Deep learning-based air temperature prediction method and system

Abstract

The application relates to the technical field of air temperature prediction, in particular to an air temperature prediction method and system based on deep learning, which identifies a city heat island key region and an adjacent mountain slope cold lake key region in a target region; meteorological observation data, geographic information data and surface coverage feature data are standardized and fused into a multi-dimensional feature tensor, and then a preliminary gridded air temperature prediction field is generated through an air temperature prediction model; a cold air flow intensity contribution item and a city heat island intensity contribution item are coupled to obtain a heat flux conduction intensity, and a final temperature difference compensation amount is generated in combination with a distance attenuation effect between the city heat island key region and the mountain slope cold lake key region; the preliminary gridded air temperature prediction field is corrected according to the final temperature difference compensation amount to obtain a final gridded air temperature prediction field; the problem that a slight temperature difference caused by the close proximity of a city heat island and a mountain slope cold lake cannot be accurately captured is effectively solved, and the accuracy of air temperature prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of temperature forecasting technology, specifically to a temperature forecasting method and system based on deep learning. Background Technology

[0002] In real-world meteorological environments, some cities are closely connected to complex mountainous terrain, resulting in urban heat islands adjacent to hillside cold lakes. The urban heat island effect, caused by dense buildings, concentrated populations, and frequent industrial activities, leads to significantly higher temperatures in urban areas compared to surrounding suburbs; while hillside cold lakes create relatively low-temperature zones near hillsides. When urban heat islands and hillside cold lakes are adjacent, a subtle temperature difference exists between them, influenced by the interaction of various factors.

[0003] However, existing temperature forecasting methods often struggle to accurately capture the subtle temperature variations caused by the proximity of urban heat islands and cold lakes on hillsides when dealing with such unique terrain and urban layouts. This inaccurate temperature forecasting can negatively impact various aspects of daily life planning and agricultural production.

[0004] Therefore, there is an urgent need for a deep learning-based temperature forecasting method to effectively address the problem of failing to accurately capture subtle temperature differences caused by the proximity of urban heat islands and cold lakes on hillsides, thereby significantly improving the accuracy of temperature forecasts. Summary of the Invention

[0005] (1) Technical problems to be solved

[0006] The purpose of this invention is to provide a deep learning-based temperature forecasting method and system to solve the problem that the subtle temperature differences caused by the proximity of urban heat islands and cold lakes on hillsides cannot be accurately captured, resulting in inaccurate temperature forecasts.

[0007] (2) Technical solution

[0008] To achieve the above objectives, on the one hand, the present invention provides a temperature forecasting method based on deep learning, the method comprising:

[0009] S1. Acquire meteorological observation data, geographic information data, and land cover characteristic data of the target area; identify and mark key urban heat island areas and adjacent cold lake areas on hillsides within the target area based on the geographic information data and land cover characteristic data; the geographic information data includes slope, altitude, and spatial information; the land cover characteristic data includes building density and vegetation coverage.

[0010] S2. The meteorological observation data, geographic information data and land cover feature data are standardized and fused into a multidimensional feature tensor. The multidimensional feature tensor is then used to generate a preliminary gridded temperature forecast field through a temperature forecast model.

[0011] S3. Obtain the relative altitude difference between the key urban heat island area and the key hillside cold lake area, and calculate the cold air flow intensity contribution term based on the relative altitude difference; determine the urban heat island intensity contribution term based on the building density of the key urban heat island area; obtain the heat flux conduction intensity by coupling the cold air flow intensity contribution term and the urban heat island intensity contribution term, and generate the final temperature difference compensation amount by combining the distance attenuation effect between the key urban heat island area and the key hillside cold lake area.

[0012] S4. Based on the final temperature difference compensation, the temperature forecast values ​​corresponding to the key urban heat island area and the key hillside cold lake area in the preliminary gridded temperature forecast field are corrected to obtain the final gridded temperature forecast field, and temperature forecasts are made based on the final gridded temperature forecast field.

[0013] Furthermore, the method for generating a preliminary gridded temperature forecast field from the multidimensional feature tensor using a temperature forecast model includes:

[0014] The multi-dimensional feature tensor is extracted by multi-layer convolution through the encoder network of the temperature forecast model to obtain a multi-scale meteorological feature map; the multi-scale meteorological feature map is then upsampled through the decoder network of the temperature forecast model to restore the spatial resolution and fused with the meteorological feature map of the corresponding scale from the encoder network to generate a grid resolution fused feature map.

[0015] The grid resolution fusion feature map is processed through a regression output layer to obtain a preliminary gridded temperature forecast field covering the target area; the temperature forecast model is a convolutional neural network with an encoder-decoder architecture.

[0016] Furthermore, the method for calculating the cold air flow intensity contribution based on the relative altitude difference and determining the urban heat island intensity contribution based on the building density of the key urban heat island area includes:

[0017] The relative altitude difference is used to calculate the cold air flow intensity contribution term using a cold air flow intensity function, which characterizes the positive correlation between the relative altitude difference and the cold air sinking velocity.

