Icing prediction method and system based on multi-source data fusion

Abstract

The invention provides an icing prediction method and system based on multi-source data fusion, and relates to the technical field of data processing, and the method comprises the steps: constructing a historical data tensor field, and constructing a real-time query tensor based on a real-time snow-melting agent residual efficacy value, calculating the similarity score of the real-time query tensor and the historical data tensor field to form a historical data subset; forming a discrete point cloud based on the historical data subset; generating geometric distribution index identification evolution states, and obtaining a freezing point based on each evolution state; generating a physical icing probability according to the freezing point and the ground temperature; generating a predicted icing probability based on a pre-constructed machine learning model; and generating a final icing probability. Accurate quantification of the real ice-resistant capability of the road surface is realized, and the accuracy and reliability of icing probability prediction under complex meteorological conditions are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for predicting icing based on multi-source data fusion. Background Technology

[0002] Winter road icing is one of the main causes of serious traffic accidents, and accurate early warning of road icing is a core aspect of road monitoring and maintenance. However, existing road icing prediction schemes have a fundamental technical flaw in practical applications: the prediction models rely solely on real-time meteorological data, completely ignoring the dynamic changes in the residual effectiveness of existing de-icing agents on the road surface, thus failing to accurately determine the road's true anti-icing capacity.

[0003] Specifically, existing methods cannot identify the specific stage of the residual effectiveness of de-icing agents. Specifically, they cannot distinguish whether the road surface is in a "dissolving state" where salt is rapidly taking effect, a "saturated transition state" where the concentration has reached critical equilibrium, an "accumulated state" where salt has been depleted or ineffective, or an "uncertain state" severely disturbed. This leads to the model's inability to differentiate the significant differences in actual icing risk across different road surfaces under the same weather conditions. This lack of understanding of the road surface's internal state directly results in severely distorted icing probability predictions: often falsely reporting icing risk when salt is still effective, or failing to report icing hazards when salt has already ineffective. Due to the lack of dynamic feedback on the actual evolution of the road surface, existing prediction systems struggle to adapt to complex and changing weather disturbances and traffic flow effects. In extreme weather or long-duration snow and ice disasters, the accuracy and reliability of their predictions significantly decrease. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art.

[0005] To achieve the above objectives, this application proposes an icing prediction method and system based on multi-source data fusion, comprising the following steps: Step 1, obtain the ground temperature; Step 2: Set the initial effectiveness value, collect cumulative rainfall and cumulative traffic flow data in real time, and calculate the total loss ratio; obtain the real-time residual effectiveness value of the snow melting agent based on the difference between the initial effectiveness value and the total loss ratio. Step 3: Construct a historical data tensor field and a real-time query tensor based on the real-time residual effectiveness value of the snow melting agent. Calculate the similarity score between each real-time query tensor and the historical data tensor field. Sort all similarity scores in descending order and select the top k groups of historical data tensor fields with the highest similarity scores to form a subset of similar historical data. Map this subset of historical data onto a two-dimensional phase space with the historical snow melting agent effectiveness value on the horizontal axis and the historical measured freezing point on the vertical axis to form a discrete point cloud. Generate several geometric distribution indices based on the discrete point cloud. Identify the evolution state based on the geometric distribution indices. Based on each evolution state and the relationship between the geometric distribution indices and a preset target value, obtain the freezing point. Step 4: Generate the physical icing probability based on the freezing point and ground temperature; construct a forecast feature vector based on real-time weather forecast data, and input the forecast feature vector into a pre-constructed machine learning model to generate a predicted icing probability; Step 5: Generate the final icing probability based on the physical icing probability and the predicted icing probability.

[0006] Furthermore, historical meteorological feature vectors, historical residual effectiveness values ​​of snow melting agents, and historical measured freezing points are obtained, and together they constitute a multidimensional historical data tensor field.

[0007] Furthermore, real-time meteorological feature vectors are obtained, and the real-time residual effectiveness value of the snow melting agent is normalized to obtain a scalar value. The scalar value is then concatenated with the real-time meteorological feature vector to generate a real-time query tensor.

[0008] Furthermore, the geometric distribution indicators include: linear concentration index, distribution dispersion index, evolution trend direction index, and end-bending characteristic index; Calculate the covariance matrix of the discrete point cloud, solve for the maximum and minimum eigenvalues ​​of the covariance matrix, construct the feature ratio based on the maximum and minimum eigenvalues, and calculate the linear concentration index based on the feature ratio. Calculate the centroid coordinates of the discrete point cloud, calculate the Euclidean distance from each data point in the discrete point cloud to the centroid coordinates, and obtain the arithmetic mean of all Euclidean distances. Use the arithmetic mean as an index of the degree of dispersion of the distribution. Find the eigenvector corresponding to the largest eigenvalue in the covariance matrix; calculate the slope of the eigenvector and use the slope as an indicator of the direction of evolution trend. The historical data subset is sorted by timestamp from smallest to largest, and the last m data points after sorting are selected; the m data points are fitted into a quadratic curve equation; the absolute value of curvature of the quadratic curve equation at the real-time residual effectiveness value of the de-icing agent is calculated, and the end bending characteristic index is obtained after normalizing the absolute value of curvature.

[0009] Furthermore, if the linear concentration index is greater than the first preset target value and the evolution trend direction index is negative, then it is in a dissolving state. A similarity calculation is performed on the historical data subset to obtain a similarity value. The similarity value is used as the basic weight, and a weighted average is calculated with the historical measured freezing point of each historical data pair in the historical data subset and used as the freezing point.

[0010] Furthermore, if the distribution dispersion index is greater than the second preset target value and the end bending characteristic index is greater than the third preset target value, then it is in a saturated precipitation transition state, and the upper quartile of all historical measured freezing points of the historical data pairs of the historical data subset is taken as the freezing point.

[0011] Furthermore, if the evolution trend direction index is greater than or equal to zero and the linear concentration index is less than the fourth preset target value, then it is in a state of failure accumulation, and the maximum value of all historical measured freezing points of the historical data pairs of the historical data subset is taken as the freezing point.

[0012] Furthermore, a first weight and a second weight are preset, and the final icing probability is obtained by weighted summation based on the physical icing probability, the first weight, the predicted icing probability, and the second weight.

