sample is constructed;Supervised learning sample is obtained, and the winter asphalt pavement surface temperature prediction model of integrated deep learning is trained;Winter asphalt pavement surface temperature prediction method based on integrated deep learning includes with LSTM neural network module, GRU neural network module, feature fusion layer and fully connected layer;The trained model is used to predict hourly pavement surface temperature.The present application is through the cumulative influence of the integration of deep learning network of multi-dimensional meteorological factor and pavement surface temperature periodicity, improves the accuracy and robustness of pavement surface temperature, with data-driven, self-learning, multi-granularity prediction and other technical characteristics.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent traffic and traffic safety, and particularly relates to a winter asphalt pavement surface temperature prediction method based on integrated deep learning. BACKGROUND

[0002] Temperature is an important influencing factor of the skid resistance, mechanical properties and durability of asphalt pavement, and it affects the safety, maneuverability and comfort of highways. Unlike the high summer road surface temperature (60℃ and above), which easily causes tire burst and rutting of the road surface, the low winter road surface temperature (0℃ and below) easily causes icy and slippery road surface, freeze-thaw damage, worsens road driving conditions, and easily induces major traffic accidents and large-scale traffic congestion. Therefore, from the perspective of actively preventing accidents from the source, how to accurately predict the winter road surface temperature to provide an important basis for road icing risk analysis, and then through timely and active travel information services, rapid road maintenance and active traffic management activities, to reduce the impact of low-temperature ice and snow weather on the highway transportation system, has become a key technical problem that needs to be solved in the intelligent traffic operation and control of digital traffic and smart highways.

[0003] In recent years, with the development of traffic meteorological data collection technology, a large number of methods for predicting road surface temperature have emerged, mainly including theoretical analysis method, statistical analysis method and machine learning method. The theoretical analysis method is based on the principles of thermodynamics and the road surface radiation energy balance condition, and uses numerical calculation to develop a mechanism model for predicting road surface temperature. However, this method involves multiple physical parameters, numerical calculation is complex, and the model has poor universality. Statistical analysis is based on historical road surface temperature data, and uses descriptive statistics to analyze the spatial and temporal distribution characteristics of road surface temperature to construct a road surface temperature prediction model. Statistical analysis method has simple calculation, and the local prediction model has high accuracy in instance analysis. However, most of the models target the daily maximum and minimum road surface temperature, and it is difficult to ensure the timeliness of the prediction. In recent years, with the rapid development of new generation information technologies such as big data and artificial intelligence, data-driven machine learning algorithms have shown excellent performance in modeling complex problems with high-dimensional and large data, which provides a new opportunity for the application of artificial intelligence technology in road surface temperature prediction. However, most existing machine learning models develop prediction models from the perspective of periodicity of road surface temperature, without combining the model with the heat transfer characteristics of the road surface, and lack of consideration of the near-term trend of road surface temperature, which may result in large prediction errors when the weather suddenly changes. SUMMARY

[0004] The purpose of the present application is to solve the problems of the prior art, and to provide a winter asphalt pavement surface temperature prediction method based on integrated deep learning.

[0005] To achieve the above-mentioned purpose, the technical solution adopted by the present application is as follows:

[0006] The winter asphalt pavement surface temperature prediction method based on integrated deep learning comprises the following steps:

[0007] Step (1), based on the Spearman correlation coefficient and the pavement thermodynamics theory, select the meteorological factors that have cumulative influence on the pavement temperature;

[0008] Step (2), use the sliding window method to cut the winter historical time series data of the meteorological factors that have cumulative influence and the winter historical time series data of the pavement surface temperature, and construct the supervised learning sample;

[0009] Step (3), use the supervised learning sample obtained in step (2) to train the winter asphalt pavement surface temperature prediction model based on integrated deep learning;

[0010] The winter asphalt pavement surface temperature prediction model based on integrated deep learning comprises an LSTM neural network module, a GRU neural network module, a feature fusion layer and a full connection layer;

[0011] The winter asphalt pavement surface temperature prediction model based on integrated deep learning uses the LSTM neural network module to extract the cumulative influence features of the meteorological factors, uses the GRU neural network module to extract the periodic features of the pavement surface temperature, then uses the feature fusion layer to fuse the extracted periodic features and cumulative influence features into a feature vector, and finally outputs the pavement surface temperature after future Δh through two full connection layers; wherein the input of the LSTM neural network module is a cumulative influence feature matrix;

[0012] The input of the GRU neural network module is a periodic feature matrix.

[0013] Step (4), collect the meteorological factor data that has cumulative influence on the pavement temperature in real time, and use the integrated deep learning model to predict the highway pavement surface temperature.

[0014] Further, preferably, in step (1), the selected meteorological factors that have cumulative influence on the pavement temperature are air temperature, humidity, rainfall and wind speed.

