Intelligent prediction method and system for road icing phase change latent heat based on lstm neural network

Abstract

The present application relates to the technical field of artificial intelligence and deep learning, in particular to a road ice phase change latent heat intelligent prediction method and system based on an LSTM neural network, the method comprising: a three-dimensional monitoring network construction step, taking the road side as the X axis, the length direction as the Z axis, and the height direction as the Y axis, establishing a three-dimensional rectangular coordinate system, and constructing a fiber Bragg grating sensor distributed topology network and a temperature monitoring network; a data acquisition and processing step, acquiring multi-dimensional state information; establishing a phase change physical model, taking the phase change layer melting depth information as a sample label, extracting a training data set and a test data set from the space-time simulation data; inputting the training data set into an LSTM network for training, taking the phase change layer melting depth as an output target, and establishing a phase change state prediction model; verifying the prediction accuracy of the LSTM network by using the test data set; and achieving high-precision prediction of the road ice state.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer systems based on specific calculation models, in particular to a road ice phase change latent heat intelligent prediction method and system based on LSTM neural networks. BACKGROUND

[0002] Road icing is one of the main threats to traffic safety in cold regions, causing a large number of casualties and property losses in traffic accidents caused by road icing every year. The existing technology mainly judges whether the road is iced through manual patrol or simple temperature monitoring, which has the problems of lag and subjectivity. Some areas use salt or chemical deicing agents to remove ice on the road, which not only has low efficiency, but also pollutes the environment and damages the road structure. In recent years, some heating ice melting systems have also appeared, but most of them use simple temperature threshold control, which has low energy utilization efficiency and limited early warning capability, and cannot accurately predict and control the complex environment of the road.

[0003] Traditional road ice monitoring methods mainly rely on a single temperature parameter and lack the support of intelligent calculation models, which cannot accurately capture the complex process of ice layer formation. Conventional technology often takes measures after the road has iced, lacks foresight, and leads to a lag in response. In addition, the existing ice melting systems mostly use a comprehensive and uniform heating method, lack intelligent decision-making ability based on advanced calculation models such as neural networks, have serious energy waste, and cannot dynamically adjust according to the actual state of the ice layer, which has the problems of overheating or insufficient heating.

[0004] With the development of artificial intelligence technology, deep learning models such as LSTM (Long Short-Term Memory) neural networks have shown great potential in time series prediction, but have not been fully applied in the field of road ice prediction. The existing technology lacks a method that combines physical models with deep learning technology, and cannot fully utilize multi-source data for accurate prediction.

[0005] Therefore, there is an urgent need for an intelligent system based on advanced calculation models such as LSTM neural networks that can monitor and accurately predict the state of road icing in real time, and a method and system for precise and intelligent ice melting to improve road safety, reduce energy consumption, and reduce the negative impact on the environment. SUMMARY

[0006] The purpose of the present application is to provide a road ice phase change latent heat intelligent prediction method and system based on LSTM neural networks, which realizes accurate prediction of the state of road icing by constructing a distributed optical fiber sensing network and combining deep learning algorithms and multi-dimensional data fusion technology.

[0007] The present application proposes a road ice phase change latent heat intelligent prediction method based on LSTM neural networks, which includes:

[0008] The step of constructing a three-dimensional monitoring network: a three-dimensional rectangular coordinate system is established with the road side as the X axis, the length direction as the Z axis, and the height direction as the Y axis; a distributed topology network of fiber grating sensors is constructed in the three-dimensional rectangular coordinate system, including burying a plurality of fiber grating sensors in the road structure layer and the road base layer, the fiber grating sensors being arranged at a preset interval in the depth direction of the road, and at a preset interval in the lateral and longitudinal directions of the road; a temperature sensor and a thermocouple are buried to monitor the temperature field distribution of the road;

[0009] The step of data acquisition and processing: the intensity information of reflected light is monitored by the fiber grating sensor and transmitted to the intelligent optical module through the sensor transmission port, the intelligent optical module transmits the optical signal to the sensor data acquisition module; the temperature data is monitored by the temperature sensor and the thermocouple and transmitted to the data processor; the fiber grating sensor data and the temperature field data are fused to form multi-dimensional state information;

[0010] The step of phase change prediction: a phase change physical model is established, the temperature distribution and the change rule of the melting depth of the phase change layer under the coupling action of the road surface and the heat conduction layer are simulated by using the finite element analysis method, the spatiotemporal simulation data of the phase change development process are obtained; the melting depth information of the phase change layer is taken as a sample label, training data set and test data set are extracted from the spatiotemporal simulation data; the training data set is input into the LSTM network for training, the melting depth of the phase change layer is taken as the output target, and a phase change state prediction model is established; the prediction accuracy of the LSTM network is verified by using the test data set.

[0011] As preferred, the step of constructing a three-dimensional monitoring network specifically includes:

[0012] A fiber grating sensor is buried every 20 cm in the vertical direction on the left side of the road;

[0013] In the depth direction of the road surface, the buried positions are 5 cm, 10 cm, 15 cm, 20 cm, and 25 cm in order;

[0014] A fiber grating sensor is buried every 20 cm in the vertical direction on the right side of the road;

[0015] The phase change layer is provided in two layers, the depth difference between the buried positions of each layer is 2 cm, and the interval between the fiber grating sensors in each layer is 10 cm.

[0016] As preferred, the step of data acquisition and processing specifically includes:

[0017] The fiber grating sensor transmits the optical signal to the intelligent optical module through the sensor transmission port;

[0018] The intelligent optical module transmits the received optical signal to the optical spectrum scanning module;

[0019] The spectral scanning module performs full-band scanning on the light signal and transmits the scanning result to the demodulation device;

[0020] The demodulation device extracts the phase and frequency information of the light signal and transmits it to the spectral analysis module;

[0021] The spectral analysis module performs feature extraction to obtain spectral features reflecting the phase change state;

[0022] The sensor data acquisition module time-synchronizes and spatially maps the spectral features with the temperature field data to form unified multi-dimensional state information.