[0018] Obtain building density data for key areas of the urban heat island, and use the building density data to determine the contribution of urban heat island intensity through an urban heat island intensity mapping function.

[0019] Furthermore, the method of obtaining the heat flux conduction intensity by coupling the cold air flow intensity contribution term and the urban heat island intensity contribution term, and generating the final temperature difference compensation amount by combining the distance attenuation effect between the key area of ​​the urban heat island and the key area of ​​the hillside cold lake, includes:

[0020] The heat flux conduction intensity of the cold air flow intensity contribution and the urban heat island intensity contribution is quantified in microscale space through nonlinear product operation.

[0021] The Euclidean distance is calculated based on the geometric center of the key area of ​​the urban heat island and the geometric center of the key area of ​​the hillside cold lake. The distance attenuation function is established based on the heat diffusion attenuation law, and the real-time attenuation coefficient is calculated.

[0022] The heat flux conduction intensity and the real-time attenuation coefficient are weighted and fused to generate the basic quantity of temperature difference compensation; the basic quantity of temperature difference compensation is corrected for terrain shading effect based on the relative positional topological relationship between the key urban heat island area and the key hillside cold lake area, and then the final temperature difference compensation quantity is output.

[0023] Furthermore, the method for quantifying the heat flux conduction intensity of the cold air flow intensity contribution and the urban heat island intensity contribution in microscale space through nonlinear product operations includes:

[0024] The cold air flow intensity contribution term is defined as the cold source equivalent thermal conductivity, and the urban heat island intensity contribution term is defined as the heat source intensity value; a nonlinear product relationship function between the cold source equivalent thermal conductivity and the heat source intensity value is established according to Fourier's law of heat conduction.

[0025] The equivalent thermal conductivity of the cold source and the intensity value of the heat source are used to calculate the microscale heat exchange flux value through the nonlinear product relationship function; the heat flux conduction intensity is generated by integrating the microscale heat exchange flux value over a unit area, and the heat flux conduction intensity characterizes the amount of heat energy transferred through a unit area per unit time.

[0026] Furthermore, the method for establishing a distance attenuation function based on the Euclidean distance according to the thermal diffusion attenuation law and calculating the real-time attenuation coefficient includes:

[0027] Based on the diffusion equation of Fourier's law of heat conduction, a distance decay function is constructed with the Euclidean distance as the independent variable, and the initial thermal diffusion decay factor is calculated through the distance decay function.

[0028] The elevation gradient data of the key area of ​​the cold lake on the hillside is obtained to generate the terrain dynamic enhancement coefficient; the initial thermal diffusion attenuation factor and the terrain dynamic enhancement coefficient are dynamically weighted and fused to obtain the terrain correction attenuation factor.

[0029] The real-time attenuation coefficient is obtained by correcting the heat exchange path of the terrain correction attenuation factor based on the relative azimuth angle data between the key urban heat island area and the key hillside cold lake area.

[0030] Furthermore, the method for correcting the temperature difference compensation baseline based on the topological relationship between the key urban heat island area and the key hillside cold lake area, and then outputting the final temperature difference compensation amount, includes:

[0031] Based on the azimuth and elevation profile data of the geometric center of the key urban heat island area and the key hillside cold lake area, the three-dimensional shading angle is calculated through three-dimensional terrain analysis, and the shading coefficient is determined based on the three-dimensional shading angle; the three-dimensional shading angle and the shading coefficient are negatively correlated.

[0032] The minimum elevation difference data between key areas of urban heat island and key areas of cold lake on hillside is obtained. Based on the minimum elevation difference, the kinetic energy loss of cold air flow is analyzed to obtain an elevation correction factor. Based on the shading coefficient and the elevation correction factor, a dynamic terrain shading factor is calculated. Based on the dynamic terrain shading factor and the basic amount of temperature difference compensation, the final corrected temperature difference compensation amount is obtained by multiplying the basic amount of temperature difference compensation.

[0033] Furthermore, the method for calculating the three-dimensional shading angle from the azimuth data and elevation profile data through three-dimensional terrain analysis includes:

[0034] A spatial straight line vector is constructed based on the three-dimensional coordinates of the geometric center of the key area of ​​the urban heat island and the three-dimensional coordinates of the geometric center of the key area of ​​the hillside cold lake; a continuous topographic elevation profile sequence along the spatial straight line vector is generated based on the altitude and spatial location information in the geographic information data.

[0035] Calculate the elevation offset of each sampling point in the continuous terrain elevation profile sequence relative to the spatial straight line vector, and form a terrain shading profile curve based on the elevation offset. Identify the elevation maxima points in the terrain shading profile curve as key shading points.

[0036] The three-dimensional occlusion angle is obtained by calculating the three-dimensional spatial angle between the line connecting the start and end points of the spatial straight line vector and the key occlusion point.