[0013] This embodiment also proposes an icing prediction system based on multi-source data fusion, which includes the following modules: First data acquisition module: used to acquire ground temperature; The second data acquisition module is used to construct a historical data tensor field and a real-time query tensor based on the real-time residual effectiveness value of the snow melting agent. It calculates the similarity score between each real-time query tensor and the historical data tensor field, sorts all similarity scores in descending order, and selects the top k groups of historical data tensor fields with the highest similarity scores to form a subset of similar historical data. This subset of historical data is then mapped onto a two-dimensional phase space with the historical snow melting agent effectiveness value on the horizontal axis and the historical measured freezing point on the vertical axis, forming a discrete point cloud. Several geometric distribution indices are generated based on the discrete point cloud. The evolutionary state is identified according to the geometric distribution indices, and the freezing point is obtained based on the relationship between the geometric distribution indices and a preset target value for each evolutionary state. Data processing module: used to construct a historical data tensor field and a real-time query tensor based on the real-time residual effectiveness value of the snow melting agent; to determine the evolution state based on the real-time query tensor and the historical data tensor field; and to generate the freezing point based on the evolution state. Prediction module: used to generate physical icing probability based on the freezing point and ground temperature; construct forecast feature vector based on real-time weather forecast data, and input the forecast feature vector into a pre-constructed machine learning model to generate predicted icing probability; Fusion module: used to generate the final icing probability based on the physical icing probability and the predicted icing probability.

[0014] Compared with existing technologies, this application provides an icing prediction method and system based on multi-source data fusion. By constructing a historical data tensor field and matching the residual effectiveness of de-icing agents in real time, it accurately identifies four evolution states of the road surface. Based on the physicochemical characteristics of different states, it dynamically selects specific statistical quantities in the geometric distribution index as the real-time freezing point, thereby fundamentally solving the prediction distortion problem caused by ignoring the dynamic changes in salinity in traditional methods. In a stable dissolved state, salt dissolves efficiently in the water film, resulting in the strongest and most uniform anti-icing ability of the road surface. At this point, the judgment logic is based on a high linear concentration index and a negative evolution trend. A weighted average is used as the freezing point to reflect the overall low freezing point characteristic under uniform salt distribution, avoiding false alarms of icing when salt is still effective. When entering the saturated precipitation transition state, the solution approaches critical equilibrium, and local salt precipitation may cause fluctuations in the lower limit of anti-icing ability. At this point, the judgment logic is based on high distribution dispersion and high end-bending characteristics, extracting the upper quartile as the freezing point to capture the higher-risk tail data in the distribution and prevent missed alarms caused by uneven local concentration. When in a failed accumulation state, salt has been depleted or only physical accumulation remains, losing its chemical anti-icing effect, and the road surface reverts to pure water icing characteristics. At this point, the judgment logic is based on a non-negative evolution trend and extremely low linear concentration. The maximum value is directly extracted as the freezing point to ensure that the highest level of icing warning is triggered immediately when salt is completely ineffective, eliminating safety hazards. In a mixed and uncertain state with severely interfered data, a robust median is used as the freezing point to avoid extreme noise interference. This mechanism, which uses state recognition to drive dynamic switching of statistics, allows freezing point calculation to no longer rely on fixed empirical values, but rather to dynamically evolve throughout the entire life cycle of the de-icing agent, from efficient dissolution to failure and accumulation. This enables precise quantification of the road surface's true anti-icing capability and significantly improves the accuracy and reliability of icing probability prediction under complex weather conditions. Attached Figure Description

[0015] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating an ice-prediction method based on multi-source data fusion provided in this application embodiment; Figure 2 A structural block diagram of an icing prediction system based on multi-source data fusion provided in an embodiment of this application; Figure 3 This is a block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0016] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0017] The following description, with reference to the accompanying drawings, illustrates an embodiment of an icing prediction method and system based on multi-source data fusion.

[0018] like Figure 1 As shown, an icing prediction method based on multi-source data fusion includes the following steps: Step 1: Obtain the ground temperature.

[0019] This embodiment uses a set of historical surface temperature observations for the same period (e.g., the same time on the same day over the past 5 years) for this road segment, and calculates its weighted average or median as the surface temperature for the predicted time of the target road segment. Surface temperature is strictly constrained by solar radiation, Earth's orbital parameters, and seasonal climate models, exhibiting extremely high temporal autocorrelation and annual cycle recurrence. Under normal circumstances excluding extreme and sudden weather events, the solar altitude angle, sunshine duration, and background climate conditions received by a specific road segment on the same date and time each year are highly similar, resulting in significant repetitive characteristics in its surface heat budget process. This makes historical data from the same period the most reliable prior distribution for predicting short-term temperatures in the future. Secondly, by aggregating observations from the same period over many years (e.g., 5 years) to construct a dataset, the random noise and accidental bias of a single year can be effectively smoothed out, and the law of large numbers can be used to restore the normal climate at that spatiotemporal point. Furthermore, using a weighted average can assign higher weight to recent years based on the principle of time decay, in order to capture long-term warming or cooling trends; while using the median can naturally eliminate extreme outliers in historical data, avoiding distortion of prediction results by extreme cold or heat cases, thus providing a more representative and robust temperature benchmark.

[0020] Step 2: Obtain the real-time residual effectiveness value of the snow melting agent.

[0021] In this embodiment, an initial set efficacy value is first preset, derived from laboratory standard operating condition calibration. This means that in a controlled environment, for a specific type of de-icing agent (such as calcium chloride or compound salt), at a specified standard spraying rate (e.g., 40 g / m²) and a specific ambient temperature (e.g., -5°C), the maximum theoretical freezing point reduction is measured using a freezing point meter, and this maximum theoretical freezing point reduction is converted into a percentage. This data is pre-entered into the de-icing agent efficacy feature database. When spraying operations are carried out on-site, based on the de-icing agent type reported by the operating vehicle and the actual spraying rate, the corresponding initial set efficacy value is directly retrieved from the de-icing agent efficacy feature database as the calculation starting point. Secondly, the real-time residual efficacy value of the de-icing agent refers to the percentage of relative de-icing capacity retained by the existing de-icing agent on the road surface after deducting the cumulative losses caused by environmental factors from the time of spraying to the current time, starting from the aforementioned initial set efficacy value. The specific implementation steps are as follows: Taking the end of spraying as the zero point, collect the cumulative rainfall from existing roadside weather stations and the cumulative traffic flow data from traffic detectors in real time; call the preset unit rainfall erosion loss coefficient (e.g., every 1mm of rainfall causes a 3% decrease in effectiveness) and unit traffic flow wear loss coefficient (e.g., every 100 standard vehicles passing through causes a 0.5% decrease in effectiveness) in the feature library, calculate the product of the real-time collected cumulative rainfall and the unit rainfall erosion loss coefficient to obtain the first loss ratio, calculate the product of the traffic flow and the unit traffic flow wear loss coefficient to obtain the second loss ratio, add the first loss ratio and the second loss ratio and normalize them to calculate the total loss ratio; finally, based on the difference between the initial set effectiveness value and the total loss ratio, obtain the real-time residual effectiveness value of the de-icing agent: real-time residual effectiveness value of de-icing agent = initial set effectiveness value - total loss ratio (if the calculation result is less than 0, it is forced to zero).