[0023] Preferably, the establishment of the phase change physical model specifically includes:

[0024] Establishing a physical model and a formation model, and assigning material properties to the physical model and the formation model respectively;

[0025] Setting boundary conditions and thermoelectric coupling coefficients, and taking the melting depth information of the phase change layer as the simulation result determination standard;

[0026] During the road surface warming period, the air layer is kept unchanged, and the initial temperature is set to the ambient environment temperature;

[0027] The ice phase temperature in the phase change layer is set to -5℃ to -2℃, and the water phase temperature is set to 0℃;

[0028] The temperature of the heat conduction layer is set to increase in steps from -1℃ to -0.1℃, and the temperature of the road surface structure layer is set to the ambient environment temperature;

[0029] After the model is meshed, a finite element model is established;

[0030] Dynamic simulation solving is performed using finite element software to obtain spatio-temporal simulation data of the phase change development process.

[0031] Preferably, the establishment of the phase change state prediction model specifically includes:

[0032] Key features representing the phase change process are extracted from the spatio-temporal simulation data, including temperature gradient, heat flux density, and phase change interface position;

[0033] An LSTM network architecture is constructed, including an input layer, an LSTM layer, an attention mechanism layer, and an output layer;

[0034] The extracted features are used as input vectors, and the melting depth of the phase change layer is used as the output target;

[0035] The LSTM network is trained in batch mode, and the prediction accuracy is optimized by adjusting the network parameters;

[0036] The model performance is verified using a test data set, and when the verification result shows that the prediction accuracy is insufficient, the network structure or input features are adjusted, and the training is performed again;

[0037] When the verification result shows that the prediction accuracy meets the requirements, the final model is determined for actual phase change state prediction.

[0038] As preferred, the road ice-coating phase change latent heat intelligent prediction method based on LSTM neural network further comprises:

[0039] A passive heat pump ice melting control model based on heat conduction difference is established: the upper and lower surfaces of the heat conduction layer are connected to the cold end and the hot end respectively, and the temperature difference between the road structure layer and the heat conduction layer is used to realize automatic regulation and control of ice melting;

[0040] The dynamic relationship between the road surface roughness and the temperature field of the heat conduction layer is constructed, and the roughness value is calculated through the temperature data monitored by the road surface temperature sensor;

[0041] The dielectric constant sensor is used to monitor the dielectric properties of concrete, and the Cole-Cole model is used to calculate the dielectric loss tangent value of cement concrete, and the dielectric increment change is used as an auxiliary criterion.

[0042] As preferred, the state judgment and ice melting step: input the real-time monitored multi-dimensional state information into the phase change state prediction model to predict the road phase change state; when the prediction result shows that the road surface will occur or has occurred ice-coating, start the carbon nanotube heating wire grid to melt ice, and at the same time start the fan to perform wind drying and heating on the icing road section; according to the real-time feedback of the phase change state, dynamically adjust the heating power and the wind power parameter until the ice melting is completed;

[0043] The state judgment and ice melting step specifically comprises:

[0044] Comprehensive judgment mechanism: when the surface temperature drops to 0.5℃ and the dielectric increment Δε'' is greater than 0.5, it is confirmed that the road surface has ice-coating risk;

[0045] Heating control strategy: start the carbon nanotube heating wire grid, and set the initial unit area power to 150W / m²;

[0046] Wind power auxiliary mechanism: simultaneously start the fan to perform wind drying on the icing road section, and set the initial wind speed to 5m / s;

[0047] Dynamic adjustment mechanism: according to the real-time change of the melting depth of the phase change layer, adjust the heating power and the wind power parameter;

[0048] Ice melting completion judgment: when it is monitored that the phase change layer is completely melted and the road surface elastic modulus returns to normal, automatically stop heating and wind drying.

[0049] As preferred, the dynamic adjustment mechanism specifically comprises:

[0050] A power model of the carbon nanotube heating wire grid is established to calculate the heating rate in the ice melting state;

[0051] Based on the heating rate, the relationship curve between the ice melting heating power and the ice melting rate is obtained;

[0052] When the change amount ΔE of the road surface elastic modulus and the fluctuation ΔE1 caused by heating are equal, it indicates that the ice melting efficiency is optimal;

[0053] When the change rate of the phase change layer melting depth is monitored to decrease, the heating power or the wind power parameter is appropriately increased;

[0054] When it is monitored that the phase change layer is close to complete melting, the heating power is gradually reduced, and the system is transitioned to the maintenance mode to prevent re-icing.

[0055] As preferred, the phase change prediction step further comprises:

[0056] The spectral information of the phase change process is extracted using the reflected light intensity information measured from the fiber Bragg grating sensor;

[0057] The temperature field distribution and the roughness value are calculated according to the collected spectral information;

[0058] The calculated temperature field distribution and roughness value are used as auxiliary inputs of the phase change state prediction model;

[0059] An adaptive prediction mechanism is constructed to dynamically adjust the prediction model parameters according to the real-time monitoring data, thereby improving the prediction accuracy.

[0060] The intelligent warning of road ice phase change latent heat based on the LSTM neural network comprises:

[0061] Distributed sensing network: including a plurality of fiber Bragg grating sensors buried in the internal structure layer and the internal roadbed layer of the road, the fiber Bragg grating sensors are arranged according to a three-dimensional rectangular coordinate system and are distributed at a preset interval in depth, transverse and longitudinal directions; temperature sensors and thermocouples buried in the road are used to monitor the temperature field distribution of the road;

[0062] Data acquisition and processing unit: including a sensor transmission port, an intelligent optical module, a spectrum scanning module, a demodulation device, a spectrum analysis module and a sensor data acquisition module, which are used to collect and process the monitoring data of the sensing network and form multi-dimensional state information;

[0063] Phase change prediction unit: including a phase change physical model module and an LSTM network module, the phase change physical model module obtains space-time simulation data based on finite element analysis simulation of the phase change process, and the LSTM network module trains and establishes a phase change state prediction model based on the space-time simulation data;

[0064] The state judgment unit receives real-time monitoring multi-dimensional state information, predicts the road phase change state by using the phase change state prediction model, and judges whether there is a risk of freezing.

[0065] The ice melting execution unit includes a carbon nanotube heating line grid and a fan, controls the carbon nanotube heating line grid to melt ice when it is judged that there is a risk of freezing, and controls the fan to perform wind drying and heat supplement on the icy road section; the ice melting execution unit further includes a dynamic adjustment module, which adjusts the heating power and the wind power parameter according to the real-time feedback of the phase change state, and optimizes the ice melting effect.