[0015] Further, preferably, in step (3), the cumulative influence feature matrix is as shown in formula (1):

[0016]

[0017] Wherein, n is the size of the cumulative influence sliding window, representing the historical Δt1 hours before, represents the air temperature observed by the meteorological station at time t, and the historical air temperature observed from time t-n to the current time t is represented as

[0018] represents the humidity observed by the weather station at time t; the historical humidity observed from time t-n to current time t is represented as

[0019] represents the wind speed observed by the weather station at time t; the historical wind speed observed from time t-n to current time t is represented as

[0020] represents the rainfall observed by the weather station at time t; the historical rainfall observed from time t-n to current time t is represented as

[0021] The periodic feature matrix is shown in formula (2):

[0022] Y t d = [y (t+Δh)-24 ,y (t+Δh)-48 ,…,y (t+Δh)-24×m ](2)

[0023] wherein m is the size of the sliding window of the periodic feature influence, representing the history of the previous m days, y t represents the road surface temperature observed by the weather station at time t, Δh represents the predicted time interval, and m represents the number of days.

[0024] Further, preferably, n is 7, m is 3; the loss function is MAE, the learning rate is 0.001, and the optimizer is Adam.

[0025] Further, preferably, assuming that the number of nodes of the GRU neural network module and the LSTM neural network module are j and k respectively, the cumulative influence feature is represented as The periodic feature is represented as Then, the cumulative influence feature and the periodic feature are input into the feature fusion module, which fuses them into a fusion feature

[0026] Further, preferably, the LSTM neural network module includes 2 LSTM layers, the first layer LSTM network has 100 neurons, and the second layer LSTM network has 200 neurons; the GRU neural network module includes 2 GRU layers, the first layer GRU network has 100 neurons, and the second layer GRU network has 200 neurons; the LSTM neural network module and the GRU neural network module both use tanh as the activation function; after the outputs of the LSTM module and the GRU module are fused through the feature fusion layer, they are input into two fully connected layers for prediction.

[0027] Further, preferably, the Ah is 1h. Those skilled in the art should know that the Ah supports multiple prediction intervals such as 1h, 3h and 6h, and preferably the Ah is 1h.

[0028] The application also provides a winter asphalt pavement surface temperature prediction system based on integrated deep learning, comprising:

[0029] A meteorological factor selection module is configured to select meteorological factors that have cumulative effects on pavement temperature based on Spearman correlation coefficients and pavement thermodynamics theory.

[0030] A first processing module is configured to divide winter historical time series data of meteorological factors that have cumulative effects and winter historical time series data of pavement surface temperature by using a sliding window method to construct supervised learning samples.

[0031] A second processing module is configured to train a winter asphalt pavement surface temperature prediction model based on integrated deep learning by using the supervised learning samples obtained by the first processing module.

[0032] The winter asphalt pavement surface temperature prediction model based on integrated deep learning comprises an LSTM neural network module, a GRU neural network module, a feature fusion layer and a full connection layer.

[0033] The winter asphalt pavement surface temperature prediction model based on integrated deep learning extracts cumulative effect features of meteorological factors by using the LSTM neural network module, extracts periodic features of pavement surface temperature by using the GRU neural network module, then fuses the extracted periodic features and cumulative effect features into a feature vector by using the feature fusion layer, and finally outputs pavement surface temperature after Ah by using two full connection layers; wherein the input of the LSTM neural network module is a cumulative effect feature matrix.

[0034] The input of the GRU neural network module is a periodic feature matrix.

[0035] A pavement surface temperature prediction module is configured to collect meteorological factor data that have cumulative effects on pavement temperature in real time, and predict highway pavement surface temperature by using the integrated deep learning model.

[0036] The application also provides an electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the winter asphalt pavement surface temperature prediction method based on integrated deep learning when executing the program.

[0037] The application further provides a non-transitory computer-readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement steps of the winter asphalt pavement surface temperature prediction method based on integrated deep learning.

[0038] In the application, it is assumed that the size of the sliding window is n, which means that the meteorological factor data of the previous [t-n, t) time period is used to predict the surface temperature at time t+Δh.

[0039] The model proposed in the application has two inputs: the first one represents the cumulative influence of the meteorological factors at the adjacent time, and the data of the previous [t-n, t) hours is input into the LSTM neural network module to predict the road surface temperature at time t+Δh. The second one represents the periodicity of the road surface temperature, and the road surface temperature at time t+1 in the previous m days is used to predict the road surface temperature at time t+Δh through the GRU neural network module.