[0022] Step 3: Construct a historical data tensor field and a real-time query tensor based on the real-time residual effectiveness value of the snow melting agent; Historical meteorological feature vectors, historical residual effectiveness values ​​of de-icing agents, historical measured freezing points, and geometric feature labels are obtained and collectively constitute a multi-dimensional historical data tensor field. Through data cleaning and time alignment algorithms, historical meteorological feature vectors, historical residual effectiveness values ​​of de-icing agents, historical measured freezing points, and geometric feature labels representing road physical properties (including roadbed and pavement type, bridge type, and tunnel entrance / exit location) for the target road segment are simultaneously acquired over a preset period (e.g., the past three winters). These four types of data are then fused on a unified time axis and spatial grid to form a multi-dimensional historical data tensor field. This historical data tensor field includes sub-tensors corresponding to each historical spatiotemporal node, and each sub-tensor is also composed of the corresponding historical meteorological feature vector, historical residual effectiveness value of de-icing agents, historical measured freezing point, and geometric feature label. The roadbed and pavement types are used to distinguish the differences in heat capacity and thermal conductivity of road sections made of different materials such as asphalt concrete and cement concrete; the bridge types are used to identify components with bidirectional heat dissipation characteristics, such as viaducts and suspension bridges, to reflect their tendency to freeze before the ground; and the tunnel entrance and exit locations are used to mark special areas near the tunnel entrance where there are alternating light and dark zones and sudden temperature changes.

[0023] Obtain real-time meteorological feature vectors, and generate real-time query tensors based on the real-time meteorological feature vectors and the real-time residual effectiveness value of snow melting agents.

[0024] The residual effectiveness value of the real-time snow melting agent is normalized to obtain a scalar value, and the scalar value is concatenated with the real-time meteorological feature vector to generate a real-time query tensor.

[0025] Based on the real-time query tensor, a similar subset of historical data is retrieved from the historical data tensor field. The subset of historical data contains k sets of historical data pairs (Mi, Ti), where Mi is the historical snow melting agent efficacy value, Ti is the historical measured freezing point, and k is greater than or equal to 2.

[0026] In the process of retrieving similar historical data subsets in the historical data tensor field based on the real-time query tensor, this application adopts a screening strategy of first geometric matching and then meteorological fitting: First, the geometric feature labels in the real-time query tensor (e.g., the current target is a cement road bridge) are compared one by one with the geometric feature labels of all samples in the historical data tensor field. Only historical data with completely identical geometric feature labels are retained, thus initially filtering out the historical subspace with isomorphic physical environments. Because roads with different geometric properties (e.g., bridges and ground, asphalt and cement) have drastically different thermodynamic response mechanisms, mixing and comparing historical data of different materials would lead to distorted meteorological data matching due to differences in basic physical properties, failing to accurately reflect the current icing trend of the road section.

[0027] Secondly, within the historical subspace, refined similarity calculations are performed using historical meteorological feature vectors, historical residual effectiveness values ​​of de-icing agents, and historical measured freezing points to ultimately determine the most similar subset of historical data. This hierarchical selection method ensures that subsequent evolutionary state identification and freezing point prediction are based on the same roadbed and structural form, effectively eliminating prediction errors caused by differences in road geometry and significantly improving the accuracy of icing prediction.

[0028] Within the historical subspace, a refined similarity calculation is performed using historical meteorological feature vectors, historical residual effectiveness values ​​of snow melting agents, and historical measured freezing points to ultimately determine the most similar subset of historical data. The specific implementation process is as follows: This embodiment is based on the cosine similarity algorithm, calculating the similarity score between each sub-tensor of the real-time query tensor and the historical data tensor field. Based on the calculated scores, all similarity scores are sorted in descending order. The system extracts the top k groups (k≥2) of historical data tensor fields with the highest similarity scores to form a subset of similar historical data, and extracts the key result pair (Mi,Ti) corresponding to each group of data, where Mi is the historical snow melting agent effectiveness value and Ti is the historical measured freezing point.

[0029] Based on the aforementioned subset of historical data, a discrete point cloud is formed by mapping it to a two-dimensional phase space with historical snow-melting agent efficacy values ​​on the horizontal axis and historical measured freezing points on the vertical axis.

[0030] A standard normalized two-dimensional phase space coordinate system is constructed based on historical data pairs from a subset of historical data. The horizontal axis (X-axis) is defined as the historical residual effectiveness value of de-icing agents, and the vertical axis (Y-axis) is defined as the historical measured freezing point. Each set of historical data pairs in the subset is treated as an independent coordinate point and mapped one by one to this two-dimensional phase space. As k data points are projected sequentially, a cluster of discrete point clouds naturally forms in the phase space. This discrete point cloud is tightly clustered in the neighborhood of the effectiveness value corresponding to the current real-time operating conditions, intuitively revealing the nonlinear statistical law and distribution density between changes in de-icing agent effectiveness and the reduction of road surface freezing point under similar meteorological backgrounds.

[0031] Several geometric distribution indices are generated based on the discrete point cloud; the geometric distribution indices include: linear concentration index, distribution dispersion index, evolution trend direction index, and end bending characteristic index.

[0032] The linear concentration index characterizes the degree to which the distribution of a point cloud approximates a straight line. The covariance matrix of the discrete point cloud is calculated, and the maximum and minimum eigenvalues ​​of the covariance matrix are determined. The maximum eigenvalue represents the maximum variance of the data distribution, and the minimum eigenvalue represents the minimum variance of the data distribution. A feature ratio I2 / I1 is constructed based on the maximum and minimum eigenvalues, where I1 is the maximum eigenvalue and I2 is the minimum eigenvalue. Based on this feature ratio, the linear concentration index L = 1 - I2 / I1 is calculated.

