A Multi-Source Heterogeneous Data Fusion Processing Method for Urban Temperature Prediction

Abstract

This invention relates to the field of electronic digital data processing technology, and discloses a multi-source heterogeneous data fusion processing method for urban temperature prediction. The method includes: first, generating a thermodynamic potential calibration map based on physical and geometric information, independent of real-time data, as a prediction benchmark; then, using sparse measured data for correction. The conduction strength of the correction process is constrained by the potential difference of the calibration map, and the system can adaptively adjust the strength of this constraint based on the prediction residual. This invention avoids the prediction system's excessive reliance on unstable real-time data streams, using a stable physical reference field as the system anchor point, so that the data correction process follows physical logic rather than purely mathematical interpolation. Therefore, even with large-scale data loss or quality degradation, it can still provide physically meaningful and reliable prediction outputs, improving the robustness and usability of the prediction system.

Description

Technical Field

[0001] This invention relates to a multi-source heterogeneous data fusion processing method for urban temperature prediction, belonging to the field of electronic digital data processing technology. Background Technology

[0002] The current mainstream technology for urban temperature forecasting is built on a data-driven logic. Its core is to integrate heterogeneous data from as many sources as possible, such as fixed weather stations, satellite remote sensing images, mobile Internet of Things networks, and social behavior data, and then use machine learning or deep learning models for training in order to obtain high-precision predictions of environmental parameters such as urban temperature.

[0003] However, when this processing method, which relies on the integrity of multi-source data, is placed in the unavoidable complex and dynamic context of actual urban operation, a hidden cost that has long been accepted by the industry becomes apparent. Specifically, in a typical summer afternoon severe convective weather event, the rapid intrusion of a cumulonimbus cloud can trigger a series of chain data failures. For example, optical remote sensing satellites may experience data interruption due to cloud cover, some low-cost IoT sensors may have abnormal readings due to rain intrusion, and sudden changes in citizens' travel patterns may cause their associated behavioral data to lose its general regularity. This large-scale nonlinear distortion of the input feature vector directly leads to the unreliable output of prediction models that rely heavily on these data for fitting, and even results that contradict physical reality.

[0004] Faced with this challenge, the most intuitive improvement approaches are, firstly, to deploy higher-specification and more robust sensing hardware, and secondly, to develop more complex algorithms to interpolate and repair missing or abnormal data. However, the former is hindered by its high cost to achieve large-scale application, while the latter, because it is essentially still performing statistical inference at the data level, often lacks real physical basis when faced with global environmental changes dominated by strong physical events, and may even amplify errors. Thus, the fundamental contradiction of existing technologies lies in: 1. The reliability of the system is bound to the data stream it relies on, which is most prone to failure at critical moments; 2. The pursuit of prediction accuracy forces the model to continuously integrate more data sources, but this in turn increases the risk of the system failing globally due to the failure of a single data source. Therefore, the technical problem to be solved by this invention is how to construct a completely new data processing method that no longer regards the diverse and uneven quality of real-time data streams as necessities for system operation, but instead establishes a physically meaningful benchmark and uses any available real-time data as a correction to that benchmark. This enables the entire prediction system to still provide a prediction output with a reliable lower limit even when the data fails or degrades on a large scale. Summary of the Invention

[0005] This invention provides a multi-source heterogeneous data fusion processing method for urban temperature prediction. Its main purpose is to solve the problem that the reliability of the prediction system is overly tied to the unstable real-time data stream under the existing data-driven approach, which leads to unreliable or even completely invalid system output results when large-scale data failure occurs.

[0006] To achieve the above objectives, this invention provides a multi-source heterogeneous data fusion processing method for urban temperature prediction, comprising:

[0007] Step a: Based on the three-dimensional geometric model of the city and the sun's position at the target time, calculate and generate a thermal potential calibration map;

[0008] Step b: Obtain m measured temperature points, and use a weighted parameter that includes the thermodynamic potential decay parameter statistically calibrated from historical data to perform preliminary correction on the thermodynamic potential calibration map through a correction function to obtain a preliminary temperature prediction field. The correction strength of the correction function depends on the difference between the geographical distance and the corresponding value on the thermodynamic potential calibration map.

[0009] Step c: Calculate the residual between the measured value at the measured temperature point and the corresponding predicted value in the preliminary temperature prediction field, and dynamically adjust the thermal potential decay parameter in the weight parameters based on the difference in the average value of the residual between the high thermal potential region and the low thermal potential region to obtain the updated weight parameters.

[0010] Step d: Using the updated weight parameters, re-execute the correction function to generate the final temperature prediction field;

[0011] Step e: When m=0, the thermodynamic potential calibration map is transformed by a linear transformation whose transformation parameters are determined by regression analysis of historical data, and then used as the final temperature prediction field output.

[0012] Preferably, the value of any point on the thermal potential calibration map in step a is obtained by the following calculation: determining the direct sunlight factor based on whether the point receives direct sunlight, and determining the sky diffuse light factor based on the sky openness of the point; weighting the direct sunlight factor and the sky diffuse light factor together; and multiplicatively correcting the weighted combination result according to the absorption coefficient of the ground material corresponding to the point.