[0037] Furthermore, the method for obtaining the elevation correction factor based on the analysis of cold air flow kinetic energy loss according to the minimum elevation difference includes:

[0038] The theoretical kinetic energy loss value of cold air ascent is calculated based on the minimum elevation difference data; the theoretical kinetic energy loss value is dimensionless to obtain the initial correction coefficient; the initial correction coefficient is corrected according to the slope of the key area of ​​the cold lake on the hillside to obtain the elevation correction factor.

[0039] On the other hand, based on the same inventive concept, this invention also provides a temperature forecasting system based on deep learning. The system includes: a data acquisition and key area marking module, a preliminary gridded temperature forecast field generation module, a final temperature difference compensation generation module, and a temperature forecast field correction and temperature forecasting module, with each module connected in a sequential communication manner.

[0040] The data acquisition and key area marking module is used to acquire meteorological observation data, geographic information data, and land cover feature data of the target area; and to identify and mark the key urban heat island area and the adjacent cold lake area on the hillside within the target area based on the geographic information data and land cover feature data; the geographic information data includes slope, altitude, and spatial information; and the land cover feature data includes building density and vegetation coverage.

[0041] The preliminary gridded temperature forecast field generation module is used to standardize and merge the meteorological observation data, geographic information data and land cover feature data into a multidimensional feature tensor, and then use the multidimensional feature tensor to generate a preliminary gridded temperature forecast field through a temperature forecast model.

[0042] The final temperature difference compensation generation module is used to obtain the relative altitude difference between the key urban heat island area and the key hillside cold lake area, calculate the cold air flow intensity contribution term based on the relative altitude difference, determine the urban heat island intensity contribution term based on the building density of the key urban heat island area, obtain the heat flux conduction intensity by coupling the cold air flow intensity contribution term and the urban heat island intensity contribution term, and generate the final temperature difference compensation term by combining the distance attenuation effect between the key urban heat island area and the key hillside cold lake area.

[0043] The temperature forecast field correction and temperature forecast module is used to correct the temperature forecast values ​​corresponding to the key urban heat island area and the key hillside cold lake area in the preliminary gridded temperature forecast field according to the final temperature difference compensation amount to obtain the final gridded temperature forecast field, and to make temperature forecasts based on the final gridded temperature forecast field.

[0044] (3) Beneficial effects

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] 1. By acquiring meteorological observation data, geographic information data, and land cover characteristic data of the target area, and after standardizing these data, they are fused into a multidimensional feature tensor. Then, a preliminary gridded temperature forecast field is generated using a convolutional neural network temperature forecast model. This fully explores the meteorological characteristics and patterns in the data to more comprehensively and accurately reflect the meteorological conditions of the target area.

[0047] 2. By coupling the contribution of cold air flow intensity and urban heat island intensity, the heat flux conduction intensity is obtained. At the same time, the final temperature difference compensation is generated by combining the distance attenuation effect. This more accurately captures the complex thermal changes between these two key areas, and then makes targeted corrections to the preliminary gridded temperature forecast field, further optimizing the temperature forecast results.

[0048] 3. By calculating the three-dimensional shading angle through three-dimensional terrain analysis, calculating the elevation correction factor based on the minimum elevation difference, and correcting the temperature difference compensation base quantity for terrain shading effect, the system can accurately reflect the influence of terrain on air flow and heat exchange. The resulting gridded temperature forecast field can more accurately reflect the actual temperature distribution of the target area, achieving high-precision temperature forecasting. Attached Figure Description

[0049] Figure 1 This is a flowchart of the deep learning-based temperature forecasting method according to Embodiment 1 of the present invention.

[0050] Figure 2 This is a schematic diagram of the module composition of the deep learning-based temperature forecasting system according to Embodiment 2 of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Before providing examples, it is necessary to describe the application scenarios of this invention. This invention is a temperature forecasting method and system based on deep learning, applied to situations where cities are closely connected with complex mountainous terrain, resulting in urban heat islands adjacent to cold lakes on hillsides. When urban heat islands and cold lakes on hillsides are adjacent, subtle temperature differences occur between them. Existing temperature forecasting methods often struggle to accurately capture these subtle temperature differences, leading to inaccurate temperature forecasts.

[0053] Example 1: As Figure 1 As shown in the figure, this embodiment provides a temperature forecasting method based on deep learning, the method including:

[0054] S1. Acquire meteorological observation data, geographic information data, and land cover characteristic data of the target area; identify and mark key urban heat island areas and adjacent cold lake areas on hillsides within the target area based on the geographic information data and land cover characteristic data; the geographic information data includes slope, altitude, and spatial information; the land cover characteristic data includes building density and vegetation cover; the meteorological observation data includes real-time temperature, humidity, air pressure, wind speed, wind direction, and other conventional meteorological elements from various observation stations, which are derived from the network of automatic weather stations deployed within the target area. The geographic information data, covering slope, altitude, and spatial information, is obtained through a digital elevation model (DEM), providing three-dimensional topographic features. The land cover characteristic data includes building density and vegetation cover. Building density is obtained by interpreting remote sensing images to determine the proportion of building area per square kilometer, and vegetation cover is calculated using the normalized difference in vegetation index (NDVI). For example, based on building density thresholds and temperature observation data, key urban heat island areas with building density greater than 60% and temperatures significantly higher than surrounding areas are identified. For the identification of key areas of cold lakes on hillsides, we analyzed hillside areas with slopes greater than 15 degrees and relatively high altitudes, while combining water distribution information and temperature observation data to determine cold lake areas with relatively low temperatures. The identification process employed a spatial clustering algorithm to cluster and label grid cells that met the criteria, forming complete boundaries for the key areas.