[0033] The dispersion index characterizes the dispersion range of a point cloud in phase space. The centroid coordinates of the discrete point cloud are calculated, the Euclidean distance from each data point in the discrete point cloud to the centroid coordinates is calculated, and the arithmetic mean of all Euclidean distances is obtained. This arithmetic mean is used as the dispersion index.

[0034] The evolution trend direction index represents the slope of the overall extension of the point cloud. The eigenvector corresponding to the largest eigenvalue in the covariance matrix is ​​obtained, consisting of the horizontal and vertical components. The slope of the eigenvector, S = vertical component / horizontal component, is calculated and used as the evolution trend direction index. The eigenvector corresponding to the largest eigenvalue points towards the principal axis direction with the largest variance in the point cloud distribution.

[0035] The end-bending characteristic index characterizes the degree of nonlinear transition of the point cloud at the end of the time series. The historical data subset is sorted by timestamp from smallest to largest, and the last m (m ≥ 2) data points after sorting are selected (the application and failure of snow melting agent is a dynamic time-varying process. The high-efficiency low-freezing-point state at an earlier moment in the historical data may no longer represent the critical characteristics of the current low-efficiency freezing-point recovery. Only the data at the end of the time closest to the current moment can best reflect the true physical law of the tail of efficiency decay). The m data points are fitted into a quadratic curve equation using the least squares method. The absolute value of the curvature of the quadratic curve equation at the real-time residual efficiency value of the snow melting agent is calculated, and the end-bending characteristic index is obtained after normalizing the absolute value of the curvature.

[0036] The evolutionary state is identified based on the geometric distribution index, and the evolutionary state includes: dissolution state, saturated precipitation transition state, failure accumulation state, and mixed uncertain state.

[0037] Based on each of the aforementioned evolutionary states, and according to the relationship between the geometric distribution index and the preset target value, the freezing point is obtained.

[0038] If the linear concentration index is greater than the first preset target value and the evolution trend direction index is negative, then it is in a dissolving state. A similarity calculation is performed on the historical data subset to obtain a similarity value. Based on the similarity value and the historical measured freezing point of each historical data pair in the historical data subset, a weighted average value is obtained and used as the freezing point.

[0039] First, a logical judgment is made on the linear concentration index and the evolution trend direction index: If the linear concentration index exceeds a preset high threshold (e.g., 0.85), it means that historical data points are highly and closely distributed near a straight line in the effectiveness-freezing point coordinate system, indicating that the physical laws are stable and noise interference is minimal under the current working conditions; if the evolution trend direction index is negative, it intuitively represents that as the effectiveness of the de-icing agent is maintained or increased, the measured freezing point of the road surface is continuously decreasing, showing a clear cooling trend. Only when both of the above conditions are met simultaneously can the current road surface be determined to be in a state of efficient melting.

[0040] This application precisely identifies the golden window period for snow-melting agents to take effect—the optimal time when the agent is rapidly dissolving and the freezing point is steadily decreasing—by combining indicators of high linearity and negative trend. Within this specific period, efficacy and freezing point exhibit a strong linear negative correlation. Therefore, a prediction algorithm based on similarity-weighted averaging is initiated: the similarity between the current real-time operating conditions and each data point in the historical subset is calculated, the similarity values ​​are normalized and used as the basic weight, and a weighted average is calculated with the historical measured freezing points of each historical data pair in the historical data subset to obtain the predicted freezing point.

[0041] This application introduces a linear concentration index, starting from the data confidence level. Only when historical similar data exhibits a high degree of linear clustering in phase space can it be proven that the current physical process follows a stable dissolution law, eliminating the interference of random noise and nonlinear disturbances. Simultaneously, it introduces an evolutionary trend direction index, starting from the physical causality level, mandating that the freezing point must strictly decrease with increasing potency (negative evolution), thereby accurately filtering out the non-dissolution stages of reagent saturation precipitation or failure recovery. This dual-index coupling mechanism fills the gap in existing technologies regarding the coordinated perception of data distribution quality and dynamic evolution direction. It ensures that the weighted prediction model is activated only during the truly efficient dissolution period when data patterns are most significant and physical logic is most explicit. This fundamentally solves the model mismatch and prediction distortion problems caused by the coarse granularity of state identification in existing technologies, achieving a technological leap from empirical time-segmentation to precise perception driven by both data mechanisms.

[0042] If the distribution dispersion index is greater than the second preset target value and the end bending characteristic index is greater than the third preset target value, then it is in a saturated precipitation transition state, and the upper quartile of all historical measured freezing points of the historical data pairs of the historical data subset is taken as the freezing point.

[0043] When the dispersion index is greater than the second preset target value (e.g., a threshold of 0.6, meaning that historical data is no longer tightly clustered but shows a significant scattered distribution, suggesting that the physical process is greatly affected by random disturbances or phase transition instability) and the end-bending characteristic index is greater than the third preset target value (e.g., a threshold of 0.15, meaning that at the current effectiveness level, the rate of change of freezing point with effectiveness undergoes a drastic change, the curve shows a sharp turn, consistent with the nonlinear characteristics of crystal precipitation in a saturated solution), the road surface of the current target section is determined to be in a saturated precipitation transition state. The physical essence of this state is that the concentration of de-icing agent has exceeded the solubility limit at the current temperature, and the excess solute begins to precipitate from the liquid phase to form solid crystals, causing the liquid phase concentration to no longer increase linearly with the increase of the application amount. The freezing point reduction effect reaches its limit or even shows a oscillating rebound. At this time, traditional linear regression or weighted average models are very prone to serious underestimation due to data fluctuations (i.e., predicting the freezing point too low, misleading people into thinking that the de-icing effect is still very good). To address this high-risk situation, this embodiment extracts the historical measured freezing points of all data pairs in the historical data subset, calculates the upper quartile (Q3, i.e., the 75th percentile) of its statistical distribution, and directly uses this value as the current predicted freezing point. The significance of choosing the upper quartile lies in the fact that during the saturation precipitation stage, the data often exhibits a right-skewed distribution (with some outliers at higher freezing points, representing local precipitation-induced de-icing failure). Using Q3 instead of the maximum value avoids extreme noise interference. However, compared to the average or median, Q3 more sensitively reflects the possibility of poor operating conditions, providing a warning value biased towards the safe side (higher freezing point). This ensures that at the critical moment when the de-icing agent's effectiveness is about to fail, the dispatch system can receive a more conservative and safer freezing point signal, thereby triggering timely instructions for re-spreading or switching de-icing schemes. This effectively avoids the risk of road icing caused by blindly trusting low freezing point predictions, reflecting the robustness and safety-first design philosophy of this system under complex phase change conditions.