[0013] Preferably, the correction strength of the correction function in steps b and d It is defined as the product of a geographical distance influence function and a thermal potential difference influence function, and its calculation rules are as follows: Where i is a measured temperature point and j is a target point. The geographical distance between the measured temperature point and the target point. and These are the values ​​of two points on the thermodynamic potential calibration diagram. It is a distance attenuation parameter. This is the thermodynamic potential decay parameter.

[0014] Preferably, in step c, the thermal potential decay parameter is dynamically adjusted as follows: when the difference in the average value of the residual between the high thermal potential region and the low thermal potential region increases, the value of the thermal potential decay parameter is increased; when the difference in the average value of the residual between the high thermal potential region and the low thermal potential region decreases, the value of the thermal potential decay parameter is decreased.

[0015] Preferably, the method further includes: before performing step b, acquiring real-time rainfall intensity data of the covered city, and generating a negative thermodynamic potential map characterizing the evaporative cooling effect of rainfall based on the real-time rainfall intensity data; superimposing the negative thermodynamic potential map and the thermodynamic potential calibration map at the pixel level to generate a comprehensive thermodynamic potential calibration map; and using the comprehensive thermodynamic potential calibration map to replace the original thermodynamic potential calibration map to perform subsequent steps.

[0016] Preferably, the method further includes: before performing step a, calculating the spatial variance of the temperature values ​​contained in the m measured temperature points; when the spatial variance is greater than a first threshold, performing the calculation of the original step a to generate a thermal potential calibration map; when the spatial variance is less than a second threshold, adjusting the calculation rules in step a, using the sky openness of each geographical location as the dominant factor to generate a thermal potential calibration map, wherein both the first threshold and the second threshold are determined based on historical data statistics.

[0017] Preferably, it also includes acquiring real-time wind speed and direction data, and adjusting the geographical distance influence function according to the wind direction and wind speed, so that the influence range of the correction intensity extends along the downwind direction and shortens in the upwind direction, forming an anisotropic correction range.

[0018] Preferably, in step e, the transformation parameters are determined by linear transformation based on historical data regression analysis. Specifically, this involves performing linear regression analysis on a set of historical measured temperature values ​​and their corresponding historical thermodynamic potential calibration map values ​​that have similar historical meteorological conditions to the target time, in order to determine the scale parameters and offset parameters.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. First, a thermodynamic potential calibration map representing the potential for receiving solar radiation at various geographical locations in the city is generated through light projection calculations, and at least one measured temperature data is obtained. Subsequently, the calibration map is not used as an isolated prediction result, but is transformed into defining the inherent physical rules for the subsequent correction process. That is, the difficulty of the transmission of the correction effect of any measured temperature depends not only on the geographical distance, but also directly on the difference in thermodynamic potential values ​​between the two points. This method makes the data correction process no longer a mathematical interpolation in a homogeneous space, but follows a non-uniform conduction path preset by stable physical laws, so that the final generated temperature prediction field maintains an inherent consistency with the physical logic of the actual thermal distribution of the city.

[0021] 2. After generating the preliminary temperature prediction field, this method calculates the residuals between the measured and predicted values ​​at each measured temperature point and analyzes the statistical characteristics of the distribution of these residuals across different thermal potential regions. These statistical characteristics directly reflect the strength of the actual thermal isolation effect formed by sunlight and shading. Based on this characteristic, the system can dynamically adjust the weight parameters related to thermal potential differences in the correction intensity and re-execute the correction steps using the adjusted parameters, so that the core parameters of the correction model can be adaptively matched according to the meteorological conditions implied in the measured data itself.

[0022] 3. When acquiring real-time rainfall intensity data, this method converts it into a negative thermodynamic potential map representing the evaporative cooling effect. This map is then overlaid with a thermodynamic potential calibration map based on light and shadow calculations to form a comprehensive thermodynamic potential calibration map that integrates both radiative heating and evaporative cooling. This comprehensive map then replaces the original calibration map to guide subsequent corrections for non-uniform heat conduction. This expands the physical basis of the entire prediction system from a single clear-sky radiation model to a mixed thermodynamic scenario that can respond to rainfall events, avoiding systematic prediction biases caused by physical model mismatches under specific weather conditions. Furthermore, by reusing the calculation process for generating the target time thermodynamic potential calibration map, a future time thermodynamic potential calibration map is generated, and the two maps are differentially calculated. The result is a map that directly represents the short-term temperature change trend of various geographical locations in the city caused by changes in the sun's position. This trend map, as an independent output alongside the final temperature prediction field, provides the system with a description of the dynamic evolution of the future thermal environment without introducing a new prediction model. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the adaptive prediction process of the technical solution of this invention.