[0040] In the application, the cumulative influence feature matrix and the periodicity feature matrix are proposed by analyzing the change rule of the road surface temperature and comprehensively considering the periodicity of the road surface temperature itself and the adjacent and cumulative influence of the multi-dimensional meteorological factors. There is a heat conduction effect between the road surfaces, and after the meteorological factors act on the upper layer of the road surface structure, it takes a certain time for heat to conduct to the deep part of the road surface, and it also takes a certain time for the deep part of the road surface to release heat to the road surface, which makes the influence of the meteorological factors on the road surface temperature cumulative and lagging. At the same time, the road surface temperature shows a clear periodic change rule, and the road surface temperature is affected by the air temperature and solar radiation during the day, the temperature gradually rises, reaches the highest value in the afternoon, and then gradually decreases until the lowest value in the morning.

[0041] The application uses the characteristics of the LSTM neural network module that can remember and update information to simulate the storage and release of heat by the road surface and extract the cumulative influence of the meteorological features. The GRU neural network module is used to extract the periodicity of the road surface temperature. Thus, a road surface temperature prediction model that comprehensively considers the periodicity of the road surface temperature and the adjacent and cumulative influence of the multi-dimensional meteorological factors is developed.

[0042] The innovation of the application lies in that:

[0043] (a) There is a heat conduction effect between the road surfaces, and after the meteorological factors act on the upper layer of the road surface structure, it takes a certain time for heat to conduct to the deep part of the road surface, and it also takes a certain time for the deep part of the road surface to release heat to the road surface. The cumulative influence of the meteorological factors on the road surface temperature is analyzed by using the characteristics of the Spearman coefficient that can analyze the nonlinear correlation and combining the road thermodynamics theory, and it is obtained that the meteorological features that cumulatively influence the road surface temperature are air temperature, humidity, rainfall and wind speed.

[0044] (b) A winter asphalt pavement road surface temperature prediction method based on integrated deep learning is proposed, wherein the winter asphalt pavement road surface temperature prediction model based on integrated deep learning comprises three sub-modules, namely a feature extraction module, a feature fusion module and a prediction module.

[0045] (c) The constructed road surface temperature prediction algorithm based on integrated deep neural network has stronger generalization ability and is more suitable for various complex environments and scenes such as night and adverse weather.

[0046] The present application mainly aims at the problem of winter low-temperature asphalt pavement icing, uses the meteorological data and road surface temperature data collected in winter, and proposes a winter asphalt pavement road surface temperature prediction method based on integrated deep learning, so as to accurately capture the dynamic change trend and characteristics of the road surface temperature and develop a high-precision pavement temperature prediction model, thereby providing a reference for winter highway operation safety management and traffic meteorological disaster prevention.

[0047] Compared with the prior art, the present application has the following beneficial effects:

[0048] The existing model mainly develops a prediction model from the periodicity of the road surface temperature, the present application discards the previous idea of predicting only the periodicity of the road surface temperature, comprehensively considers the periodicity of the road surface temperature and the thermodynamic characteristics of the pavement, proposes an integrated deep learning model that fuses the two, simulates the heat storage and release process of the pavement by using the 'cell state' of the LSTM, and on this basis, extracts the periodicity of the road surface temperature by using the GRU, then fuses it by using the feature fusion layer, thereby improving the accuracy and robustness of the road surface temperature prediction, having the technical characteristics of data-driven, self-learning and multi-granularity prediction, and being suitable for intelligent travel services for winter highway icing risk short-term early warning in various terrain environments such as plateau mountainous areas, plains and hilly areas, and more suitable for various complex environments and scenes such as night and adverse weather. The performance of the proposed integrated deep learning model is verified in the prediction of the road surface temperature in the future 1h, 3h and 6h, and the average absolute error is 0.345, 0.833 and 1.743, which has higher prediction accuracy than the CNN-LSTM network, SVR and BP network in all prediction intervals. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 Structure diagram of the road surface temperature prediction method based on integrated deep learning

[0050] Figure 2A schematic diagram of the process of absorbing and releasing heat for asphalt pavement;

[0051] Figure 3 Figures are comparison charts of different models; (a) prediction interval 1h; (b) prediction interval 3h; (c) prediction interval 6h;

[0052] Figure 4 Figure is a structural schematic diagram of the road surface temperature prediction system based on integrated deep learning of the present application;

[0053] Figure 5 Figure is a structural schematic diagram of the electronic device of the present application. DETAILED DESCRIPTION

[0054] The present application will be further described in detail below with reference to examples.

[0055] Those skilled in the art will understand that the following examples are for illustration only and should not be taken as limiting the scope of the present application. If a specific technique or condition is not specified in the examples, it is performed according to the technique or condition described in the literature in the art or according to the product manual. If the manufacturer of the material or equipment is not specified, it is a conventional product that can be obtained by purchase.