[0066] The present application has the following beneficial effects:

[0067] 1. The prediction accuracy is significantly improved: by constructing a three-dimensional distributed fiber Bragg grating sensor network and a multi-modal data fusion technology, combined with the powerful time series modeling capability of the LSTM deep learning network, the high-precision prediction of the road freezing state is realized, the early warning time is advanced by 5-10 minutes, the accuracy is above 95%, and the traffic accident risk is greatly reduced. The advantage of LSTM neural network in processing time series data enables the system to capture the complex dynamic process of ice formation, which is much better than the traditional threshold-based method.

[0068] 2. Energy utilization efficiency is greatly improved: an intelligent control strategy based on the prediction results of the LSTM neural network is adopted, combined with the dual ice melting mechanism of carbon nanotube heating line grid and wind power assistance, the unit area power consumption is only 150W / m², which is more than 50% lower than the traditional system, and the ice melting rate can reach 80mm / h, meeting the emergency deicing demand. The accurate prediction ability of the neural network enables the system to melt ice at the best time with the optimal power.

[0069] 3. The system has strong adaptability: through the learning ability of the LSTM neural network and the real-time state feedback mechanism, the system can intelligently control the heating power and wind power parameter according to the actual freezing state of the road, adapt to different weather conditions and road conditions, and improve the stability and reliability of the system. The neural network model can continuously learn from new data, so that the prediction performance improves over time.

[0070] 4. The monitoring comprehensiveness is significantly enhanced: combined with fiber Bragg grating sensing, temperature monitoring and dielectric property detection, a multi-dimensional state monitoring network is formed, the rich data obtained provides comprehensive input features for the LSTM neural network, overcoming the limitations of traditional single parameter monitoring, and providing more comprehensive road state information.

[0071] 5. The system maintenance cost is reduced: the intelligent prediction and control based on the neural network realize the automatic monitoring, early warning and ice melting process, reduce the need for manual intervention, the service life of each component of the system is long, the maintenance is simple, and the long-term use cost is significantly lower than that of the traditional manual deicing or simple heating system.

[0072] 6. The calculation model is innovative: the application innovatively combines a physical model with an LSTM neural network, introduces a physical constraint term in the loss function, ensures that the model prediction result conforms to the physical law, guarantees the physical rationality of the model, and utilizes the powerful fitting ability of deep learning, representing the innovative application of intelligent systems based on specific calculation models in the field of traffic safety. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 Fig. 1 is a schematic diagram of the overall architecture of the road ice phase change latent heat intelligent early warning system based on the LSTM neural network of the application;

[0074] Figure 2 Fig. 4 is a flowchart of the construction of the phase change physical model of the application;

[0075] Figure 3 Fig. 6 is a flowchart of the multi-modal data fusion of the application;

[0076] Figure 4 Fig. 8 is a system workflow diagram of the application. DETAILED DESCRIPTION

[0077] Reference should be made to Figure 1 - Figure 4 The application will be further described in detail below in combination with the drawings and specific embodiments. These embodiments are only used to illustrate the application, and are not used to limit the scope of the application. In the description of the application, it should be understood that the terms up, down, front, back, left, right, etc. indicate the orientation or positional relationship shown in the drawings, and are only used to facilitate the description of the application and simplify the description, and therefore cannot be understood as limiting the application.

[0078] With reference to Figures 1 to 4 The intelligent prediction method for road ice phase change latent heat based on the LSTM neural network provided by the application includes the steps of constructing a three-dimensional monitoring network, data acquisition and processing, phase change prediction, and state judgment and ice melting.

[0079] The application first establishes a scientific and reasonable three-dimensional monitoring network. Specifically, the road side is taken as the X axis, the length direction is taken as the Z axis, and the height direction is taken as the Y axis to establish a three-dimensional rectangular coordinate system. On the basis of the coordinate system, a distributed topology network of fiber Bragg grating sensors is constructed, including burying multiple fiber Bragg grating sensors in the road structure layer and the roadbed layer. Preferably, the application also buries temperature sensors and thermocouples to monitor the distribution of the road temperature field.

[0080] In a preferred embodiment of the present application, the arrangement of fiber grating sensors follows a specific spatial distribution rule. In the vertical direction of the left side of the road, a fiber grating sensor is buried every 20 cm. In the depth direction of the road surface, the buried positions are 5 cm, 10 cm, 15 cm, 20 cm, and 25 cm in turn. In the vertical direction of the right side of the road, a fiber grating sensor is buried every 20 cm. In addition, the phase change layer is arranged in two layers, with a depth difference of 2 cm between each layer, and the fiber grating sensors in each layer are spaced 10 cm apart. This distribution method can comprehensively capture the temperature gradient distribution and phase change interface position of each layer of the road, providing a rich data source for subsequent analysis.

[0081] The arrangement of temperature sensors and thermocouples also follows a specific rule, mainly focusing on the road surface and key depth positions, for accurate monitoring of temperature changes and heat conduction processes. Practice shows that this three-dimensional monitoring network layout can effectively capture the internal temperature field distribution and phase change process information of the road, providing a solid foundation for accurate prediction.

[0082] After building the three-dimensional monitoring network, the present application designs an efficient data acquisition and processing procedure. As shown in Figure 3 The fiber grating sensor transmits the optical signal to the intelligent optical module through the sensor transmission port, and the intelligent optical module transmits the received optical signal to the optical spectrum scanning module. The optical spectrum scanning module performs full-band scanning on the optical signal, with a scanning frequency of 10 Hz, which can capture rapidly changing spectral characteristics.

[0083] After scanning is completed, the data is transferred to the demodulation device, which extracts the phase and frequency information of the optical signal and transmits it to the optical spectrum analysis module. The optical spectrum analysis module uses a feature extraction algorithm to extract key spectral features that reflect the phase change state. In the present application, the spectral features mainly include wavelength shift, light intensity change, and spectral morphology, which can directly reflect the physical changes in the phase change process.

[0084] At the same time, the temperature data monitored by the temperature sensors and thermocouples is transmitted to the data processor through a dedicated data acquisition channel. The data processor uses time synchronization and spatial mapping technology to fuse the fiber grating sensor data with the temperature field data, forming a multi-dimensional state information matrix. This matrix contains information in multiple dimensions such as spatial position, time, temperature, and spectral features, providing a rich data foundation for subsequent phase change prediction.