[0024] Figure 2 This is a comparison chart of the dynamic response and prediction error of the core parameters of this invention;

[0025] Figure 3This is the system architecture and closed-loop control flow diagram for multi-source data fusion of the present invention;

[0026] Figure 4 This is a schematic diagram of the non-uniform correction mechanism based on physical benchmarks of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0028] A multi-source heterogeneous data fusion processing method for urban temperature prediction is proposed. The data processing flow begins with the construction of a thermodynamic potential calibration map generated based on physical and geometric information, which serves as the prediction benchmark. Then, using any available sparse measured temperature data, a non-uniform thermal conduction correction is applied to this benchmark. The core parameters of this correction process are further adjusted through a feedback loop based on the prediction residual, ultimately generating the urban temperature prediction field. To address the issue of reduced reliability of the prediction model output due to real-time data stream interruptions or quality degradation in existing data-driven methods, this method first performs step a, which calculates and generates a thermodynamic potential calibration map based on the urban 3D geometric model and the solar position at the target time. The input data for this step includes a one-time offline urban 3D geometric model data, which can be generated from open map data or publicly available government geographic information system data, and a temperature... The system predicts the target time and geographical location information. The process first calculates the solar altitude and azimuth angles at the target time using standard astronomical algorithms. Then, it employs a ray projection algorithm from computer graphics to project the sun as a parallel light source onto a 3D city model. The output of this process is a quantified thermodynamic potential calibration map. The value at any point on the map is obtained according to a deterministic physical rule. Specifically, the system determines the direct sunlight factor based on whether the location receives direct sunlight and the sky diffuse light factor based on the sky's openness. The direct sunlight factor and the sky diffuse light factor are weighted and combined, and the weighted combination result is multiplicatively corrected based on the absorptivity coefficient of the corresponding land cover material. The resulting thermodynamic potential calibration map, with its pixel values ​​representing the potential for solar radiation received at that geographical location at that time, provides a prediction benchmark independent of real-time data quality for subsequent processing. (Direct sunlight factor) With sky-scattered light factor The specific calculation can be achieved as follows: direct light factor The calculation, taking into account shading, solar geometry, and atmospheric conditions, can be expressed by the following formula: Here, S is the shading factor (the result of light projection calculation, 1 for sunlight and 0 for shadow). The solar constant, The zenith angle of the sun. Atmospheric transmittance, Cloud attenuation coefficient; Sky scattering factor The calculation is mainly based on the degree of sky openness, and its formula can be expressed as: Here, SVF represents the sky's openness at that point. The total sky diffuse radiation intensity is related to atmospheric conditions and cloud cover. By introducing the specific calculation formulas that include atmospheric physical parameters, the accuracy and repeatability of the thermal potential calibration map as a physical benchmark are ensured. After obtaining the thermal potential calibration map, in order to correlate this physical potential field with the actual temperature value, the system then executes step b, obtains m measured temperature points, and uses a weight parameter that includes a thermal potential attenuation parameter statistically calibrated from historical data to perform preliminary correction on the thermal potential calibration map through a correction function to obtain a preliminary temperature prediction field. The core of the correction process is that the correction intensity of each measured temperature point i relative to any target point j... It is defined as the product of a geographical distance influence function and a thermal potential difference influence function, and its calculation rules are as follows: ,in, This represents the geographical distance between the measured temperature point and the target point. and These are the values ​​of the two points on the thermodynamic potential calibration diagram, respectively. For distance attenuation parameters, For thermodynamic potential decay parameters; correction intensity It is a dimensionless weight value, geographical distance. and its corresponding distance attenuation parameters The unit is usually meters (m), and the value of the thermodynamic potential calibration diagram is... , and its corresponding thermodynamic potential decay parameters These are dimensionless relative values ​​that have been normalized; this correction method ensures that the data correction process follows physical boundaries determined by illumination and shadow, rather than mathematical interpolation within a homogeneous space; it takes into account a fixed thermodynamic potential decay parameter. Unable to adapt to the dynamic changes in the actual thermal isolation effect caused by factors such as wind speed that are not directly measured, the system further executes step c, calculating the residual between the measured value at the measured temperature point and the corresponding predicted value in the preliminary temperature prediction field. Based on the difference in the average value of the residual between high and low thermal potential regions, the system dynamically adjusts the thermal potential attenuation parameter in the weighting parameters. The adjustment procedure is as follows: when the difference in the average value of the residual between high and low thermal potential regions increases, it indicates that the thermal isolation effect caused by light and shadow is more significant under the current weather conditions, and the system accordingly increases the thermal potential attenuation parameter. The value of the average value; conversely, when the difference in the average value decreases, it indicates that the air mixing effect is stronger, the thermal isolation effect is weakened, and the system decreases. The value; this adjustment process constitutes a feedback adjustment loop that does not rely on external meteorological sensors, enabling the model parameters to self-calibrate based on the conditions inherent in the measured data itself; after obtaining the updated weight parameters, the system executes step d, using the updated weight parameters to re-execute the correction function to generate the final temperature prediction field.