[0055] S2. The meteorological observation data, geographic information data, and land cover characteristic data are standardized and fused into a multidimensional feature tensor. This multidimensional feature tensor is then used to generate a preliminary gridded temperature forecast field through a temperature forecast model. Data standardization eliminates dimensional differences between different data sources. For example, temperature data is normalized to the 0-1 range in degrees Celsius, altitude is normalized according to the maximum elevation difference in the region, and building density and vegetation cover are already proportional data. The standardized data are aligned according to a spatial grid to form a multidimensional feature tensor, with each spatial grid unit containing multiple feature dimensions such as meteorology, geography, and land cover.

[0056] S3. Obtain the relative altitude difference between the key urban heat island region and the key hillside cold lake region; calculate the cold air flow intensity contribution term based on the relative altitude difference; determine the urban heat island intensity contribution term based on the building density of the key urban heat island region; obtain the heat flux conduction intensity by coupling the cold air flow intensity contribution term and the urban heat island intensity contribution term, and generate the final temperature difference compensation amount by combining the distance attenuation effect between the key urban heat island region and the key hillside cold lake region; calculate the relative altitude difference between the key urban heat island region and the key hillside cold lake region, i.e., the altitude difference between the geometric center points of the two regions. The final temperature difference compensation amount reflects the degree of influence of the interaction between the urban heat island and the hillside cold lake on the local temperature distribution.

[0057] S4. Based on the final temperature difference compensation, the temperature forecast values ​​corresponding to the key urban heat island area and the key hillside cold lake area in the preliminary gridded temperature forecast field are corrected to obtain the final gridded temperature forecast field. Temperature forecasts are then made based on this final gridded temperature forecast field. Appropriate cooling corrections are made to the forecast temperature of the urban heat island area, and corresponding heating or cooling adjustments are made to the hillside cold lake area. The correction range is determined based on the final temperature difference compensation. After correction, the final gridded temperature forecast field accurately reflects the complex temperature distribution characteristics of the adjacent areas of the urban heat island and hillside cold lake, achieving high-precision temperature forecast services.

[0058] The method for generating a preliminary gridded temperature forecast field from the multidimensional feature tensor using a temperature forecast model includes:

[0059] The multi-dimensional feature tensor is extracted into multi-scale meteorological feature maps through multi-layer convolution in the encoder network of the temperature forecast model. These multi-scale meteorological feature maps are then upsampled through the decoder network of the temperature forecast model to restore spatial resolution and fused with meteorological feature maps of corresponding scales from the encoder network to generate a grid-resolution fused feature map. The specific implementation of the temperature forecast model employs a convolutional neural network with an encoder-decoder architecture, which effectively extracts spatial features from meteorological data while maintaining the spatial continuity of forecast results. The encoder network contains multiple convolutional and pooling layers, extracting meteorological features at different scales from the input multi-dimensional feature tensor through layer-by-layer convolution operations. For example, the first convolutional kernel size is 7×7, used to capture large-scale meteorological patterns, while subsequent layers gradually reduce the kernel size to 3×3, focusing on extracting local detail features. Each convolutional layer is followed by batch normalization and a ReLU activation function to ensure the stability of the training process. During the generation of the multi-scale meteorological feature maps, the encoder generates feature maps of different resolutions at different depths. Shallow feature maps maintain high spatial resolution and mainly contain spatial structural information such as terrain contours and building distribution. While deep feature maps have lower spatial resolution, they contain more abstract meteorological model information, such as temperature gradient variation patterns and thermal circulation characteristics. This multi-scale feature extraction method ensures that the model can simultaneously perceive local details and global meteorological patterns. The decoder network gradually restores spatial resolution through upsampling operations, while fusing meteorological feature information from the encoder network at the corresponding scale. Upsampling is achieved using transposed convolution, avoiding information loss that may occur with simple interpolation. The feature fusion process employs a skip connection mechanism, directly connecting the feature map of a layer in the encoder network to a corresponding layer in the decoder network. This allows the decoder network to utilize more refined and localized feature information extracted from the encoder network while restoring spatial resolution, thus preserving high-resolution spatial details and fusing deep semantic information. The generation of grid-resolution fused feature maps ensures the accurate spatial correspondence of the final forecast results.