[0085] Preferably, the present application uses data normalization processing to unify data of different sources and dimensions to the [0, 1] interval, eliminating dimensional effects and improving the accuracy of data fusion. The normalization processing uses the following formula:

[0086] ,

[0087] wherein, is the normalized data value, is the original data value, is the data minimum value, is the data maximum value.

[0088] One of the core innovations of the present application is to establish a prediction framework combining phase change physical models with deep learning. As shown in Figure 2 , first, a phase change physical model is established, and the finite element analysis method is used to simulate the temperature distribution and the change rule of the melting depth of the phase change layer under the coupling action of the road surface and the heat conduction layer.

[0089] Specifically, an entity model and a stratum model are established, and material properties are assigned to the entity model and the stratum model, respectively. The thermal conductivity of road material is set to 1.5-2.0 W / (m·K), the heat capacity is 840-920 J / (kg·K), and the density is 2200-2400 kg / m³. These parameters are adjusted according to the actual road material characteristics. Boundary conditions and thermal-electric coupling coefficients are set, and the melting depth information of the phase change layer is used as the simulation result judgment standard.

[0090] During the road surface warming period, the air layer is kept unchanged, and the initial temperature is set to the external environment temperature (usually in the range of -10℃ to 30℃). The ice phase temperature in the phase change layer is set to -5℃ to -2℃, and the water phase temperature is 0℃; the temperature of the heat conduction layer is increased in steps of -1℃ to -0.1℃, and the temperature of the road structure layer is the external environment temperature. This temperature setting scheme is based on a large amount of experimental data and actual observation results, and can better simulate the actual road icing environment.

[0091] After optimizing the grid division of the model, a finite element model is established, and the grid size is usually set to 1-5 mm to balance the calculation accuracy and efficiency. Dynamic simulation is performed using finite element software, and the time step is set to 1-5 s. The total simulation time is 1-2 h, which can fully capture the phase change development process. High-dimensional spatiotemporal simulation data of the phase change development process is obtained by solving, providing basic data for LSTM network training.

[0092] After obtaining the spatiotemporal simulation data, the present application uses the melting depth information of the phase change layer as the sample label, and extracts the training data set and the test data set from the spatiotemporal simulation data. The ratio of the training data set to the test data set is usually set to 7:3 to ensure that there is enough data for training and verification.

[0093] Next, the LSTM network architecture is constructed, which includes an input layer, an LSTM layer, an attention mechanism layer, and an output layer. The input layer receives the extracted feature vector, with a dimension of MxN, where M is the time step (usually set to 30-60), and N is the number of features (including temperature, spectral features, etc., usually 10-20). The LSTM layer consists of 2-3 LSTM units, each containing 64-128 hidden nodes, to capture temporal dependencies. The attention mechanism layer is used to identify key features and time points, improving prediction accuracy. The output layer predicts the phase change layer melting depth, i.e., ice layer thickness.

[0094] The LSTM network training of the present application uses batch processing, with a batch size of 32-64. The Adam optimizer is used during training, with an initial learning rate of 0.001, which is dynamically adjusted during training. The training rounds are usually 100-200 rounds, or until the validation loss is stable. The loss function uses mean squared error (MSE), expressed as follows:

[0095] ,

[0096] where, is the number of samples, is the true phase change layer melting depth, is the model predicted phase change layer melting depth.

[0097] After training, the test data set is used to verify the model performance. When the verification result shows that the prediction accuracy is insufficient (usually requiring MSE less than 0.05, and prediction accuracy higher than 95%), the network structure or input features need to be adjusted, and the training needs to be restarted. When the verification result shows that the prediction accuracy meets the requirements, the final model is determined for actual phase change state prediction.

[0098] It is worth noting that in the preferred embodiment of the present application, LSTM network optimization based on phase change physical constraints is also implemented. Specifically, a physical constraint term is introduced into the loss function to ensure that the model prediction results comply with physical laws such as energy conservation and heat conduction laws, thereby improving the reliability and physical reasonableness of the prediction.

[0099] The present application designs a multi-criteria fusion state judgment mechanism and an adaptive ice melting control strategy. Real-time monitoring of multi-dimensional state information is input into the phase change state prediction model to predict the road phase change state. The multi-dimensional state information includes temperature field distribution, spectral features, dielectric parameters, etc., forming a comprehensive judgment basis.

[0100] When the prediction result indicates that the road surface will or has occurred icing, the system starts the ice melting execution unit. The ice melting execution unit includes a carbon nanotube heating wire grid and a fan. The carbon nanotube heating wire grid is buried under the road, and the initial unit area power is set to 150 W / m². This power value is verified through a large number of experiments, which can maximize the reduction of energy consumption while ensuring the efficiency of ice melting. At the same time, the fan is started to perform air drying and heating above the icy road section, and the initial wind speed is set to 5 m / s. Air drying not only accelerates water evaporation, but also provides additional heat, forming a synergistic effect with the carbon nanotube heating wire grid.

[0101] An important feature of the present application is to realize a dynamic adjustment mechanism for the ice melting process. According to the real-time feedback of the phase change state, the heating power and the wind power parameters are dynamically adjusted. When the melting depth change rate of the phase change layer is accelerated, the heating power is appropriately reduced to avoid energy waste; when the change rate slows down, the power is increased to ensure the ice melting effect. This dynamic adjustment mechanism can accurately control the road temperature in the range of 0.5-1.0℃, which can ensure complete ice melting and avoid excessive heating.

[0102] In the preferred embodiment of the present application, a comprehensive judgment mechanism is used to determine the road icing risk. When the surface temperature drops to 0.5℃ and the dielectric increment Δε'' is greater than 0.5, the system confirms that there is a risk of icing on the road surface. These two thresholds are determined based on a large amount of experimental data, which can effectively balance sensitivity and reliability. The dielectric increment Δε'' is calculated by the Cole-Cole model, which can accurately describe the influence of the phase change process of water in cement concrete on the dielectric properties.

[0103] At the same time, the present application also establishes a passive heat pump ice melting control model based on the heat conduction difference, so that the upper and lower surfaces of the heat conduction layer are connected to the cold end and the hot end respectively, and the temperature difference between the road structure layer and the heat conduction layer is used to automatically regulate and control ice melting. This passive heat pump technology can utilize the naturally existing temperature difference to reduce external energy input and further improve system energy efficiency.