[0029] To ensure system availability in data-deficient scenarios, this method is configured to execute step e when no measured temperature data points are available (m=0). In this case, the system outputs the final temperature prediction field after performing a linear transformation on the thermodynamic potential calibration map, the transformation parameters of which are determined by regression analysis of historical data. The scale and offset parameters of this linear transformation are determined by performing linear regression analysis on a set of historical measured temperature values ​​under similar historical meteorological conditions to the target time and their corresponding historical thermodynamic potential calibration map values. This mechanism allows the system to switch to outputting a prediction benchmark based on physical meaning and historical statistics when all real-time data is unavailable, thus eliminating dependence on easily lost data streams. To improve the physical completeness of the model for specific weather events, without conflict, this method can also acquire real-time rainfall intensity data covering the city before executing step b. The system then uses this rainfall intensity data, through a non-linear transformation rule, to generate a representation of... The negative thermodynamic potential map of the evaporative cooling effect of rainfall is generated and superimposed pixel-wise with the original thermodynamic potential calibration map to produce a comprehensive thermodynamic potential calibration map. Subsequent correction steps will be based on this comprehensive map that integrates both radiative heating and evaporative cooling to simulate the cooling effect of rainfall. Similarly, to enable the system's physical model to adapt to both sunny and cloudy radiation patterns, this method can first calculate the spatial variance of the temperature values ​​contained in m measured temperature points before executing step a. When the spatial variance is greater than a first threshold determined based on historical data statistics, it indicates that the urban temperature difference is significant and the weather pattern is sunny, and the system executes the original calculation in step a. When the spatial variance is less than a second threshold, it indicates that the urban temperature is uniform and the weather pattern is cloudy, and the system adjusts the calculation rules in step a, using the sky openness of each geographical location as the dominant factor to generate the thermodynamic potential calibration map. This mechanism allows the system to dynamically switch its core modeling logic according to data characteristics.

[0030] Furthermore, to address the impact of wind fields on heat transport, this method can also acquire real-time wind speed and direction data, and adjust the correction intensity based on wind direction and speed. The geographical distance influence function is adjusted so that the influence range of the correction intensity extends along the downwind direction and shortens in the upwind direction, forming an anisotropic correction range to simulate the heat distribution caused by advection transport. This method can also reuse the calculation process of step a, based on a future time relative to the target time. The system uses solar position information to generate a future thermal potential calibration map. Then, it performs a difference calculation between the future and target thermal potential calibration maps, resulting in a thermal change trend map representing the increase or decrease in heat at various geographical locations in the city due to changes in solar position. This trend map can be used as an independent output alongside the final temperature prediction field, providing city managers with a description of the dynamic evolution of the future thermal environment. Finally, to characterize the differences in thermodynamic properties across different underlying surface areas, this method may also include a parameter inversion and regional adaptive adjustment process. During a period selected based on stable thermal input in historical meteorological data, the system applies a standardized virtual thermal pulse to the physical parameters of a region within the calculation model, and records and analyzes the predicted temperature response curves for multiple subsequent time steps triggered by this pulse. Based on the characteristics of the response curves, the system can invert and calculate an index that quantifies the equivalent heat capacity of the region, and use this index to perform regional adaptive adjustments to the parameters in the correction function, enabling the model to reflect the thermal response characteristics of different regions.

[0031] In practical implementation, the negative thermal potential map generated based on real-time rainfall intensity data has the following pixel values: Through conversion function The calculation results in which The rainfall intensity at that point, and the heat conversion coefficient C and the nonlinear response index. The two internal parameters are determined by optimizing the sum of squared simulation errors from historical rainfall samples according to localized calibration procedures. Correspondingly, based on adjustments to the influence function of wind direction and speed on geographical distance, the specific procedure involves using an effective wind influence distance. Euclidean distance in the alternative correction function This effective distance is calculated. Obtain, among which The vector pointing from the measured point i to the target point j. With real-time wind field vector The included angle, For wind speed, and the wind force influence coefficient. According to the pre-deployment calibration procedure, calibration is performed through interpolation of pre-calculated wind impact maps or statistical analysis of historical data. To achieve a quantitative characterization of the thermodynamic properties of different underlying surface areas, the process of inverting and calculating the equivalent heat capacity index of the region involves a standardized virtual heat pulse index applying a fixed increment to the absorptivity coefficient of the target area's land cover material within a single time step within the model, and recording the predicted temperature response curves for multiple subsequent time steps. Subsequently, the peak increment was extracted from the curve. Ascent time With decay time Based on these characteristics, the equivalent heat capacity index is calculated. Its operation rules can be as follows: Where k is a normalization constant, which is subsequently used to adjust the distance decay parameter in the correction function. Regional adaptation adjustments are performed; furthermore, to address potential systematic biases in the measured data, a data consistency verification step is included before correction, which calculates the standardized thermodynamic potential deviation for each measured point i. Deviation from standardized measured temperature When the signs of the two data points are opposite and the absolute value of their difference exceeds the threshold determined based on historical data statistics, the weight of that measurement point data in subsequent correction calculations will be reduced or temporarily removed.

[0032] Meanwhile, in this invention, the thermodynamic potential calibration map generated in step a is essentially a quantitative two-dimensional data field representing the potential of each location point in the urban geographic space to receive total solar radiation at a specific target time. The generation of this map does not depend on any real-time meteorological measurement data, but is based on a deterministic physical model. The core input of this model is the static three-dimensional geometric structure information of the city and the geometric position information of the sun at the target time. The value of any location point on the map (i.e., thermodynamic potential) is calculated after comprehensively considering physical factors such as whether the point receives direct sunlight, the ability to receive scattered light determined by the openness of the sky, and the absorption rate of radiation by the corresponding ground material. Therefore, this calibration map provides a stable prediction benchmark with clear physical meaning and independent of dynamic data flow for all subsequent data correction and prediction processes.