[0060] The fused feature map at the grid resolution is processed through a regression output layer to obtain a preliminary gridded temperature forecast field covering the target area. The temperature forecast model is a convolutional neural network with an encoder-decoder architecture. The regression output layer uses a 1×1 convolutional kernel to convert the fused feature map at the grid resolution into single-channel temperature forecast values, directly outputting continuous temperature values ​​without using an activation function. Each grid cell in the preliminary gridded temperature forecast field corresponds to a fixed geographical location in the target area. The grid resolution is typically set to 1 km × 1 km, which meets the accuracy requirements for urban-scale temperature forecasts.

[0061] The method for calculating the cold air flow intensity contribution based on the relative altitude difference and determining the urban heat island intensity contribution based on the building density of the key urban heat island area includes:

[0062] The relative altitude difference is used to calculate the cold air flow intensity contribution term using a cold air flow intensity function. This function characterizes the positive correlation between the relative altitude difference and the cold air descending velocity. The cold air flow intensity contribution term quantifies the direct impact of cold air descending on the temperature of the target area. The cold air flow intensity function describes the quantitative relationship between the relative altitude difference and the cold air descending velocity. According to physical laws, the two are usually positively correlated; that is, the greater the relative altitude difference, the greater the cold air descending velocity.

[0063] Building density data for key areas of the urban heat island is obtained, and the contribution factors to urban heat island intensity are determined using the urban heat island intensity mapping function. The urban heat island intensity mapping function accurately reflects the relationship between building density and urban heat island intensity. It can take the form of a linear function, a polynomial function, or a piecewise function. The function parameters are determined through field observation data or simulation experiments, enabling the urban heat island intensity mapping function to reasonably quantify the impact of building density on urban heat island intensity.

[0064] The method for obtaining heat flux conduction intensity by coupling analysis of the cold air flow intensity contribution term and the urban heat island intensity contribution term, and generating the final temperature difference compensation amount by combining the distance attenuation effect between the key area of ​​the urban heat island and the key area of ​​the hillside cold lake, includes:

[0065] The heat flux conduction intensity of the cold air flow intensity contribution and the urban heat island intensity contribution is quantified in microscale space through nonlinear product operation.

[0066] The Euclidean distance is calculated based on the geometric center of the key area of ​​the urban heat island and the geometric center of the key area of ​​the hillside cold lake. The distance attenuation function is established based on the heat diffusion attenuation law, and the real-time attenuation coefficient is calculated.

[0067] The heat flux conduction intensity and the real-time attenuation coefficient are weighted and fused to generate the basic quantity of temperature difference compensation; the basic quantity of temperature difference compensation is corrected for terrain shading effect based on the relative positional topological relationship between the key urban heat island area and the key hillside cold lake area, and then the final temperature difference compensation quantity is output.

[0068] The method for quantifying the heat flux conduction intensity of the cold air flow intensity contribution and the urban heat island intensity contribution in microscale space through nonlinear multiplication includes:

[0069] The cold air flow intensity contribution term is defined as the equivalent thermal conductivity of the cold source, and the urban heat island intensity contribution term is defined as the heat source intensity value. A nonlinear product function is established between the equivalent thermal conductivity of the cold source and the heat source intensity value based on Fourier's law of heat conduction. The cold air flow intensity contribution term is defined as the equivalent thermal conductivity of the cold source, reflecting the ability of cold air to carry and conduct heat. The urban heat island intensity contribution term is defined as the heat source intensity value, characterizing the intensity of heat radiation from the urban area. The mathematical expression of the nonlinear product function is: Heat exchange flux value = m × × Where m is the coupling coefficient. and It is a non-linear exponent, and its value is usually between 0.6 and 0.8.

[0070] The microscale heat exchange flux is calculated by applying the nonlinear product function to the equivalent thermal conductivity of the cold source and the intensity of the heat source. The heat flux conduction intensity is then generated by integrating this microscale heat exchange flux over a unit area. This heat flux conduction intensity characterizes the amount of heat energy transferred per unit time through a unit area. The calculation of the microscale heat exchange flux considers the heat conduction mechanism at the air molecular level. When cold and hot air masses meet at the boundary layer, collisions and diffusion between molecules lead to heat transfer from the high-temperature region to the low-temperature region. The higher the equivalent thermal conductivity of the cold source, the stronger the ability of the cold air to carry away heat; the higher the intensity of the heat source, the more heat can be conducted. The nonlinear product relationship between the two reflects the synergistic effect and saturation characteristics in the heat conduction process. The heat flux conduction intensity is generated by integrating over a unit area, and the integration process considers the geometry of the contact interface and the temporal evolution of heat exchange. In practical calculations, the contact boundary between the urban heat island and the hillside cold lake is divided into several micro-area elements. The micro-scale heat exchange flux value on each element is multiplied by the area of ​​the element, and then summed to obtain the total heat flux conduction intensity. The physical meaning of heat flux conduction intensity is the amount of heat energy transferred through a unit area per unit time.