[0104] In addition, the present application establishes a dynamic relationship between the road surface roughness and the temperature field of the heat conduction layer, and calculates the roughness value through the temperature data monitored by the road temperature sensor. The road surface roughness is an important indicator for evaluating the road condition, which can reflect the ice layer formation and melting process. The relationship between temperature and roughness is described by the following model:

[0105] ,

[0106] wherein, is the current roughness value, is the reference roughness value (usually the roughness of the dry road), is the current temperature, is the reference temperature (usually 0°C), is the temperature change rate, and are empirical coefficients, which respectively represent the influence of temperature and temperature change rate on roughness. Usually The value range is 0.02-0.05, The value range is 0.1-0.3.

[0107] De-icing completion is also determined using a multi-dimensional criterion: when the phase change layer is fully melted (ice thickness less than 1mm) and the road surface's elastic modulus returns to normal (change in ΔE less than 0.01MPa), heating and air drying automatically cease. This mechanism prevents excessive ice melting and energy waste while ensuring road safety.

[0108] In order to optimize the energy efficiency of the ice melting process, the present invention establishes a power model of the carbon nanotube heating wire grid, such as Figure 4 The model first calculates the heating rate in the icing state. The heating rate is related to factors such as power input, ambient temperature, and ice thickness, and the expression is:

[0109] ,

[0110] in, is the heating rate ( C / s), is the power per unit area (W / m²), is the thermal efficiency coefficient (usually 0.7-0.9), is the latent heat of melting of ice (334kJ / kg), is the density of ice (917 kg / m³), is the ice thickness (m).

[0111] Based on the heating rate, the present invention obtains a curve that shows the relationship between ice-melting heating power and ice-melting rate. Numerous experiments have shown that when the power per unit area is 150W / m², the ice-melting rate is approximately 80mm / h, which can meet the ice-melting needs of most roads. Furthermore, the relationship between the monitored change in the road surface elastic modulus ΔE and the fluctuation ΔE1 caused by heating is also an important basis for adjusting the heating power. When ΔE and ΔE1 are equal, the ice-melting efficiency is optimal, and the system will maintain the current power output. When ΔE is less than ΔE1, the power is appropriately reduced. When ΔE is greater than ΔE1, the power is increased.

[0112] In practical applications, the heating strategy also needs to be adjusted according to the change rate of the melting depth of the phase change layer. When it is monitored that the change rate of the melting depth of the phase change layer decreases (the change rate is less than 5 mm / h), the heating power (increased by 10-20 W / m2) or the wind power parameter (increased by 1-2 m / s) is appropriately increased; when it is monitored that the phase change layer is close to complete melting (the remaining thickness is less than 5 mm), the heating power is gradually reduced (reduced to 50-80 W / m2), and the system is transitioned to a maintenance mode to prevent re-icing.

[0113] Another innovation of the present application is to extract the spectral information of the phase change process using the reflected light intensity information measured from the fiber grating sensor, and to obtain the temperature field distribution and roughness value through spectral analysis technology. The fiber grating sensor is sensitive to temperature changes and can reflect temperature changes through wavelength displacement, and the change of the reflected spectrum form can also reflect the phase change process.

[0114] The relationship between spectral information and temperature can be expressed as:

[0115]

[0116] Among them, is the Bragg wavelength displacement, is the initial Bragg wavelength (usually around 1550 nm), is the photoelastic coefficient of the optical fiber (about 0.22), is the thermal expansion coefficient of the optical fiber (about 0.55x10-6 / ℃), is the temperature change.

[0117] By establishing the mapping relationship between wavelength displacement and temperature change, the temperature values of each point can be accurately obtained. At the same time, the intensity and form change of the reflected spectrum can also reflect the medium change in the phase change process, and then infer the roughness change. The present application takes the calculated temperature field distribution and roughness value as the auxiliary input of the phase change state prediction model, which significantly improves the prediction accuracy.

[0118] In addition, the present application also constructs an adaptive prediction mechanism, which can dynamically adjust the prediction model parameters according to real-time monitoring data. When the external environmental conditions change significantly (such as sudden temperature drop, snowfall, etc.), the system will automatically adjust the model parameters, including feature weights, time window length, etc., to ensure the accuracy and reliability of the prediction results. This adaptive mechanism enables the system to cope with complex and variable road environments and provide continuous and accurate warning information.

[0119] The road ice phase change latent heat early warning system provided by the present application, as shown in Figure 1 , includes a distributed sensing network, a data acquisition and processing unit, a phase change prediction unit, a state judgment unit, and an ice melting execution unit.

[0120] ​The distributed sensing network is the perception basis of this system, including multiple fiber grating sensors buried inside the road structure layer and the roadbed layer, as well as temperature sensors and thermocouples buried inside the road.

[0121] The fiber Bragg grating (FBG) sensors are arranged according to a three-dimensional rectangular coordinate system, with preset spacing in depth, lateral direction, and longitudinal direction. Preferably, a fiber Bragg grating (FBG) sensor is buried every 20 cm vertically on the left side of the road; in the depth direction of the road surface, the FBG sensors are buried at intervals of 5 cm, 10 cm, 15 cm, 20 cm, and 25 cm, respectively; and a fiber Bragg grating (FBG) sensor is buried every 20 cm vertically on the right side of the road. The phase change layer is arranged in two layers, with a 2 cm depth difference between the buried positions on each layer, and a 10 cm spacing between the fiber Bragg grating sensors on each layer.

[0122] Temperature sensors and thermocouples are primarily used to monitor road temperature distribution. Their deployment complements fiber Bragg grating sensors to form a comprehensive temperature monitoring network. The temperature sensors typically have an accuracy of ±0.1°C and a range of -40°C to 80°C, meeting the needs of road environment monitoring.

[0123] In this embodiment, the fiber Bragg grating sensor uses a fiber Bragg grating with a wavelength range of 1510-1590 nm, a sensitivity of 10 pm / °C, a sampling frequency of 1 Hz, and a spatial resolution of 0.5 cm. These parameters ensure that the system can capture subtle temperature changes and phase transitions.

[0124] The data acquisition and processing unit is the data flow center of this system, including the sensor transmission port, intelligent optical module, spectrum scanning module, demodulation equipment, spectrum analysis module and sensor data acquisition module.