[0033] Example 1: On a summer afternoon, a temperature prediction system for city A is in operation. At this time, a cumulonimbus cloud rapidly moves into the city and causes data failure. Specifically, optical remote sensing satellites experience data interruption due to cloud cover, and some IoT sensors show abnormal readings due to rain intrusion. A prediction model that relies on this data for fitting suffers large-scale distortion of its input feature vector, resulting in an output temperature field that does not match physical reality. When faced with the same data failure situation, the technical solution of this invention does not interrupt its processing flow. The system first calls a thermodynamic potential calibration map generated according to step a of the aforementioned specific implementation method, which does not rely on any real-time data. This map is calculated based on the solar geometric position at the target time and the static three-dimensional structural information of the city, and it itself represents the thermal distribution pattern of the city surface caused by sunlight and shading. At this time, although a large number of IoT sensors have failed, a few high-level meteorological stations located in critical infrastructure are still working normally. The system then executes step b, using these sparse measured temperature data points as correction sources to correct the thermodynamic potential calibration map.

[0034] This correction process is achieved through a non-uniform correction function constrained by a thermodynamic potential calibration map. The correction effect of a measured high-temperature data point located in an open square will be transmitted through the correction function to other streets also directly exposed to sunlight. However, its influence is suppressed when attempting to cross into adjacent areas in the shadow of buildings due to the difference in thermodynamic potential (P) values ​​between the two points. This approach transforms the traditional challenge of statistical interpolation when data is extensively missing into a problem of logically correcting a stable reference field using sparse anchor points within a physical framework, thus maintaining prediction accuracy even with incomplete data. Furthermore, during the feedback adjustment in step c, the system calculates the residuals between the measured values ​​and the preliminary predicted values ​​at each effective temperature measurement point. It finds that the mean residual value in high thermodynamic potential areas is significantly higher than that in low thermodynamic potential areas. Based on this, the system determines that the thermal isolation effect caused by light and shadow under current weather conditions remains strong and accordingly increases the thermodynamic potential attenuation parameter. The value of makes the subsequent secondary correction more closely fit the physical boundary formed by the shadows of buildings. At the same time, the system acquires real-time rainfall intensity data covering the city and generates a negative thermodynamic potential map characterizing the evaporative cooling effect of rainfall. This negative thermodynamic potential map is superimposed on the original thermodynamic potential calibration map to generate a comprehensive thermodynamic potential calibration map. Subsequently, all correction calculations are based on this comprehensive reference field that integrates the three physical effects of sunlight shading and rainfall evaporation. Finally, when other models that rely on multi-source real-time data output temperature fields that contradict physical reality due to input feature distortion, the final temperature prediction field generated by this method not only reflects the temperature values ​​of a few measured points, but also presents a temperature distribution that is consistent with the spatial distribution of the city's geometric structure, solar position, and rainfall range in areas without direct data coverage. Even when the data quality is degraded on a large scale, it still provides a physically meaningful output.

[0035] Example 2: To verify the predictive robustness of the technical solution of the present invention under different real-time data integrity conditions, this example sets up a set of comparative experiments. The experimental environment is a 2km x 2km urban center area with high-density buildings and complex underlying surface features. 100 calibrated temperature sensors are deployed in this area as ground truth references. The software environment of the experimental platform is a server that deploys the technical solution of the present invention and a data-driven model based on gradient boosting decision trees. The core independent variable of the experiment is the integrity of real-time data. The design aims to simulate the progressive failure of sensor networks in a real urban environment. The procedure is set to start from 100% data integrity and gradually reduce the data integrity in discrete steps. In each step, a portion of sensors are randomly selected so that their data is invisible to the prediction model until the data integrity drops to 0%. The mean absolute error of the temperature prediction field output by the experimental group and the control group relative to the ground truth values ​​of all 100 sensors is recorded at each integrity level.

[0036] Experimental results show that as the real-time data completeness decreases from 100%, the mean absolute error (MAE) of the experimental group increases gradually, while the error of the control group increases sharply, indicating a difference in prediction robustness between the two. At 100% data completeness, the MAE of the control group is 0.52°C, and that of the experimental group is 0.65°C. When the data completeness decreases to 50%, the MAE of the control group rises to 1.35°C, while that of the experimental group is only 0.82°C. When the data completeness further decreases to 10%, the error of the control group expands to 3.58°C, while the error of the experimental group remains at 1.15°C. Finally, when the number of available sensors (m) is 0, the control group model fails due to lack of effective input, while the experimental group predicts based on the output of step e, with an MAE of 2.53°C. Analysis of the above experimental phenomena indicates that the performance of the control group is significantly different from the performance of the input it relies on. The integrity of the feature vectors is strongly coupled. When a large number of features are missing, the learned statistical laws become invalid, leading to increased errors. In contrast, the performance of the experimental group is anchored to a physical benchmark—the thermodynamic potential calibration map generated in step a, which is independent of real-time data. Any available real-time data is used as a correction to this stable benchmark, rather than being essential for system operation. Even when the number of available sensors m is extremely low, the correction process still takes place in a physically meaningful non-uniform field, thus suppressing the rapid divergence of errors. The results of this experiment confirm that the technical solution of this invention has higher stability in prediction accuracy when faced with large-scale missing or degraded real-time input data. This characteristic stems from its data processing architecture, which uses a constant physical benchmark field as the system anchor and utilizes real-time data as a correction, thus possessing higher availability in real-world application environments with unstable data flows.