[0071] The method for establishing a distance attenuation function based on the Euclidean distance according to the thermal diffusion attenuation law and calculating the real-time attenuation coefficient includes:

[0072] Based on the diffusion equation according to Fourier's law of heat conduction, a distance attenuation function is constructed with the Euclidean distance as the independent variable. The initial thermal diffusion attenuation factor is then calculated using this distance attenuation function. The establishment of the distance attenuation function is based on the diffusion solution of the Fourier heat conduction equation in three-dimensional space. The constructed distance attenuation function is as follows: ,in, The normalization constant is For Euclidean distance, The distance decay exponent, The characteristic diffusion length is denoted as . The distance decay function considers both power-law and exponential decay mechanisms, providing a more accurate description of the attenuation law of heat diffusion under complex terrain conditions. The initial heat diffusion decay factor is obtained by substituting the actual Euclidean distance into the distance decay function.

[0073] The elevation gradient data of the key area of ​​the hillside cold lake is obtained to generate a topographic dynamic enhancement coefficient. The initial heat diffusion attenuation factor and the topographic dynamic enhancement coefficient are dynamically weighted and fused to obtain a topographic correction attenuation factor. The generation of the topographic dynamic enhancement coefficient considers the promoting effect of the elevation gradient of the key area of ​​the hillside cold lake on cold air flow. The elevation gradient data is obtained by gradient calculation of the digital elevation model, reflecting the steepness of the terrain. When the elevation gradient is large, the sinking speed of cold air under gravity increases, enhancing the heat exchange intensity with the urban heat island. The dynamic weighted fusion process combines the initial heat diffusion attenuation factor with the topographic dynamic enhancement coefficient, and the weight coefficient is dynamically adjusted according to the complexity of the terrain; the more complex the terrain, the greater the weight of the topographic dynamic effect.

[0074] The real-time attenuation coefficient is obtained by correcting the heat exchange path of the topographic correction attenuation factor based on the relative azimuth data between the key areas of the urban heat island and the key areas of the hillside cold lake. The heat exchange path correction takes into account the relative azimuth between the urban heat island and the hillside cold lake. When the two are located in the same valley direction, the heat exchange efficiency is high; when there is lateral obstruction, the exchange intensity needs to be reduced accordingly.

[0075] The method for correcting the temperature difference compensation baseline based on the topological relationship between the key urban heat island area and the key hillside cold lake area, and then outputting the final temperature difference compensation amount, includes:

[0076] Based on the azimuth and elevation profile data of the geometric centers of the key urban heat island area and the key hillside cold lake area, a three-dimensional shading angle is calculated using three-dimensional terrain analysis. The shading coefficient is then determined based on this angle. The three-dimensional shading angle and the shading coefficient are negatively correlated. The calculation of the three-dimensional shading angle employs a three-dimensional terrain analysis method, comprehensively considering the spatial relationship between the azimuth and elevation profile data. The azimuth data determines the direction from the geometric center of the key urban heat island area to the geometric center of the key hillside cold lake area, while the elevation profile data provides information on the terrain undulations along this direction. During the three-dimensional terrain analysis, sampling points are set along the straight path connecting the two geometric centers, and the elevation of each sampling point is obtained, forming a continuous terrain profile. The shading coefficient is determined based on the negative correlation between the three-dimensional shading angle and the actual degree of heat exchange obstruction. When the three-dimensional shading angle is large, it indicates a significant terrain barrier between the two key areas, resulting in a greater obstruction to the exchange of hot and cold air currents, and a correspondingly smaller shading coefficient.

[0077] The minimum elevation difference data between key urban heat island areas and key hillside cold lake areas is obtained. An elevation correction factor is obtained by analyzing the kinetic energy loss of cold air flow based on this minimum elevation difference. A dynamic terrain shading factor is calculated based on the shading coefficient and the elevation correction factor. The final corrected temperature difference compensation is obtained by multiplying the dynamic terrain shading factor with the basic temperature difference compensation amount. The minimum elevation difference refers to the minimum elevation obstacle that cold air needs to overcome as it flows from the hillside cold lake area to the urban heat island area. When there is a significant terrain obstacle, the cold air needs to consume additional kinetic energy to overcome the gravitational potential energy difference, resulting in a weakened flow intensity when reaching the urban heat island area. The elevation correction factor is calculated using the law of conservation of energy. The calculation of the dynamic terrain shading factor combines the effects of the shading coefficient and the elevation correction factor: Dynamic terrain shading factor = Shading coefficient × Elevation correction factor × Terrain complexity adjustment coefficient. The terrain complexity adjustment coefficient considers the overall complexity of the terrain, including the frequency and amplitude of terrain undulations. The final temperature difference compensation amount ensures that the basic amount of temperature difference compensation can accurately reflect the impact of terrain shading on the heat exchange process.