[0056] The winter asphalt pavement road surface temperature prediction method based on integrated deep learning comprises the following steps:

[0057] Step (1), based on the Spearman correlation coefficient and the road surface thermodynamics theory, selecting meteorological factors that have cumulative influence on the road surface temperature;

[0058] Step (2), using the sliding window method to cut the winter historical time series data of the meteorological factors that have cumulative influence and the winter historical time series data of the road surface temperature, and constructing supervised learning samples;

[0059] Step (3), using the supervised learning samples obtained in step (2) to train the winter asphalt pavement road surface temperature prediction model based on integrated deep learning;

[0060] The winter asphalt pavement road surface temperature prediction model based on integrated deep learning comprises an LSTM neural network module, a GRU neural network module, a feature fusion layer and a fully connected layer;

[0061] The winter asphalt pavement road surface temperature prediction model based on integrated deep learning uses the LSTM neural network module to extract the cumulative influence features of the meteorological factors, uses the GRU neural network module to extract the periodic features of the road surface temperature, then uses the feature fusion layer to fuse the extracted periodic features and cumulative influence features into a feature vector, and finally outputs the road surface temperature after the future Ah through two fully connected layers; wherein the input of the LSTM neural network module is a cumulative influence feature matrix;

[0062] The input of the GRU neural network module is a periodic feature matrix.

[0063] In step (4), meteorological factor data having a cumulative effect on the road surface temperature is collected in real time, and an integrated deep learning model is used to predict the highway road surface temperature.

[0064] In step (1), the selected meteorological factors having a cumulative effect on the road surface temperature are air temperature, humidity, rainfall, and wind speed.

[0065] In step (3), the cumulative influence feature matrix is shown in formula (1):

[0066]

[0067] wherein, represents the air temperature observed by the meteorological station at time t, and the historical air temperature observed from time t-n to the current time t is represented as

[0068] represents the humidity observed by the meteorological station at time t, and the historical humidity observed from time t-n to the current time t is represented as

[0069] represents the wind speed observed by the meteorological station at time t, and the historical wind speed observed from time t-n to the current time t is represented as

[0070] represents the rainfall observed by the meteorological station at time t, and the historical rainfall observed from time t-n to the current time t is represented as

[0071] The periodic feature matrix is shown in formula (2):

[0072]

[0073] wherein, y t represents the road surface temperature observed by the meteorological station at time t, Δh represents the predicted time interval, and m represents the number of days.

[0074] n is 7, m is 3, the loss function is MAE, the learning rate is 0.001, and the optimizer is Adam.

[0075] Suppose the number of nodes of the GRU neural network module and the LSTM neural network module are j and k respectively, the cumulative influence feature is represented as The periodic feature is represented as Then, the cumulative influence feature and periodic features an input feature fusion module, the feature fusion module fuses them into a fused feature that contains both cumulative effects and periodicity

[0076] The LSTM neural network module includes 2 LSTM layers, the first layer of LSTM network is 100 neurons, and the second layer of LSTM network is 200 neurons; the GRU neural network module includes 2 GRU layers, the first layer of GRU network is 100 neurons, and the second layer of GRU network is 200 neurons; the LSTM neural network module and the GRU neural network module both adopt tanh as the activation function; after the outputs of the LSTM module and the GRU module are fused through the feature fusion layer, they are input into two fully connected layers for prediction.

[0077] Those skilled in the art should know that Δh supports multiple prediction intervals such as 1h, 3h and 6h, and the preferred Δh is 1h.

[0078] The present application proposes a brand-new winter highway road surface temperature prediction method based on integrated deep learning (see Figure 1 ), which mainly includes three parts in the model: (1) feature extraction module; (2) feature fusion module; (3) prediction module, and the following is a detailed description of the three modules.

[0079] (1) Feature extraction module

[0080] There is a heat conduction effect between the road surface, after the meteorological factors act on the upper layer of the road surface structure, it takes a certain time for heat to conduct to the deep part of the road surface, and it also takes a certain time for the deep part of the road surface to release heat to the road surface, which makes the influence of meteorological factors on the road surface temperature cumulative and lagging, and the asphalt pavement structure is as shown in Figure 2 In order to more accurately predict the road surface temperature, it is tried to use LSTM neural network to simulate the storage and release of heat by the road surface, and the unique "cell state" of LSTM is similar to the asphalt pavement structure in principle, LSTM controls the "cell state" through the input gate, output gate and forget gate, while the asphalt pavement absorbs heat through heat conduction and exchanges heat with the surrounding environment through convection and evaporation. Therefore, the meteorological factors (air temperature, humidity, rainfall and wind speed) that have cumulative effects on the road surface temperature are selected, and a cumulative effect feature matrix is constructed as the input of the LSTM neural network. The calculation process of LSTM is as follows:

[0081] The forget gate uses equation (3) to determine the information that needs to be discarded in the old cell state; the input gate uses equation (4) to determine the proportion of new information that needs to be updated in the old cell state, and at the same time, equation (6) is used to determine the data information that needs to be updated; the new cell state consists of two parts: the information retained after passing through the forget gate and the new information updated by the input gate, which is calculated using equation (7); finally, the final output of the LSTM is determined by the output gate equation (5) and the new cell state, which is calculated using equation (8).