[0037] Example 3: This example combines Figures 1 to 4 This paper describes a multi-source heterogeneous data fusion processing method for urban temperature prediction, as follows: Figure 1 As shown, the process begins with two core inputs: a 3D geometric model of the city representing its static geometry and a target time and location specifying the prediction requirements. Based on these two inputs, the system first generates a thermal potential calibration map to construct a physical prediction benchmark independent of real-time data. After acquiring this benchmark, the system enters a data correction phase, acquiring m sparse measured data points and performing preliminary thermal conduction corrections on the aforementioned calibration map based on these data points, generating a preliminary temperature prediction field. To improve the adaptability of the prediction, the system then performs residual analysis and parameter adaptive adjustment steps, that is, based on the differences in the residuals of the preliminary prediction in different thermal potential regions, adjusting the parameters accordingly. The thermal potential decay parameters are dynamically adjusted during the process, and the final thermal conduction correction is re-executed using updated weight parameters to generate the final temperature prediction field and output the urban temperature distribution map. This process also includes processing logic for two special cases. First, when real-time rainfall intensity data is obtained, the negative thermal potential map representing rainfall cooling can be superimposed on the original calibration map by fusing the rainfall evaporation cooling effect step to form a comprehensive thermal potential map that is more responsive to weather changes. Second, when no measured data can be obtained, i.e. when m=0, the system can directly output the thermal potential calibration map as the prediction benchmark through a transformation determined by historical regression linear transformation to maintain the availability of the system.

[0038] like Figure 2 As shown in the figure, the left vertical axis represents the thermodynamic potential decay parameter. The values ​​of are shown, with the right-hand vertical axis representing the prediction error in °C. The graph also shows the parameters for sunny weather conditions. The curve shows that under sunny conditions, The value increases significantly during the day, peaking in the afternoon, which is consistent with the physical reality of enhanced thermal isolation effect caused by changes in solar radiation, while the cloud condition parameter... The curve remained at a relatively low level throughout the day. Correspondingly, the two curves for sunny weather prediction error and cloudy weather prediction error showed that, despite the dynamic changes in external meteorological conditions and internal core parameters, this method could stably control the prediction error within a low range in both modes, thus verifying the adaptability and robustness of the technical solution.

[0039] like Figure 3As shown, the data sources for the entire system include urban geographic information systems, forecast task designations, weather radar systems, and real-time sensor networks. These data collectively support the operation of three core processing modules. Module 1.0 generates a physical reference field, receiving initial information and generating the D1 thermodynamic potential calibration map as the physical reference. This map can also be fused with rainfall intensity data from the weather radar system to form a comprehensive thermodynamic potential map. Module 2.0, which integrates real-time data correction, is the core of the processing flow. It receives measured temperature data m from the real-time sensor network and combines it with the initial calibration parameters from the D2 historical meteorological database or the transformation parameters under the m=0 scenario to correct the physical reference field and output the final temperature forecast field to city managers / users. Module 3.0 performs feedback adjustment, updating the core parameters in the correction process based on the preliminary forecast field and residuals generated by Module 2.0 and feeding them back to Module 2.0, forming a closed-loop adjustment loop.

[0040] like Figure 4 As shown, the physical reference field on the left illustrates how solar radiation forms shadows and eclipses based on urban building geometry such as B1, B2, B3, and B4, thereby constructing a thermal potential calibration map with high and low potential regions on the ground. The graph, as the core input, is fed into the data fusion processing flow on the right, which receives measured temperature data as input. ), and through a core correction function When performing data fusion, the strength of this function is determined by both geographical distance attenuation and thermal potential difference attenuation, and its mathematical expression is: Subsequently, the system uses a feedback mechanism based on the residual difference between high and low potential regions to adjust the parameters in the thermodynamic potential difference decay term. Through dynamic adjustments, a final temperature prediction field is generated that integrates physical logic and measured data. The prediction field can be generally described as a functional relationship. Where P represents the complete thermodynamic potential calibration diagram. Let represent the data set of m measured temperature points, and W represent the result of the correction function. A series of weight values ​​were calculated.