[0078] The method for calculating the three-dimensional shading angle from the azimuth data and elevation profile data through three-dimensional terrain analysis includes:

[0079] A spatial straight-line vector is constructed based on the three-dimensional coordinates of the geometric centers of the key urban heat island area and the key hillside cold lake area. A continuous sequence of topographic elevation profiles along these spatial straight-line vectors is generated based on the elevation and spatial location information from the geographic information data. The spatial straight-line vectors are constructed using the three-dimensional coordinates of the geometric centers of the key urban heat island area and the key hillside cold lake area, including longitude, latitude, and elevation information. The geometric center coordinates are obtained by calculating the area-weighted average of the coordinates of all grid cells within the area, ensuring that the center position represents the spatial characteristics of the entire area. The direction cosine of the spatial straight-line vector determines the direction of subsequent topographic profile analysis. During the generation of the continuous topographic elevation profile sequence, a sampling point is set every 50 meters along the spatial straight-line vector, and the elevation of the corresponding location is obtained from the digital elevation model using bilinear interpolation. The sampling density is chosen to balance computational accuracy and efficiency; excessively dense sampling increases the computational burden, while excessively sparse sampling may miss important topographic features. The continuous topographic elevation profile sequence forms a one-dimensional elevation change curve, fully describing the topographic undulations between the two key areas.

[0080] The elevation offset of each sampling point in the continuous terrain elevation profile sequence relative to the spatial straight line vector is calculated, and a terrain shading profile curve is formed based on the elevation offset. The elevation maxima in the terrain shading profile curve are identified as key shading points. The calculation of the elevation offset considers the degree of deviation of the actual terrain from the ideal straight path. The ideal straight path is the theoretical elevation of the straight line connecting two geometric centers at different locations; the difference between the actual terrain elevation and the theoretical elevation is the elevation offset. A positive offset indicates that the terrain is higher than the ideal path and may form a shading obstacle; a negative offset indicates that the terrain is lower than the ideal path and will not produce a shading effect. The terrain shading profile curve is formed by connecting all positive offset points, clearly showing the distribution of terrain obstacles along the path. The identification of key shading points uses an extreme value detection algorithm to find the elevation maxima in the terrain shading profile curve, representing the locations of terrain obstacles that have the greatest impact on the exchange of hot and cold air currents.

[0081] The three-dimensional occlusion angle is obtained by calculating the three-dimensional spatial angle between the line connecting the starting and ending points of the spatial straight-line vector and the key occlusion point. The three-dimensional occlusion angle is calculated using a three-dimensional spatial geometry method: Three-dimensional occlusion angle = arccos[(vector OA·vector OB) / (|vector OA|×|vector OB|)], where O is the geometric center of the starting point, A is the geometric center of the ending point, and B is the key occlusion point. The three-dimensional occlusion angle reflects the degree of obstruction by terrain obstacles when observing the ending point from the starting point.

[0082] The method for obtaining the elevation correction factor based on the analysis of cold air flow kinetic energy loss according to the minimum elevation difference includes:

[0083] The theoretical kinetic energy loss of cold air during its ascent is calculated based on the minimum elevation difference data. This theoretical kinetic energy loss is then dimensionlessly processed to obtain an initial correction coefficient. The initial correction coefficient is then adjusted based on the slope of the key cold lake area on the hillside to obtain an elevation correction factor. When cold air rises from a lower to a higher altitude, it needs to overcome gravity, and according to the principle of kinetic energy to gravitational potential energy conversion in physics, its kinetic energy will decrease accordingly. Given the minimum elevation difference data, i.e., the vertical height the cold air needs to climb, the theoretical kinetic energy loss during the ascent can be calculated using the kinetic energy theorem, combined with the cold air's mass, initial velocity, and other physical quantities. Dimensionlessness can be achieved using various methods, such as dividing the theoretical kinetic energy loss by a reference value with the same dimensions, such as the initial total kinetic energy of the cold air. The dimensionless initial correction coefficient eliminates the influence of dimensions, allowing for unified comparative analysis under different meteorological conditions and terrain environments. The slope of the key cold lake area on the hillside affects the ease with which cold air can ascend. The steeper the slope, the greater the gravitational force that cold air needs to overcome as it rises, resulting in faster and more severe kinetic energy loss. Conversely, the gentler the slope, the easier it is for cold air to rise, leading to relatively less kinetic energy loss. Therefore, the initial correction coefficient needs to be adjusted based on the slope. This can be achieved by establishing a functional relationship between the slope and the adjustment amount of the correction coefficient. For example, suppose the initial correction coefficient is... The slope is Then the elevation correction factor ,in It's about slope The function is determined based on actual observation data or theoretical analysis, and generally increases with the increase of slope.

[0084] Example 2: Based on the same inventive concept, such as Figure 2 As shown, this embodiment also provides a temperature forecasting system based on deep learning. The system includes: a data acquisition and key area marking module, a preliminary gridded temperature forecast field generation module, a final temperature difference compensation generation module, and a temperature forecast field correction and temperature forecasting module. The modules are connected in sequence.

[0085] The data acquisition and key area marking module is used to acquire meteorological observation data, geographic information data, and land cover characteristic data of the target area; and to identify and mark the key urban heat island area and the adjacent cold lake area on the hillside within the target area based on the geographic information data and land cover characteristic data; the geographic information data includes slope, altitude, and spatial information; the land cover characteristic data includes building density and vegetation coverage.