[0044] If the evolution trend direction index is greater than or equal to zero and the linear concentration index is less than the fourth preset target value, then it is in a state of failure accumulation, and the maximum value of all historical measured freezing points of the historical data pairs of the historical data subset is taken as the freezing point.

[0045] When the trend direction indicator is greater than or equal to zero (indicating that the freezing point no longer decreases with the increase of the de-icing agent's effectiveness, but instead shows a flat or even rising trend, physically meaning that the agent can no longer lower the freezing point) and the linear concentration indicator is less than the fourth preset target value (for example, a threshold of 0.4, meaning that historical data points are extremely scattered in phase space, completely losing linearity and showing a disordered clustered distribution, indicating that there is no definite functional relationship between the effectiveness value and the freezing point), the current road surface is determined to be in a failed accumulation state. The physical essence of this state is that the ambient temperature is too low or the amount of de-icing agent applied far exceeds the solubility limit, causing the de-icing agent to be unable to dissolve and absorb heat, instead accumulating on the ice and snow surface in the form of solid salt particles. This not only fails to lower the freezing point, but may also exacerbate the risk of road icing by forming a heat insulation layer or hindering mechanical snow removal. At this time, any prediction model based on regression or averaging will fail and produce dangerously optimistic estimates. For this highest risk state, this embodiment directly extracts the maximum value of the historical measured freezing point of all data pairs in the historical data subset (i.e., the freezing point value with the highest temperature) and forces it as the current predicted freezing point. The strategic significance of choosing the maximum value is that, during the failure accumulation stage, the data is highly discrete, and the average or median may mask the fact that the freezing point is high in some areas of complete failure, thus misleading decision-makers into believing that there is still some de-icing effect on the road surface; while taking the maximum value is to capture the worst working condition and issue the highest level of safety warning to the dispatch system, that is, the freezing point of the current road section may be close to the ambient temperature, and the de-icing operation has actually been completely failed.

[0046] When any of the above determination logics are not met, the road surface of the target road segment is determined to be in a mixed uncertain state, and the median of the historical measured freezing point of the historical data subset is determined as the current road surface freezing point.

[0047] This application constructs a technical approach that uses query tensors to retrieve similar subsets in historical data tensor fields, integrates point cloud geometric analysis to generate multidimensional indicators to identify evolutionary states, and then outputs the freezing point in a targeted manner. This approach breaks through the dependence of traditional time series forecasting on temporal continuity and pioneers a working condition perception based on phase space topological isomorphism. Through high-dimensional tensor retrieval, this application can transcend the time dimension and accurately locate "historical mirror" subsets in the vast ocean of historical data that are highly similar to the current instantaneous topological structure in multidimensional feature spaces (efficiency, temperature, humidity, etc.), essentially finding the most accurate "physical twin" of the current working condition. Based on this, point cloud analysis is introduced to calculate linear concentration and curvature. Geometric distribution indicators such as evolution trends, and mathematical reconstruction of the dissolution kinetic mechanism (e.g., linearity corresponding to the applicable range of Henry's law, curvature corresponding to the critical point of solubility limit), not only solve the overfitting problem under small sample conditions, but also achieve a qualitative leap from blindly fitting data to actively identifying physical mechanisms. It ensures that the corresponding prediction strategy is activated only when the data distribution conforms to specific physical laws (e.g., strong linear negative correlation in the efficient dissolution period), and automatically switches to a conservative extreme value strategy in non-steady states (e.g., precipitation, accumulation). This approach of first identifying the physical state through geometric topology and then matching the prediction logic according to the state fills the gap in the lack of fine-grained state perception capability in complex phase transition processes in existing technologies.

[0048] Step 4: Generate the physical icing probability based on the freezing point and ground temperature; construct a forecast feature vector based on real-time weather forecast data, and input the forecast feature vector into a pre-constructed machine learning model to generate a predicted icing probability.

[0049] In this embodiment, the weighted average or median obtained in step 1 is used as the baseline expected value of the ground temperature at the current prediction time, representing the most likely state of the road surface temperature under similar seasonal and temporal conditions. Next, to quantify the reliability and fluctuation range of this expected value, the system does not process the current data in isolation, but rather traces back the performance records of the prediction logic in all historical samples from the same period in the past five years. Specifically, it compares the baseline predicted ground temperature for each year in the past five years with the actual measured ground temperature for that year, calculating the prediction residual for each instance. This constructs a historical residual sequence of ground temperature that reflects the inherent error characteristics of the model at a specific seasonal time. Subsequently, the system calculates the standard deviation of the historical residual sequence of ground temperature. This indicator objectively characterizes the natural temperature fluctuation amplitude caused by un-modeled factors such as wind speed, cloud cover, and surface heat capacity at that specific spatiotemporal node. Finally, based on the central limit theorem, this embodiment constructs a normal distribution model with the mean of the baseline expected ground temperature and the standard deviation of the historical ground temperature residuals as the standard deviation to describe the evolution characteristics of the road surface temperature. Therefore, the ground temperature follows a normal distribution, and the probability of physical icing can be calculated by combining the ground temperature and the freezing point. :

[0050] , The standard deviation of historical residuals for ground temperature. The cumulative distribution function of the standard normal distribution. Freezing point This refers to the ground temperature.

[0051] This embodiment, based on a normal distribution model, calculates the probability that the actual road surface temperature has dropped below freezing point under current meteorological conditions and historical error. Even if the predicted temperature is slightly above freezing point, if historical data for the same period indicates significant temperature fluctuations (i.e., a large standard deviation), a higher probability of icing is still output, thus prompting early preventative measures. Conversely, if the predicted temperature is close to freezing point but historical fluctuations are minimal, the probability value is low, avoiding unnecessary resource investment. This mechanism effectively compensates for the blind spots of traditional methods regarding uncertainties such as microclimate abrupt changes, sensor delays, or model residuals, enabling the decision-making system to predict risks. Before actual icing is observed, it identifies potential high-risk scenarios based on statistical patterns, thereby achieving earlier response, more accurate intervention, and fewer false alarms, significantly improving the safety, economy, and intelligence of road anti-icing operations.