[0041] Example 4: Before deploying the technical solution of the present invention, its correction function is calibrated. The core parameter in the process is the distance attenuation parameter. With thermodynamic potential decay parameter The system is configured to execute a standardized offline parameter determination procedure. The input to this procedure is a dataset of hourly meteorological observations for the target city over at least one full year, encompassing measured temperature, wind speed, and total radiation intensity information from multiple ground sensors. The procedure first filters out all daytime periods in the dataset that meet clear and calm conditions to isolate the underlying physical scenario dominated by solar radiation. Subsequently, the system performs a grid search within a predefined two-dimensional parameter space. The search range was set to 50 meters to 500 meters. The search range was set to 0.05 to 0.5; for each set in the parameter space To determine the value, the system performs a cross-validation calculation based on the filtered dataset. At each time step, the system uses a leave-one-out method, which involves taking each sensor as the target point to be predicted in turn, and using the measured temperatures of all other sensors at that moment, based on the current... Parameter combination through correction function Perform temperature prediction and calculate the error between the predicted value and the actual measured value of the sensor.

[0042] After traversing all time steps and all sensors, the system calculates the current... The global mean absolute error corresponding to the parameter combination; after completing the grid search of the entire parameter space, which set minimizes the global mean absolute error? The combination of these parameters is then determined as the initial weight parameters for the urban environment and configured into the system as the default value for subsequent real-time predictions. Based on this, to further determine the dynamic adjustment logic in step c, the system utilizes the residual data generated during the optimization process to analyze the difference in the average residual values ​​between high and low thermal potential regions at different wind speed levels. The relationship between the values ​​is used to establish a mapping of the difference in the residual mean to... The functional relationship of the adjustment quantities is determined. Simultaneously, to determine the linear transformation parameters under the m=0 condition in step e, the system uses the same historical dataset to extract the thermodynamic potential calibration map values ​​P and corresponding measured temperature values ​​T for all selected time periods. A linear regression analysis is then performed on the (P,T) data set, and the slope and intercept obtained from the regression are determined as the scale and offset parameters of the linear transformation. By executing the above offline parameter determination procedure, the core algorithm parameters of the scheme are determined before deployment. and The linear transformation parameters in step e are all determined, so that the implementation of this technical solution does not depend on empirical settings.

[0043] Example 5: When the optional function of correcting based on real-time rainfall intensity data in the technical solution of this invention is enabled, in order to calibrate the internal parameters of the negative thermodynamic potential conversion rule in this function, the system is configured to execute a localized calibration procedure. This procedure retrieves multiple complete rainfall process samples from the local historical dataset. Each sample contains minute-level meteorological radar rainfall intensity maps from one hour before the rainfall to one hour after the rainfall stabilizes, as well as measured temperature data from high-density ground sensors. For each rainfall sample, the system uses the measured temperature field before the rainfall as the initial condition, and simulates the temperature field change after the rainfall by systematically adjusting the parameter combination in the negative thermodynamic potential conversion rule. The simulation results are compared with the measured ground temperature during that period. The parameter combination that minimizes the sum of squared simulation errors of all rainfall samples is determined as the configuration parameters of this functional module in the urban environment.

[0044] Accordingly, when enabling the function of dynamically switching weather modes based on the spatial variance of measured temperature, the system is configured to execute a classification statistical procedure to determine the first and second thresholds. This procedure is based on the local historical dataset. First, it automatically filters out two types of sample sets according to historical meteorological records: one is clear sky samples with more than 6 consecutive hours of sunshine, and the other is cloudy samples with a total cloud cover that is consistently higher than 80%. Then, for each time point in these two sample sets, the system randomly selects virtual measured temperature points of different numbers and spatial distributions multiple times and calculates their spatial variance, thereby obtaining the spatial variance distribution intervals belonging to the clear sky and cloudy sky modes respectively. The system takes the 10% quantile of the variance distribution interval of the clear sky mode as the first threshold and the 90% quantile of the variance distribution interval of the cloudy sky mode as the second threshold. This setting introduces hysteresis near the critical switching point of the mode to suppress high-frequency switching of the physical model.

[0045] Example 6: Before deploying the technical solution of this invention in a new urban area and enabling its advanced correction function, to enable the system to cope with the unique micro-meteorological environment and data quality challenges of the local area, the system is configured to execute a standardized pre-deployment calibration procedure. The first part of this procedure aims to pre-calculate the impact of wind field disturbance on heat conduction. Based on the three-dimensional geometric model of the target city, it uses a simplified computational fluid dynamics model to simulate the detailed airflow distribution inside the city when the external wind field blows through the city with a set of offline, fixed directions and standardized wind speeds, thereby generating a set of wind force impact maps corresponding to each standard wind direction. During the real-time operation phase of the system, after acquiring real-time wind speed and direction data, the system first determines one or two standard wind directions that are closest to the current real-time wind direction and retrieves the corresponding pre-calculated wind force impact map. Through interpolation, a dynamic impact map adapted to the current real-time wind conditions is generated. The values ​​on this map are used to adjust the correction function. The calculation of geographical distance effectively shortens it along the downwind direction and effectively lengthens it against the wind direction, thus coupling the corrected transmission intensity with the wind's transport effect. The second part of this pre-deployment calibration procedure aims to establish a logic for online data quality verification to address potential systematic biases at a few available measured temperature points. Before executing the correction process in step b, this verification procedure performs a data consistency check on m measured temperature points. It first calculates the average and standard deviation of the thermodynamic potential calibration map values ​​P and the average and standard deviation of the measured temperature values ​​T corresponding to the m measured temperature points. Subsequently, for each measured point i, the system calculates its standardized thermodynamic potential deviation. Deviation from standardized measured temperature When the system detects that the standardized temperature deviation of a certain measuring point has the opposite sign to its standardized thermodynamic potential deviation or the difference exceeds a threshold determined based on historical data statistical analysis, the measuring point is judged as an abnormal data point. Its weight in subsequent correction calculations will be reduced or its data will be temporarily removed. This online verification mechanism enables the system to identify and suppress erroneous data inputs that contradict the thermodynamic distribution laws revealed by the physical benchmark model and are caused by sensor malfunctions or abnormal microenvironment before performing thermodynamic conduction correction.