[0082]

[0083]

[0084]

[0085]

[0086]

[0087]

[0088] In the formula, The input of the LSTM layer at time t. In cellular state, W f W i W o and W c U is the weight matrix used to map the input layer to the forget gate, input gate, output gate, and cell state; f U i U o and U c b is the weight matrix used to map the previous output layer to the forget gate, input gate, output gate, and cell states; f b i b o and b c The bias vectors are calculated for the forget gate, input gate, output gate, and cell state. σ(·) is the sigmoid activation function, and tanh(·) is the hyperbolic tangent function. The output of the LSTM layer at time t.

[0089] Meanwhile, road surface temperature exhibits a clear periodic variation pattern. During the day, influenced by air temperature and solar radiation, the temperature gradually increases, reaching its peak in the afternoon, and then gradually decreases until reaching its lowest value in the early morning. As a variant of the LSTM neural network, GRU is easier to compute and train than LSTM, and can quickly capture the periodic pattern of road surface temperature. In GRU, the hidden layer unit structure consists of a reset gate and an update gate. The "cell state" and output gate in LSTM are removed, and the input gate and forget gate of LSTM are merged into a single update gate. Since information is directly passed through the hidden layer state, the LSTM model structure is simplified, and the network complexity is reduced. The computation process of GRU is as follows:

[0090]

[0091]

[0092]

[0093]

[0094] In the formula, Y t d r is the input of the GRU layer at time t. t and z t Reset the gate and update the gate's output when t is the time; W z W r and W h The input layer is mapped to the weight matrix of the reset gate, update gate, and candidate hidden states; U z U r and U c b is the weight matrix used to map the previous output layer to the reset gate, update gate, and candidate hidden states; z and b r The bias vectors for updating and resetting the gate. The output of the GRU layer at time t.

[0095] (2) Feature fusion module

[0096] Different features possess different attributes and express their impact on road surface temperature from different perspectives. Therefore, effectively combining different features is a crucial step in improving prediction performance. After a feature extraction module composed of LSTM and GRU, the cumulative impact features were obtained. and periodic characteristics Assuming the GRU neural network module and the LSTM neural network module have j and k nodes respectively, the cumulative influence feature can be expressed as: Periodicity can be represented as Then, the cumulative impact characteristics and periodic features an input feature fusion module, which fuses them into a fused feature that contains both cumulative effects and periodicity

[0097] (3) a prediction module

[0098] Finally, the obtained fused feature H t An input two-layer fully connected layer is used to realize the prediction of the road surface temperature.

[0099] The LSTM and GRU neural network modules used in the application respectively include 2 LSTM layers and 2 GRU layers; the first layer of the LSTM neural network module has 100 neurons, and the second layer of the LSTM neural network module has 200 neurons, and the GRU neural network module is the same as above, and tanh is used as the activation function; the outputs of the LSTM module and the GRU module are fused through a feature fusion layer, and then input into two layers of fully connected layers for prediction. Training parameters: (1) the best parameter n of cumulative effect is 7, and the best parameter m of periodicity is 3; (2) the loss function is MAE, the learning rate is 0.001, and the optimizer is Adam.

[0100] As shown in FIG. Figure 4 The winter asphalt pavement road surface temperature prediction system based on integrated deep learning includes:

[0101] The meteorological factor selection module 101 is used to select meteorological factors that have cumulative effects on the road surface temperature based on the Spearman correlation coefficient and the road surface thermodynamics theory.

[0102] The first processing module 102 is used to cut the winter historical time series data of the meteorological factors that have cumulative effects and the winter historical time series data of the road surface temperature by using the method of the sliding window, and construct a supervised learning sample.

[0103] The second processing module 103 is used to train the winter asphalt pavement road surface temperature prediction model based on integrated deep learning by using the supervised learning sample obtained by the first processing module.

[0104] The winter asphalt pavement road surface temperature prediction model based on integrated deep learning includes an LSTM neural network module, a GRU neural network module, a feature fusion layer and a fully connected layer.

[0105] The winter asphalt pavement surface temperature prediction model based on integrated deep learning extracts the cumulative influence characteristics of meteorological factors by using an LSTM neural network module, extracts the periodic characteristics of the surface temperature by using a GRU neural network module, then fuses the extracted periodic characteristics and cumulative influence characteristics into a feature vector by using a feature fusion layer, and finally outputs the surface temperature after future Δh through two fully connected layers; wherein the input of the LSTM neural network module is a cumulative influence feature matrix;

[0106] The input of the GRU neural network module is a periodic feature matrix.