[0052] In this embodiment, the process of generating a predicted ice-forming probability based on a pre-built machine learning model is divided into two independent stages: offline model training and online prediction. The specific implementation method is as follows: Phase 1: Offline Model Training (Building a Predictive Model) In the initialization phase, the machine learning model is trained first. Historical minute-level meteorological observation records for the same road section and season over the past three years are retrieved from a local database, including historical data on road surface temperature, air temperature, relative humidity, wind speed and direction, precipitation, and radiation. The historical data is cleaned, denoised, and normalized before being concatenated to construct a historical meteorological feature matrix. Simultaneously, historical measured road surface conditions (iced / non-iced) for the corresponding time period are extracted as labels. Using this data, a time-series classification model based on a deep neural network (such as LSTM or Transformer architecture) is trained. During training, the model learns the complex nonlinear coupling relationships between meteorological elements and the patterns of microclimate disturbances (such as icing patterns caused by sudden temperature drops during rainfall or increased radiative cooling due to high humidity and no wind) through backpropagation. Ultimately, it converges to obtain a pre-built machine learning model that can map meteorological characteristics to icing risk. After training, the model parameters are saved for online prediction.

[0053] Phase Two: Online Prediction (Generating Predicted Ice-Freezing Probabilities) During the real-time operation phase, high-resolution weather forecast data for the next 24 hours is acquired in real time via a data interface. Key forecast factors are extracted from the forecast data, including hourly temperature, hourly relative humidity, hourly wind speed, and hourly precipitation. These key forecast factors are converted into numerical arrays according to the same preprocessing rules (such as normalization and sliding window slicing) used during model training. This set of discrete numerical arrays for the next 24 hours is arranged in chronological order to construct a one-dimensional time series vector as the forecast feature vector. The constructed forecast feature vector is input into the pre-constructed machine learning model. Based on learned physical laws, the model infers future meteorological scenarios, and the output layer directly maps to the estimated icing probability at each future time step, i.e., predicts the icing probability.

[0054] Step 5: Generate the final icing probability based on the physical icing probability and the predicted icing probability.

[0055] The final icing probability is obtained by weighted summing of the physical icing probability and the predicted icing probability.

[0056] The physical icing probability, based on thermodynamic principles and historical statistical distribution, provides a benchmark anchor for the current road surface condition, ensuring the physical interpretability and baseline safety of risk assessment in the absence of future forecasts or under extreme and sudden conditions. The predicted icing probability, on the other hand, uses a deep learning model to capture the nonlinear disturbances and future evolution trends of meteorological elements such as precipitation and wind speed, giving the system a forward-looking perception capability for potential risks. Through weighted fusion, the final icing probability can dynamically adjust its weights according to the evolution of the meteorological environment: during periods of stable meteorological conditions and dominated by historical patterns, the physical icing probability is preset as the first weight, and the predicted icing probability as the second weight, with the sum of the first and second weights being 1 and the first weight being greater than the second weight; preferably, the first weight is 0.7 and the second weight is 0.3. During critical periods of drastic weather system changes (such as the passage of a cold front or the transition from rain to snow), the second weight is greater than the first weight; preferably, the second weight is 0.7 and the first weight is 0.3. The final icing probability is obtained by weighted summation of the physical icing probability, the first weight, the predicted icing probability, and the second weight.

[0057] In the above embodiments, the icing prediction process for roads where de-icing agents have been sprayed is described in detail. For roads where no de-icing agents have been sprayed, the prediction process is as follows: Steps 1 to 4 are executed in the same way, except that in step 2, since no de-icing agent has been sprayed on the current road section, the initial set efficacy value is zero, and there is no rain erosion or traffic wear, so the real-time residual efficacy value of the de-icing agent is constant at zero. Subsequently, when constructing the real-time query tensor in step 3, this zero value is used as an efficacy feature in the calculation; when retrieving the historical data subset in step 4, data samples with historical residual efficacy values ​​of zero (i.e., no de-icing agents were sprayed historically) are automatically matched in the historical data tensor field. Based on these historical data subsets under purely natural conditions, the identified evolution state will only reflect the influence of ambient temperature, humidity, and the road surface's own thermal properties on icing, and the final predicted road surface freezing point is the physical freezing point of the road section under natural conditions.

[0058] This embodiment generates the supplementary snow-melting agent spraying amount per unit area based on the determined evolution state and corresponding geometric distribution characteristic indicators, specifically including: If the road surface of the target section is determined to be in a melting state: extract the absolute value of the evolution trend direction index as the consumption rate index, determine the basic spraying threshold and perform a proportional calculation with the consumption rate index to generate a high incremental spraying volume: set the basic spraying threshold as Dbase; read the evolution trend direction index Ktr to characterize the rate at which the effectiveness of the de-icing agent decreases over time under similar historical conditions; calculate the high incremental spraying volume D1=Dbase×(1+a×|Ktr|); a is a preset response gain coefficient (e.g., set to 0.5) used to linearly amplify the response strength to the rapid melting trend.

[0059] In this embodiment, when it is determined that the target road section is in a melting state (i.e., the sprayed de-icing agent is being rapidly consumed and there is a risk of ice reformation), a high-incremental spraying strategy is immediately initiated: First, the absolute value of the evolution trend direction index at the current moment is extracted as the consumption rate index. This index accurately quantifies the rate at which the effectiveness of the de-icing agent decays over time under similar historical conditions. If the absolute value of the evolution trend direction index is large, it indicates that the ambient temperature is rising rapidly or the traffic volume is high, causing the agent to be lost quickly. Subsequently, the preset basic spraying threshold (e.g., 200 grams per square meter under standard conditions) is read and directly coupled with the consumption rate index to obtain the high-incremental spraying amount.

[0060] If the target road section is determined to be in a saturated precipitation transition state: the distribution dispersion degree is extracted as the balance fluctuation coefficient. When the distribution dispersion degree is lower than the uniformity threshold, a zero spraying amount is generated; when the distribution dispersion degree is greater than or equal to the first preset threshold but the linear concentration degree is higher than the second preset threshold, a compensation spraying amount based on the precipitation forecast is generated: the predicted precipitation Pp is obtained through the meteorological port; the distribution dispersion degree s is read, and the compensation spraying amount D2 = b × Pp × s is calculated; b is the dilution compensation coefficient, specifically the number of grams of de-icing agent required to be added per millimeter of precipitation per square meter of road surface, in grams / (mm·m²).