[0046] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for fusing and processing multi-source heterogeneous data for urban temperature prediction, characterized in that, include: Step a: Based on the three-dimensional geometric model of the city and the sun's position at the target time, calculate and generate a thermal potential calibration map; Step b: Obtain m measured temperature points, and use a weighted parameter that includes the thermodynamic potential decay parameter statistically calibrated from historical data to perform preliminary correction on the thermodynamic potential calibration map through a correction function to obtain a preliminary temperature prediction field. The correction strength of the correction function depends on the difference between the geographical distance and the corresponding value on the thermodynamic potential calibration map. Step c: Calculate the residual between the measured value at the measured temperature point and the corresponding predicted value in the preliminary temperature prediction field, and dynamically adjust the thermal potential decay parameter in the weight parameters based on the difference in the average value of the residual between the high thermal potential region and the low thermal potential region to obtain the updated weight parameters. Step d: Using the updated weight parameters, re-execute the correction function to generate the final temperature prediction field; Step e: When m=0, the thermodynamic potential calibration map is transformed by a linear transformation whose transformation parameters are determined by regression analysis of historical data, and then used as the final temperature prediction field output.

2. The method for multi-source heterogeneous data fusion processing for urban temperature prediction according to claim 1, characterized in that, The value of any point on the thermal potential calibration map in step a is obtained by the following calculation: the direct sunlight factor is determined based on whether the point receives direct sunlight, and the sky diffuse light factor is determined based on the sky openness of the point. The direct sunlight factor and the sky diffuse light factor are weighted and combined, and the weighted combination result is multiplicatively corrected according to the absorption coefficient of the ground material corresponding to the point.

3. The method for multi-source heterogeneous data fusion processing for urban temperature prediction according to claim 1, characterized in that, The correction strength of the correction function in steps b and d It is defined as the product of a geographical distance influence function and a thermal potential difference influence function, and its calculation rules are as follows: Where i is a measured temperature point and j is a target point. The geographical distance between the measured temperature point and the target point. and These are the values ​​of two points on the thermodynamic potential calibration diagram. It is a distance attenuation parameter. This is the thermodynamic potential decay parameter.

4. The method for multi-source heterogeneous data fusion processing for urban temperature prediction according to claim 1, characterized in that, In step c, the thermal potential decay parameter is dynamically adjusted as follows: when the difference in the average value of the residual between the high thermal potential region and the low thermal potential region increases, the value of the thermal potential decay parameter is increased; when the difference in the average value of the residual between the high thermal potential region and the low thermal potential region decreases, the value of the thermal potential decay parameter is decreased.

5. The method for multi-source heterogeneous data fusion processing for urban temperature prediction according to claim 1, characterized in that, Also includes: Before performing step b, real-time rainfall intensity data of the covered city is obtained, and a negative thermodynamic potential map characterizing the evaporative cooling effect of rainfall is generated based on the real-time rainfall intensity data; the negative thermodynamic potential map and the thermodynamic potential calibration map are superimposed at the pixel level to generate a comprehensive thermodynamic potential calibration map; and the comprehensive thermodynamic potential calibration map is used to replace the original thermodynamic potential calibration map to perform subsequent steps.

6. The method for multi-source heterogeneous data fusion processing for urban temperature prediction according to claim 1, characterized in that, Also includes: Before performing step a, calculate the spatial variance of the temperature values ​​contained in the m measured temperature points; When the spatial variance is greater than the first threshold, the calculation in the original step a is performed to generate a thermal potential calibration map; when the spatial variance is less than the second threshold, the calculation rules in step a are adjusted to use the sky openness of each geographical location as the dominant factor to generate a thermal potential calibration map, wherein the first threshold and the second threshold are both determined based on historical data statistics.

7. The method for multi-source heterogeneous data fusion processing for urban temperature prediction according to claim 3, characterized in that, It also includes acquiring real-time wind speed and direction data, and adjusting the geographical distance influence function based on wind direction and speed, so that the influence range of the correction intensity extends along the downwind direction and shortens in the upwind direction, forming an anisotropic correction range.

8. The method for multi-source heterogeneous data fusion processing for urban temperature prediction according to claim 1, characterized in that, In step e, the transformation parameters are determined by linear transformation based on historical data regression analysis. Specifically, the scale parameters and offset parameters are determined by performing linear regression analysis on a set of historical measured temperature values ​​and their corresponding historical thermodynamic potential calibration map values ​​that have similar historical meteorological conditions to the target time.

Images