[0107] The surface temperature prediction module 104 is used for collecting meteorological factor data which has a cumulative influence on the road surface temperature in real time, and predicting the highway surface temperature by using the integrated deep learning model.

[0108] The winter asphalt pavement surface temperature prediction system based on integrated deep learning provided by the embodiment of the application can realize real-time prediction of the winter asphalt pavement surface temperature, improve the accuracy and robustness of the surface temperature prediction, has technical features such as data driving and self-learning, and is more suitable for various complex environments and scenes such as night and adverse weather.

[0109] The system provided by the embodiment of the application is used for executing the above-mentioned method embodiments, and the specific process and detailed content are referred to the above-mentioned embodiments, which will not be described here.

[0110] Figure 5 The electronic device structure schematic diagram provided by the embodiment of the application is referred to Figure 5 The electronic device can include a processor 201, a communications interface 202, a memory 203 and a communications bus 204, wherein the processor 201, the communications interface 202 and the memory 203 complete mutual communication through the communications bus 204. The processor 201 can call the logic instructions in the memory 203 to execute the following method:

[0111] Based on the Spearman correlation coefficient and the road surface thermodynamics theory, meteorological factors which have a cumulative influence on the road surface temperature are selected;

[0112] The winter historical time series data of the meteorological factors which have a cumulative influence and the winter historical time series data of the surface temperature are cut by using the sliding window method, and a supervised learning sample is constructed;

[0113] The obtained supervised learning sample is used for training the winter asphalt pavement surface temperature prediction model based on integrated deep learning;

[0114] The winter asphalt pavement surface temperature prediction model based on integrated deep learning comprises an LSTM neural network module, a GRU neural network module, a feature fusion layer and a full connection layer.

[0115] The winter asphalt pavement surface temperature prediction model based on integrated deep learning extracts the cumulative influence features of meteorological factors by using the LSTM neural network module, extracts the periodic features of the pavement surface temperature by using the GRU neural network module, then fuses the extracted periodic features and cumulative influence features into a feature vector by using the feature fusion layer, and finally outputs the pavement surface temperature after future Dh by using two full connection layers; wherein the input of the LSTM neural network module is a cumulative influence feature matrix.

[0116] The input of the GRU neural network module is a periodic feature matrix.

[0117] Real-time meteorological factor data that has a cumulative influence on the pavement temperature is collected, and the integrated deep learning model is used to predict the highway pavement surface temperature.

[0118] In addition, the logic instructions in the above-mentioned memory 203 can be implemented in the form of a software function unit and sold or used as a stand-alone product, and can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0119] On the other hand, the embodiments of the present application also provide a non-transitory computer readable storage medium having a computer program stored thereon, which is executed by a processor to implement the winter asphalt pavement surface temperature prediction method based on integrated deep learning provided by the above-mentioned embodiments, for example, comprising:

[0120] Based on the Spearman correlation coefficient and the pavement thermodynamics theory, meteorological factors that have a cumulative influence on the pavement temperature are selected;

[0121] The winter historical time series data of the meteorological factors that have a cumulative influence and the winter historical time series data of the pavement surface temperature are divided by using the sliding window method to construct supervised learning samples;

[0122] The obtained supervised learning samples are used to train a winter asphalt pavement surface temperature prediction model based on integrated deep learning;

[0123] The winter asphalt pavement surface temperature prediction model based on integrated deep learning comprises an LSTM neural network module, a GRU neural network module, a feature fusion layer and a full connection layer.

[0124] The winter asphalt pavement surface temperature prediction model based on integrated deep learning extracts cumulative influence features of meteorological factors by using the LSTM neural network module, extracts periodic features of the surface temperature by using the GRU neural network module, then fuses the extracted periodic features and cumulative influence features into a feature vector by using the feature fusion layer, and finally outputs the surface temperature after future Ah by using two full connection layers; wherein the input of the LSTM neural network module is a cumulative influence feature matrix.

[0125] The input of the GRU neural network module is a periodic feature matrix.

[0126] Real-time meteorological factor data that has a cumulative influence on the pavement temperature is collected, and the integrated deep learning model is used to predict the highway surface temperature.

[0127] The device embodiments described above are only schematic, wherein the units shown as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment. Those skilled in the art can understand and implement without creative labor.

[0128] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary general hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0129] Application examples

[0130] The data used in the experiment is the meteorological element and road surface temperature data collected by the traffic meteorological station, the collection place is near Xinyi River Bridge of Jiangsu Ningsu Xu Expressway, and the time span is from 2016 to 2020 winter. There are problems such as data anomaly and data loss in the original highway meteorological data, in order to ensure the data quality. The noise data is judged by the Lai Da criterion, if the absolute value of the difference between the value and the average value is greater than three times the standard deviation, the value is replaced by the average value of the two values on both sides, that is, the average value of the two values on both sides. Then, in order to improve the convergence speed of the model and reduce the influence of different dimensions between different characteristics on the spatial complexity of the model, the method of z-score standardization is used for linear transformation of the original data. Finally, the sliding window method is used to cut the time series data and construct the supervised learning sample.