[0061] In this embodiment, when the target road section is determined to be in a saturated precipitation transition state, a refined compensation strategy based on concentration uniformity and precipitation prediction is initiated. The preset uniformity threshold is set to 0.15 (derived from the minimum variance statistics of historical no-icing accident data). If the distribution dispersion is lower than the uniformity threshold, it indicates that the road surface agent distribution is extremely uniform and has reached saturation equilibrium, at which point zero spraying is generated. Conversely, if the distribution dispersion is higher than the first preset threshold of 0.35 (a critical point representing local concentration unevenness) and the linear concentration is higher than the second preset threshold of 0.8 (an empirical value representing the overall maintenance of a high concentration level), a compensation mechanism is triggered to obtain a compensation spraying amount. If 0.15 ≤ distribution dispersion < 0.35, it is considered slightly uneven, and a basic maintenance spraying amount (preferably set to Dbase × 0.2) is generated. If the distribution dispersion ≥ 0.35 but the linear concentration ≤ 0.8, it indicates that the de-icing agent distribution is chaotic and the linear correlation between concentration and freezing point is disrupted, indicating a potential failure risk. At this time, the current compensation calculation is immediately terminated, and the process transitions to the failure state handling procedure. Specifically, instead of relying on the current real-time freezing point prediction, the system directly calls the maximum freezing point value (most unfavorable working condition) from historical data as the control benchmark and generates a full-volume de-icing agent spraying instruction (e.g., setting the spraying amount to 1.5 times the Dbase or the standard amount covering the entire road surface) to eliminate icing hazards to the greatest extent and ensure driving safety.

[0062] If the target road section is determined to be in a state of failure and accumulation: generate zero spray volume directly and output mechanical snow removal or chemical replacement command.

[0063] In this embodiment, when it is determined that the target road section is in a state of ineffective accumulation (i.e., the residual de-icing agent on the road surface has lost its de-icing ability due to excessive moisture absorption and deliquescence, impurity mixing, or low temperature failure, and has formed an ineffective salt mud accumulation), an interruption control strategy is immediately executed: a zero-spraying command is directly generated to completely cut off the new chemical agent injection, so as to prevent the waste of resources and the aggravation of road surface corrosion caused by ineffective accumulation; at the same time, the control unit outputs high-level operation commands, specifically including sending a mechanical snow removal request to the dispatch center to remove the accumulated ineffective salt mud mixture on the road surface, or issuing an agent replacement warning to remind maintenance personnel to clean the ineffective layer first and repave an effective de-icing agent; the vicious cycle of spraying-ineffective-re-spraying is forcibly interrupted, ensuring that no chemical intervention is carried out before the physical condition of the road surface is restored to the effective threshold, thereby switching the treatment mode from simple dosage adjustment to physical removal and media reset, ensuring the basic effectiveness of subsequent de-icing operations.

[0064] If the target road segment is determined to be in a mixed uncertain state: extract the ratio of the degree of dispersion of the distribution to the degree of linear concentration as the uncertainty confidence factor U, multiply the basic spraying threshold by a safety redundancy coefficient greater than 1, and the safety redundancy coefficient dynamically increases with the increase of the uncertainty confidence factor to generate a defensive spraying amount; Set the safety redundancy coefficient to 1 + ln(1 + U); multiply the basic spraying threshold by the safety redundancy coefficient to generate the defensive spraying amount.

[0065] In this embodiment, the evolutionary state serves as a dynamic benchmark, depicting the entire lifecycle characteristics of the road surface agent from effective penetration to saturation precipitation and then to degradation in real time. Based on this constantly changing benchmark, it accurately determines whether the current stage is a dilution phase requiring minor replenishment or an ineffective phase requiring complete blocking, ensuring that each spray strictly matches the actual activity of the agent. Secondly, the evolutionary state acts as a control switch, directly determining the direction of the operation path. When the evolutionary trajectory shows that the agent is still in the high-efficiency range, a compensation mode based on uncertainty factors is automatically activated to maintain the optimal concentration balance. Once the evolutionary state crosses the critical point and enters the ineffective accumulation range, the state signal immediately cuts off the spraying command and forcibly switches to mechanical removal mode, fundamentally preventing the vicious cycle of continuing to apply the agent on ineffective substrates. This technology, based on the evolutionary state, not only eliminates resource waste and environmental pollution caused by ignoring the timeliness of the agent, but also intervenes in advance by predicting the evolutionary trend, elevating de-icing operations from post-remediation to pre-emptive defense, ensuring the continuity and reliability of road anti-icing capabilities under complex weather conditions.

[0066] In summary, this embodiment involves multiple thresholds, each of which is obtained through analysis of historical data, and will not be described in detail here.

[0067] like Figure 2As shown, this embodiment also discloses an icing prediction system based on multi-source data fusion, including the following modules: First data acquisition module: used to acquire ground temperature; The second data acquisition module is used to preset the initial effectiveness value, collect cumulative rainfall and cumulative traffic flow data in real time, calculate the total loss ratio, and obtain the real-time residual effectiveness value of the snow melting agent based on the difference between the initial effectiveness value and the total loss ratio. The data processing module is used to construct a historical data tensor field and a real-time query tensor based on the real-time residual effectiveness value of the snow melting agent. It calculates the similarity score between each real-time query tensor and the historical data tensor field, sorts all similarity scores in descending order, and selects the top k groups of historical data tensor fields with the highest similarity scores to form a subset of similar historical data. Based on this subset, it maps it to a two-dimensional phase space with the historical snow melting agent effectiveness value on the horizontal axis and the historical measured freezing point on the vertical axis, forming a discrete point cloud. Based on the discrete point cloud, it generates several geometric distribution indices. It identifies the evolution state based on the geometric distribution indices, and based on each evolution state and the relationship between the geometric distribution indices and a preset target value, it obtains the freezing point. Prediction module: used to generate physical icing probability based on the freezing point and ground temperature; construct forecast feature vector based on real-time weather forecast data, and input the forecast feature vector into a pre-constructed machine learning model to generate predicted icing probability; Fusion module: used to generate the final icing probability based on the physical icing probability and the predicted icing probability.