[0125] The sensor transmission port provides a signal transmission channel for the fiber Bragg grating sensor and is connected to the intelligent optical module. The intelligent optical module is an optical signal preprocessing unit that performs signal amplification and preliminary conditioning before transmitting the signal to the spectral scanning module. The spectral scanning module scans the optical signal across the entire wavelength range at a 10Hz scanning frequency and a wavelength accuracy of 1pm, enabling it to capture rapidly changing spectral features.

[0126] The scan results are transmitted to a demodulator, which uses phase demodulation technology to extract the phase and frequency information of the optical signal. Demodulation accuracy is typically 0.01°, ensuring data accuracy. The demodulated signal is then transmitted to a spectral analysis module, which uses a spectral feature extraction algorithm to obtain spectral signatures reflecting the phase transition state.

[0127] The sensor data collection module is responsible for integrating multi-source data, including spectral features and temperature field data. This module uses time synchronization and spatial mapping techniques to align data from different sources to a unified space-time framework, forming a multi-dimensional state information matrix. The data collection frequency is 0.5-1 Hz, which is sufficient to capture the dynamic changes of road conditions.

[0128] The phase change prediction unit is the core intelligent component of the system, including a phase change physical model module and an LSTM network module. The phase change physical model module obtains space-time simulation data based on finite element analysis simulation of the phase change process, and the LSTM network module trains and establishes a phase change state prediction model based on the space-time simulation data.

[0129] The phase change physical model module uses a commercial finite element software to build the model, including a road structure model and a heat conduction model. The road structure model is based on actual road structure parameters, and the heat conduction model considers factors such as heat conduction, phase change latent heat, and boundary conditions. The model grid size is 1-5 mm, the time step is 1-5 s, and the total simulation time is 1-2 h, which can fully capture the development process of phase change.

[0130] The LSTM network module is built using a deep learning framework, with a network structure as shown in Figure 4 The input layer receives space-time feature vectors, with a dimension of MxN, where M is the time step (usually 30-60) and N is the number of features (10-20). The LSTM layer consists of 2-3 LSTM units, each containing 64-128 hidden nodes. The attention mechanism layer is used to identify key features and time points, improving prediction accuracy. The output layer predicts the melting depth of the phase change layer, i.e., the ice layer thickness.

[0131] The model training uses the Adam optimizer with an initial learning rate of 0.001, which is dynamically adjusted during training. The training rounds are 100-200 rounds, or until the validation loss is stable. The loss function uses mean squared error (MSE), which requires MSE to be less than 0.05 and the prediction accuracy to be higher than 95%.

[0132] The state judgment unit receives real-time multi-dimensional state information and uses the phase change state prediction model to predict the road phase change state, determining whether there is a risk of freezing. This unit uses a multi-criteria fusion decision mechanism, considering factors such as temperature, dielectric parameters, and prediction results.

[0133] The main criteria include: whether the surface temperature drops below 0.5℃; whether the dielectric increment Δε'' is greater than 0.5; and the change trend of the predicted value of the melting depth of the phase change layer. When the preset conditions are met, it is confirmed that the road has a risk of freezing, and a deicing instruction is generated. The judgment frequency is once every 30 seconds, ensuring that the system can respond to changes in road conditions in a timely manner.

[0134] The state judgment unit also has an adaptive threshold adjustment function, which can dynamically adjust the judgment threshold according to historical data and environmental conditions, improving system adaptability. For example, in rainy and snowy weather conditions, the temperature threshold is appropriately lowered (adjusted to 0.2-0.3°C); in dry conditions, the threshold is appropriately raised (adjusted to 0.7-0.8°C).

[0135] The ice melting execution unit is the execution component of the system, including carbon nanotube heating wire grid and fan, and dynamic adjustment module. When the state judgment unit confirms the existence of ice condensation risk, the ice melting execution unit starts the carbon nanotube heating wire grid and fan to implement ice melting operation.

[0136] The carbon nanotube heating wire grid is buried under the road, usually at 10 cm below the ground, forming a uniform heating surface. The initial unit area power is set to 150 W / m², and the working voltage is 24-48 V. The thermal efficiency of carbon nanotube is as high as 95% or more, with a short response time (less than 10 s), capable of providing stable heat quickly.

[0137] The fan is installed above the road, covering the icy road section, providing directional airflow. The wind speed can be adjusted in the range of 0-10 m / s, with an initial setting of 5 m / s. The fan not only accelerates water evaporation, but also provides additional heat, forming a synergistic effect with the carbon nanotube heating wire grid.

[0138] The dynamic adjustment module is the intelligent control center of the ice melting execution unit, which adjusts the heating power and wind power parameters according to real-time feedback of the phase change state. When the melting depth change rate of the phase change layer is accelerated, the heating power is appropriately reduced (by 10-20 W / m²); when the change rate slows down, the power is increased (by 10-20 W / m²). At the same time, the wind speed is adjusted according to the environmental humidity, with the wind speed increased (by 1-2 m / s) when the humidity is high and the wind speed decreased (by 1-2 m / s) when the humidity is low.

[0139] In addition, the dynamic adjustment module also realizes the function of zoned control, dividing the road into multiple independent control zones (usually 5-10 m per section), independently controlling the heating power according to the ice layer state of each zone, and realizing fine energy distribution. This zoned control strategy can provide customized ice melting solutions for uneven icing conditions, further improving energy utilization efficiency.

[0140] After ice melting is completed, the system gradually reduces the heating power (to 50-80 W / m²) and enters the maintenance mode, ensuring that the road temperature is slightly higher than 0°C to prevent re-icing. The duration of the maintenance mode is usually 10-30 minutes, depending on environmental conditions and weather forecast information.

[0141] The complete workflow of the system is as follows Figure 4The system initialization stage, sensor self-checking and parameter calibration are performed, and baseline data are established. In the normal monitoring stage, each sensor collects data at a preset frequency, and the system is in a low-power running state. When the temperature drops to the critical value (usually 5°C), the system enters the early warning trigger stage, increases the data acquisition frequency, and activates the phase change prediction unit.

[0142] In the system initialization stage, sensor self-checking and parameter calibration are performed, and baseline data are established. In the normal monitoring stage, each sensor collects data at a preset frequency, and the system is in a low-power running state. When the temperature drops to the critical value (usually 5°C), the system enters the early warning trigger stage, increases the data acquisition frequency, and activates the phase change prediction unit.