[0131] The integrated deep learning model and the classical time series prediction model proposed in the application are compared to verify the superiority of the integrated deep learning model to other models in performance. The models participating in the comparison include BP (backpropagation) neural network, support vector machine (Support Vector Regression, SVR) and CNN-LSTM (convolutional neural networks-long-short-term memory). For the BP network, the size of the training sample is set to 64, the architecture is 3 layers, and the epoch number is set to 200. For SVR, we use radial basis function (RBF) as the kernel function, and the penalty parameter is set to the experience value 1. The CNN-LSTM model is an integrated model of CNN neural network and LSTM neural network. For the CNN-LSTM model, the default setting is used for training. The division ratio of the training set and the test set of all models is 7:3.

[0132] To quantitatively evaluate the model performance, the mean absolute error (MAE), mean square error (MSE) and root mean square error (RMSE) are selected to evaluate the error between the predicted value and the observed value. The evaluation index is defined as follows:

[0133]

[0134]

[0135]

[0136] In the formula: y i y is the observed value of the road surface temperature, y is the predicted value of the road surface temperature, y is the average value of the observed value. The smaller the values of MAE, MSE and RMSE, the better the model prediction performance.

[0137] From Table 1, when the prediction interval is 1h, the prediction performance of the LSTM module is slightly worse than that of the ensemble deep learning model, while the performance difference between the GRU module and the ensemble deep learning model is larger. However, when the prediction interval gradually increases, the prediction performance of the ensemble deep learning model is more obvious.

[0138] Table 1 Comparison of performance of ensemble deep learning model and GRU, LSTM modules

[0139]

[0140] From Table 2, among all the prediction algorithms of all three indicators at all prediction time intervals, SVR has the largest prediction error. This is because SVR uses a kernel function to map a large amount of uncertain road surface temperature data to a high-dimensional space, which cannot fully utilize the road surface storage and release process for prediction, resulting in poor prediction performance. Compared with SVR, the BP network achieves better prediction performance, with an average reduction of 6.9%, 16.1%, and 86.4% in MAE, MSE, and MAPE, respectively. However, the BP network still cannot capture inherent features such as periodic characteristics and thermal inertia of the road surface. Compared with the above two methods, CNN-LSTM further reduces the prediction error in MAE, MSE, and MAPE. This performance improvement may be due to the fact that CNN-LSTM can extract the lag and cumulative effects of meteorological factors on road surface temperature, but it is worth mentioning that when the prediction time increases, the prediction error increases, which is due to the fact that CNN-LSTM cannot capture the periodic characteristics of road surface temperature. Unlike all the above algorithms, the proposed ensemble deep learning model (Ensemble model) can more effectively extract the thermal inertia and periodic characteristics of the road surface. Compared with CNN-LSTM, the MAE, MSE, and MAPE of the ensemble deep learning model are reduced by an average of 24.4%, 53.6%, and 65%, respectively. It can be seen that the ensemble deep learning model is superior to all the above algorithms in all three indicators.

[0141] Table 2 Comparison of prediction performance of different algorithms

[0142]

[0143] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the surface temperature of asphalt pavement in winter based on ensemble deep learning, characterized in that, Includes the following steps: Step (1): Based on Spearman correlation coefficient and pavement thermodynamics theory, select meteorological factors that have a cumulative impact on pavement temperature; Step (2): Use the sliding window method to segment the historical time series data of meteorological factors that have a cumulative impact in winter and the historical time series data of road surface temperature in winter, and construct supervised learning samples; Step (3): The supervised learning samples obtained in step (2) are used to train the winter asphalt pavement surface temperature prediction model based on ensemble deep learning. The winter asphalt pavement surface temperature prediction model based on integrated deep learning includes an LSTM neural network module, a GRU neural network module, a feature fusion layer, and a fully connected layer. The winter asphalt pavement surface temperature prediction model based on ensemble deep learning employs an LSTM neural network module to extract the cumulative influence features of meteorological factors and a GRU neural network module to extract the periodic features of the pavement surface temperature. Then, a feature fusion layer is used to fuse the extracted periodic features and cumulative influence features into a single feature vector. Finally, two fully connected layers output the future... The subsequent road surface temperature; where the input of the LSTM neural network module is the cumulative influence feature matrix; The input to the GRU neural network module is a periodic feature matrix; Step (4): Collect meteorological data that have a cumulative impact on road surface temperature in real time, and use an integrated deep learning model to predict the road surface temperature of the highway. In step (3), the cumulative influence feature matrix is ​​shown in equation (1): （1） in, n The sliding window size represents the cumulative effect, indicating the historical period before... n Hour, Indicates the time of the weather station t Observed temperature, from time tn up to the current time t Observed historical temperatures are expressed as ; Indicates the time of the weather station t Observed humidity; from time tn up to the current time t The observed historical humidity is expressed as ; Indicates the time of the weather station t Observed wind speed; from time tn up to the current time t The observed historical wind speed is expressed as ; Indicates the time of the weather station t Observed rainfall; from time tn up to the current time t Historical rainfall observed is expressed as ; The periodic characteristic matrix is ​​shown in equation (2): （2） in, m The sliding window size, influenced by periodic characteristics, characterizes the historical period. m sky, Indicates the time of the weather station t Observed road surface temperature, Indicates the time interval for prediction.