[0068] To implement the above embodiments, this application also proposes an electronic device. Please see [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram of the structure of the electronic device provided in an embodiment of this application. For example... Figure 3 As shown, the electronic device 500 includes: a processor 501 and a memory 502 communicatively connected to the processor 501; the memory 502 stores computer-executable instructions; the processor 501 executes the computer-executable instructions stored in the memory to implement the method provided in the foregoing embodiments.

[0069] To implement the above embodiments, this application also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided in the foregoing embodiments.

[0070] To implement the above embodiments, this application also proposes a computer program product, including a computer program that, when executed by a processor, implements the methods provided in the foregoing embodiments.

[0071] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for predicting icing based on multi-source data fusion, characterized in that, Includes the following steps: Step 1, obtain the ground temperature; Step 2: Preset the initial effectiveness value, collect cumulative rainfall and cumulative traffic flow data in real time, and calculate the total loss ratio; The real-time residual effectiveness value of the snow melting agent is obtained based on the difference between the initial set effectiveness value and the total loss ratio. Step 3: Construct a historical data tensor field and a real-time query tensor based on the real-time residual effectiveness value of the snow melting agent. Calculate the similarity score between each real-time query tensor and the historical data tensor field. Sort all similarity scores in descending order and select the top k groups of historical data tensor fields with the highest similarity scores to form a subset of similar historical data. Map this subset of historical data onto a two-dimensional phase space with the historical snow melting agent effectiveness value on the horizontal axis and the historical measured freezing point on the vertical axis to form a discrete point cloud. Generate several geometric distribution indices based on the discrete point cloud. Identify the evolution state based on the geometric distribution indices. Based on each evolution state and the relationship between the geometric distribution indices and a preset target value, obtain the freezing point. Step 4: Generate the physical icing probability based on the freezing point and ground temperature; construct a forecast feature vector based on real-time weather forecast data, and input the forecast feature vector into a pre-constructed machine learning model to generate a predicted icing probability; Step 5: Generate the final icing probability based on the physical icing probability and the predicted icing probability.

2. The icing prediction method based on multi-source data fusion according to claim 1, characterized in that, Historical meteorological feature vectors, historical residual effectiveness values ​​of snow melting agents, and historical measured freezing points are obtained and together they form a multidimensional historical data tensor field.

3. The icing prediction method based on multi-source data fusion according to claim 1, characterized in that, The real-time meteorological feature vector is obtained, and the real-time residual effectiveness value of the snow melting agent is normalized to obtain a scalar value. The scalar value is then concatenated with the real-time meteorological feature vector to generate a real-time query tensor.

4. The icing prediction method based on multi-source data fusion according to claim 1, characterized in that, The geometric distribution indices include: linear concentration index, distribution dispersion index, evolution trend direction index, and end-bending characteristic index; Calculate the covariance matrix of the discrete point cloud, solve for the maximum and minimum eigenvalues ​​of the covariance matrix, construct the feature ratio based on the maximum and minimum eigenvalues, and calculate the linear concentration index based on the feature ratio. Calculate the centroid coordinates of the discrete point cloud, calculate the Euclidean distance from each data point in the discrete point cloud to the centroid coordinates, and obtain the arithmetic mean of all Euclidean distances. Use the arithmetic mean as an index of the degree of dispersion of the distribution. Find the eigenvector corresponding to the largest eigenvalue in the covariance matrix; calculate the slope of the eigenvector and use the slope as an indicator of the direction of evolution trend. The historical data subset is sorted by timestamp from smallest to largest, and the last m data points after sorting are selected; the m data points are fitted into a quadratic curve equation; the absolute value of curvature of the quadratic curve equation at the real-time residual effectiveness value of the de-icing agent is calculated, and the end bending characteristic index is obtained after normalizing the absolute value of curvature.

5. The icing prediction method based on multi-source data fusion according to claim 4, characterized in that, If the linear concentration index is greater than the first preset target value and the evolution trend direction index is negative, then it is in a dissolving state. A similarity calculation is performed on the historical data subset to obtain a similarity value. The similarity value is used as the basic weight, and a weighted average is calculated with the historical measured freezing point of each historical data pair in the historical data subset and used as the freezing point.

6. The icing prediction method based on multi-source data fusion according to claim 4, characterized in that, If the distribution dispersion index is greater than the second preset target value and the end bending characteristic index is greater than the third preset target value, then it is in a saturated precipitation transition state, and the upper quartile of all historical measured freezing points of the historical data pairs of the historical data subset is taken as the freezing point.

7. The icing prediction method based on multi-source data fusion according to claim 4, characterized in that, If the evolution trend direction index is greater than or equal to zero and the linear concentration index is less than the fourth preset target value, then it is in a state of failure accumulation, and the maximum value of all historical measured freezing points of the historical data pairs of the historical data subset is taken as the freezing point.

8. The icing prediction method based on multi-source data fusion according to claim 1, characterized in that, The first and second weights are preset, and the final icing probability is obtained by weighted summation based on the physical icing probability, the first weight, the predicted icing probability, and the second weight.

9. An icing prediction system based on multi-source data fusion, used to execute the icing prediction method based on multi-source data fusion as described in any one of claims 1-8, characterized in that, Includes the following modules: First data acquisition module: used to acquire ground temperature; The second data acquisition module is used to preset the initial effectiveness value, collect cumulative rainfall and cumulative traffic flow data in real time, and calculate the total loss ratio. The real-time residual effectiveness value of the snow melting agent is obtained based on the difference between the initial set effectiveness value and the total loss ratio. The data processing module is used to construct a historical data tensor field and a real-time query tensor based on the real-time residual effectiveness value of the snow melting agent. It calculates the similarity score between each real-time query tensor and the historical data tensor field, sorts all similarity scores in descending order, and selects the top k groups of historical data tensor fields with the highest similarity scores to form a subset of similar historical data. Based on this subset, it maps it to a two-dimensional phase space with the historical snow melting agent effectiveness value on the horizontal axis and the historical measured freezing point on the vertical axis, forming a discrete point cloud. Based on the discrete point cloud, it generates several geometric distribution indices. It identifies the evolution state based on the geometric distribution indices, and based on each evolution state and the relationship between the geometric distribution indices and a preset target value, it obtains the freezing point. Prediction module: used to generate physical icing probability based on the freezing point and ground temperature; construct forecast feature vector based on real-time weather forecast data, and input the forecast feature vector into a pre-constructed machine learning model to generate predicted icing probability; Fusion module: used to generate the final icing probability based on the physical icing probability and the predicted icing probability.

Images