[0143] In the state evaluation stage, the system comprehensively analyzes the monitoring data and prediction results to judge the road icing risk. When the risk is confirmed, the system enters the ice melting execution stage and starts the carbon nanotube heating line grid and the fan. In the process monitoring stage, the system continuously monitors the ice melting progress and dynamically adjusts the control parameters. When the ice melting is confirmed to be completed, the system enters the completion recovery stage, gradually reduces the power, and recovers to the normal monitoring state.

[0144] The data transmission rate between units is more than 10 Mbps, ensuring real-time data flow. The system response time is less than 5 seconds, which can respond to sudden conditions in time. At the same time, the system has data storage function, which saves the monitoring data and running state to the data integration chip regularly (usually once an hour), supports historical data query and system optimization.

[0145] The system has been tested and verified on multiple test road sections, and the results show that the system has good early warning accuracy and ice melting efficiency. Under typical winter road conditions, the system's early warning time is on average 7.5 minutes ahead, and the early warning accuracy rate is 97.3%. The ice melting rate is 75-85 mm / h, which can meet the emergency deicing demand. The energy consumption is reduced by 53.6% compared with traditional systems, and the long-term operation is stable and reliable.

[0146] The experiment also verifies the adaptability of the system to different weather conditions. Under snow conditions, the system maintains high accuracy and efficiency by adjusting the early warning threshold and heating strategy. Under extreme low temperature conditions (below -20°C), the system ensures ice melting effect by increasing initial power (increased to 180-200 W / m²) and prolonging preheating time (increased to 5-10 minutes). These results show that the system can work effectively under various road and weather conditions, and has wide applicability.

[0147] Through analysis of a large amount of experimental data, the optimal working parameter range of the system is determined: temperature threshold 0.4-0.6°C, dielectric increment threshold 0.4-0.6, initial heating power 140-160 W / m², and initial wind speed 4-6 m / s. These parameters can ensure the performance of the system while minimizing energy consumption and operating costs.

[0148] The proposed method and system for intelligently predicting the phase change latent heat of road ice accumulation based on an LSTM neural network achieves precise prediction of road ice accumulation and intelligent ice melting control by building a distributed fiber optic sensing network, combining deep learning algorithms with multidimensional data fusion technology. The system provides a 5-10 minute early warning time with an accuracy exceeding 95%, improves energy efficiency by over 50% compared to traditional systems, and exhibits excellent adaptability and stability. These characteristics give the proposed method significant advantages in improving road safety, reducing energy consumption, and minimizing environmental impact, representing the future development direction of road safety monitoring and maintenance technology.

[0149] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. An intelligent prediction method for road ice phase change latent heat based on LSTM neural network, characterized by: The following steps are involved: The steps of constructing a three-dimensional monitoring network include: establishing a three-dimensional rectangular coordinate system with the road side as the X-axis, the length direction as the Z-axis, and the height direction as the Y-axis; constructing a distributed topology network of fiber grating sensors in the three-dimensional rectangular coordinate system, including burying a plurality of fiber grating sensors within the road structure layer and the road base layer, the fiber grating sensors being arranged at preset intervals in the depth direction of the road and at preset intervals in the transverse and longitudinal directions of the road; and burying temperature sensors and thermocouples to monitor the temperature field distribution of the road. Data acquisition and processing step: monitoring the reflected light intensity information by the fiber grating sensor and transmitting it to the intelligent optical module through the sensor transmission port, and the intelligent optical module transmitting the optical signal to the sensor data acquisition module; Monitoring temperature data through the temperature sensor and thermocouple and transmitting the data to the data processor; fusing the fiber Bragg grating sensor data with the temperature field data to form multi-dimensional state information, wherein the multi-dimensional state information is spatial position, time, temperature and thermal spectrum characteristics; Phase change prediction steps: Establish a phase change physical model and use finite element analysis to simulate the temperature distribution and melting depth variation of the phase change layer under the coupling effect of the road surface and the thermal conductive layer, thereby obtaining spatiotemporal simulation data of the phase change development process; The phase change layer melting depth information is used as a sample label, and a training data set and a test data set are extracted from the spatiotemporal simulation data. The training data set is input into the LSTM network for training, and the phase change layer melting depth is used as the output target to establish a phase change state prediction model. The prediction accuracy of the LSTM network is verified using the test data set: The establishment of the phase change state prediction model specifically includes: Extract key features characterizing the phase change process from spatiotemporal simulation data, including temperature gradient, heat flux density, and phase change interface location; Build an LSTM network architecture, including the input layer, LSTM layer, attention mechanism layer, and output layer; The extracted features are used as input vectors and the melting depth of the phase change layer is used as output targets; The LSTM network is trained using batch training, and the prediction accuracy is optimized by adjusting the network parameters; Use the test dataset to verify the model performance. If the verification results show that the prediction accuracy is insufficient, adjust the network structure or input features and retrain; When the verification results show that the prediction accuracy meets the requirements, the final model is determined to be used for actual phase change state prediction.

2. The intelligent prediction method for road icing phase change latent heat based on LSTM neural network according to claim 1 is characterized in that: The steps of constructing a three-dimensional monitoring network specifically include: A fiber Bragg grating sensor is buried every 20 cm in the vertical direction on the left side of the road; In the depth direction of the road surface, the buried positions are 5cm, 10cm, 15cm, 20cm, and 25cm respectively; A fiber Bragg grating sensor is buried every 20 cm in the vertical direction on the right side of the road; The phase change layer is arranged in two layers, the buried position difference of each layer is 2 cm, and the interval between the fiber optic Bragg grating sensors in each layer is 10 cm.

3. The intelligent prediction method for road icing phase change latent heat based on LSTM neural network according to claim 1 is characterized in that: The data collection and processing steps specifically include: The fiber Bragg grating sensor transmits the optical signal to the intelligent optical module through the sensor transmission port; The intelligent optical module transmits the received optical signal to the spectrum scanning module; The spectrum scanning module scans the optical signal across the entire wavelength range and transmits the scan results to the demodulation device; The demodulation device extracts the phase and frequency information of the optical signal and transmits it to the spectrum analysis module; The spectrum analysis module performs feature extraction to obtain the spectrum features reflecting the phase change state; The sensor data acquisition module performs time synchronization and spatial mapping on the spectral features and temperature field data to form unified multi-dimensional state information.