2. The method for predicting the surface temperature of asphalt pavement in winter based on ensemble deep learning according to claim 1, characterized in that, In step (1), the meteorological factors that have a cumulative impact on road surface temperature are air temperature, humidity, rainfall and wind speed.

3. The method for predicting the surface temperature of asphalt pavement in winter based on ensemble deep learning according to claim 1, characterized in that, n It is 7. m The value is 3; the loss function is MAE, the learning rate is 0.001, and the optimizer is Adam.

4. The method for predicting the surface temperature of asphalt pavement in winter based on ensemble deep learning according to claim 1, characterized in that, Assuming the GRU neural network module and the LSTM neural network module have j and k nodes respectively, the cumulative influence feature is represented as: The periodicity characteristic is represented as Then, the cumulative impact characteristics and periodic characteristics The input feature fusion module merges the features into a fused feature that includes both cumulative effects and periodicity. .

5. The method for predicting the surface temperature of asphalt pavement in winter based on ensemble deep learning according to claim 1, characterized in that, The LSTM neural network module consists of two LSTM layers, with the first LSTM layer having 100 neurons and the second LSTM layer having 200 neurons. The GRU neural network module consists of two GRU layers, with the first GRU layer having 100 neurons and the second GRU layer having 200 neurons. Both the LSTM and GRU neural network modules use tanh as the activation function. The outputs of the LSTM and GRU modules are fused through a feature fusion layer and then input into two fully connected layers for prediction.

6. The method for predicting the surface temperature of asphalt pavement in winter based on ensemble deep learning according to claim 1, characterized in that, It takes 1 hour.

7. A winter asphalt pavement surface temperature prediction system based on ensemble deep learning, characterized in that, include: The meteorological factor selection module is used to select meteorological factors that have a cumulative impact on road surface temperature based on Spearman correlation coefficient and road surface thermodynamics theory. The first processing module is used to segment the historical time series data of meteorological factors and road surface temperature in winter that have a cumulative impact using the sliding window method, and to construct supervised learning samples. The second processing module is used to train the winter asphalt pavement surface temperature prediction model based on ensemble deep learning using the supervised learning samples obtained by the first processing module. The winter asphalt pavement surface temperature prediction model based on integrated deep learning includes an LSTM neural network module, a GRU neural network module, a feature fusion layer, and a fully connected layer. The winter asphalt pavement surface temperature prediction model based on ensemble deep learning employs an LSTM neural network module to extract the cumulative influence features of meteorological factors and a GRU neural network module to extract the periodic features of the pavement surface temperature. Then, a feature fusion layer is used to fuse the extracted periodic features and cumulative influence features into a single feature vector. Finally, two fully connected layers output the future... The subsequent road surface temperature; where the input of the LSTM neural network module is the cumulative influence feature matrix; The input to the GRU neural network module is a periodic feature matrix; The road surface temperature prediction module is used to collect meteorological factor data that has a cumulative impact on road surface temperature in real time, and to predict the road surface temperature of highways using an integrated deep learning model. The cumulative influence feature matrix is ​​shown in equation (1): （1） in, n The sliding window size represents the cumulative effect, indicating the historical period before... n Hour, Indicates the time of the weather station t Observed temperature, from time tn up to the current time t Observed historical temperatures are expressed as ; Indicates the time of the weather station t Observed humidity; from time tn up to the current time t The observed historical humidity is expressed as ; Indicates the time of the weather station t Observed wind speed; from time tn up to the current time t The observed historical wind speed is expressed as ; Indicates the time of the weather station t Observed rainfall; from time tn up to the current time t Historical rainfall observed is expressed as ; The periodic characteristic matrix is ​​shown in equation (2): （2） in, m The sliding window size, influenced by periodic characteristics, characterizes the historical period. m sky, Indicates the time of the weather station t Observed road surface temperature, Indicates the time interval for prediction.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method for predicting the surface temperature of asphalt pavement in winter based on integrated deep learning as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method for predicting the surface temperature of asphalt pavement in winter based on integrated deep learning as described in any one of claims 1 to 6.

Images