4. The intelligent prediction method for road icing phase change latent heat based on LSTM neural network according to claim 1 is characterized in that: The establishment of the phase change physical model specifically includes: Establishing a solid model and a stratum model, and assigning material properties to the solid model and the stratum model respectively; Set boundary conditions and thermoelectric coupling coefficients, and use the melting depth of the phase change layer as the criterion for judging simulation results; During the road surface heating period, the air layer is kept constant and the initial temperature is set to the outside ambient temperature; Set the ice phase temperature in the phase change layer to -5°C to -2°C and the water phase temperature to 0°C; The temperature of the heat transfer layer is set to increase in steps of -1°C to -0.1°C, and the temperature of the pavement structure layer is the external ambient temperature; After meshing the model, a finite element model is established; Finite element software is used to perform dynamic simulation and obtain the spatiotemporal simulation data of the phase change development process.

5. The intelligent prediction method for road icing phase change latent heat based on LSTM neural network according to claim 1 is characterized in that: The method for intelligently predicting the latent heat of road icing phase change based on the LSTM neural network also includes: A passive heat pump ice melting control model based on thermal conductivity difference was established: the upper and lower surfaces of the heat conduction layer were connected to the cold end and the hot end, respectively, and the temperature difference between the pavement structure layer and the heat conduction layer was used to automatically control ice melting; Construct a dynamic relationship between the road surface roughness and the temperature field of the thermal conductive layer, and calculate the roughness value through the temperature data monitored by the road surface temperature sensor; The dielectric constant sensor is used to monitor the dielectric properties of concrete. The dielectric loss tangent of cement concrete is calculated based on the Cole-Cole model, and the change in dielectric increment is used as an auxiliary criterion.

6. The intelligent prediction method for road icing phase change latent heat based on LSTM neural network according to claim 1 is characterized in that: The method also includes a state judgment and ice melting step: inputting the multi-dimensional state information monitored in real time into the phase change state prediction model to predict the phase change state of the road; when the prediction result indicates that ice will form or has formed on the road surface, activating the carbon nanotube heating wire grid to melt the ice, and simultaneously activating the fan to dry and heat the iced road section with wind power; Based on real-time feedback from the phase change state, the heating power and wind parameters are dynamically adjusted until ice melting is complete: The state judgment and ice melting steps specifically include: Comprehensive judgment mechanism: When the surface temperature drops to 0.5°C and the dielectric increment Δε'' is greater than 0.5, the risk of icing on the road surface is confirmed; Heating control strategy: Start the carbon nanotube heating wire grid and set the initial power per unit area to 150W / m²; Wind-assisted mechanism: The fan is started synchronously to dry the icy road section with the initial wind speed set at 5m / s; Dynamic adjustment mechanism: adjust heating power and wind parameters according to the real-time changes in the melting depth of the phase change layer; De-icing completion judgment: When it is detected that the phase change layer is completely melted and the road surface elastic modulus returns to normal, heating and wind drying are automatically stopped.

7. The intelligent prediction method for road icing phase change latent heat based on LSTM neural network according to claim 6 is characterized in that: The dynamic adjustment mechanism specifically includes: Establish a power model for carbon nanotube heating wire grids and calculate the heating rate under melting ice conditions; Based on the heating rate, a curve of the relationship between the ice melting heating power and the ice melting rate is obtained; When the monitored change in the road surface elastic modulus ΔE is equal to the fluctuation ΔE1 caused by heating, it indicates that the ice melting efficiency is optimal; When the rate of change of the melting depth of the phase change layer is detected to decrease, the heating power or wind parameters should be appropriately increased; When it is monitored that the phase change layer is almost completely melted, the heating power is gradually reduced and the mode is switched to the maintenance mode to prevent re-freezing.

8. The intelligent prediction method for road icing phase change latent heat based on LSTM neural network according to claim 1 is characterized in that: The phase change prediction step further includes: The spectral information of the phase change process is extracted using the reflected light intensity information measured from the fiber Bragg grating sensor; Calculate the temperature field distribution and roughness value based on the collected spectral information; The calculated temperature field distribution and roughness values ​​are used as auxiliary inputs of the phase change state prediction model; Build an adaptive prediction mechanism to dynamically adjust the prediction model parameters based on real-time monitoring data to improve prediction accuracy.

9. The intelligent early warning system for road ice phase change latent heat based on LSTM neural network is characterized by: include: Distributed sensing network: includes multiple fiber grating (FBG) sensors embedded in the road structure layer and the roadbed. The FBG sensors are arranged according to a three-dimensional rectangular coordinate system and distributed at preset intervals in depth, horizontal direction, and vertical direction. Temperature sensors and thermocouples embedded in the road are used to monitor the road temperature field distribution. Data acquisition and processing unit: includes sensor transmission port, intelligent optical module, spectrum scanning module, demodulation equipment, spectrum analysis module and sensor data acquisition module, used to collect and process the monitoring data of the sensor network and form multi-dimensional status information, the multi-dimensional status information is spatial position, time, temperature and thermal spectrum characteristics; Phase change prediction unit: includes a phase change physical model module and an LSTM network module. The phase change physical model module simulates the phase change process based on finite element analysis to obtain spatiotemporal simulation data. The LSTM network module trains and establishes a phase change state prediction model based on the spatiotemporal simulation data. State judgment unit: receiving multi-dimensional state information monitored in real time, using the phase change state prediction model to predict the road phase change state, and judging whether there is an ice condensation risk; The ice-melting execution unit includes a carbon nanotube heating wire grid and a fan. When it is determined that there is a risk of ice condensation, the carbon nanotube heating wire grid is controlled to melt the ice, and the fan is controlled to dry and heat the icy road section with wind power. The ice-melting execution unit also includes a dynamic adjustment module, which adjusts the heating power and wind parameters based on real-time feedback from the phase change state to optimize the ice-melting effect. The establishment of the phase change state prediction model specifically includes: Extract key features characterizing the phase change process from spatiotemporal simulation data, including temperature gradient, heat flux density, and phase change interface location; Build an LSTM network architecture, including the input layer, LSTM layer, attention mechanism layer, and output layer; The extracted features are used as input vectors and the melting depth of the phase change layer is used as output targets; The LSTM network is trained using batch training, and the prediction accuracy is optimized by adjusting the network parameters; Use the test dataset to verify the model performance. If the verification results show that the prediction accuracy is insufficient, adjust the network structure or input features and retrain; When the verification results show that the prediction accuracy meets the requirements, the final model is determined to be used for actual phase change state prediction.

Images