[0086] The preliminary gridded temperature forecast field generation module is used to standardize and fuse the meteorological observation data, geographic information data and land cover characteristic data into a multidimensional feature tensor, and then use the multidimensional feature tensor to generate a preliminary gridded temperature forecast field through a temperature forecast model.

[0087] The final temperature difference compensation generation module is used to obtain the relative altitude difference between the key urban heat island area and the key hillside cold lake area, calculate the cold air flow intensity contribution term based on the relative altitude difference, determine the urban heat island intensity contribution term based on the building density of the key urban heat island area, obtain the heat flux conduction intensity by coupling the cold air flow intensity contribution term and the urban heat island intensity contribution term, and generate the final temperature difference compensation term by combining the distance attenuation effect between the key urban heat island area and the key hillside cold lake area.

[0088] The temperature forecast field correction and temperature forecast module is used to correct the temperature forecast values ​​corresponding to the key urban heat island area and the key hillside cold lake area in the preliminary gridded temperature forecast field according to the final temperature difference compensation amount to obtain the final gridded temperature forecast field, and to make temperature forecasts based on the final gridded temperature forecast field.

[0089] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0090] Finally, it should be noted that although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A temperature forecasting method based on deep learning, characterized in that, The method includes: Acquire meteorological observation data, geographic information data, and land cover characteristic data of the target area; identify and mark key urban heat island areas and adjacent cold lake areas on hillsides within the target area based on the geographic information data and land cover characteristic data; the geographic information data includes slope, altitude, and spatial information; the land cover characteristic data includes building density and vegetation coverage. The meteorological observation data, geographic information data and land cover characteristic data are standardized and fused into a multidimensional feature tensor. The multidimensional feature tensor is then used to generate a preliminary gridded temperature forecast field through a temperature forecast model. The relative elevation difference between the key urban heat island region and the key hillside cold lake region is obtained, and the cold air flow intensity contribution is calculated based on the relative elevation difference. The urban heat island intensity contribution is determined based on the building density of the key urban heat island region. The cold air flow intensity contribution and the urban heat island intensity contribution are coupled and analyzed to obtain the heat flux conduction intensity. Finally, the temperature difference compensation is generated by combining the distance attenuation effect between the key urban heat island region and the key hillside cold lake region. The temperature forecast values ​​corresponding to the key urban heat island area and the key hillside cold lake area in the preliminary gridded temperature forecast field are corrected according to the final temperature difference compensation amount to obtain the final gridded temperature forecast field, and the temperature is forecasted according to the final gridded temperature forecast field. The method for obtaining heat flux conduction intensity by coupling analysis of the cold air flow intensity contribution term and the urban heat island intensity contribution term, and generating the final temperature difference compensation amount by combining the distance attenuation effect between the key area of ​​the urban heat island and the key area of ​​the hillside cold lake, includes: The heat flux conduction intensity of the cold air flow intensity contribution and the urban heat island intensity contribution is quantified in microscale space by nonlinear product operation. The Euclidean distance is calculated based on the geometric center of the key area of ​​the urban heat island and the geometric center of the key area of ​​the hillside cold lake. The distance attenuation function is established based on the heat diffusion attenuation law, and the real-time attenuation coefficient is calculated. The heat flux conduction intensity and the real-time attenuation coefficient are weighted and fused to generate the basic amount of temperature difference compensation; the basic amount of temperature difference compensation is corrected for terrain shading effect based on the relative positional topological relationship between the key urban heat island area and the key hillside cold lake area, and then the final temperature difference compensation amount is output. The method for correcting the temperature difference compensation baseline based on the topological relationship between the key urban heat island area and the key hillside cold lake area, and then outputting the final temperature difference compensation amount, includes: Based on the azimuth and elevation profile data of the geometric center of the key urban heat island area and the key hillside cold lake area, the three-dimensional shading angle is calculated through three-dimensional terrain analysis, and the shading coefficient is determined based on the three-dimensional shading angle; the three-dimensional shading angle and the shading coefficient are negatively correlated. The minimum elevation difference data between the key areas of the urban heat island and the key areas of the cold lake on the hillside is obtained. The elevation correction factor is obtained by analyzing the kinetic energy loss of cold air flow based on the minimum elevation difference. The dynamic terrain shading factor is calculated based on the shading coefficient and the elevation correction factor. The final corrected temperature difference compensation amount is obtained by multiplying the dynamic terrain shading factor with the basic amount of temperature difference compensation. The method for generating a preliminary gridded temperature forecast field from the multidimensional feature tensor using a temperature forecast model includes: The multi-dimensional feature tensor is extracted by multi-layer convolution through the encoder network of the temperature forecast model to obtain a multi-scale meteorological feature map; the multi-scale meteorological feature map is then upsampled through the decoder network of the temperature forecast model to restore the spatial resolution and fused with the meteorological feature map of the corresponding scale from the encoder network to generate a grid resolution fused feature map. The grid resolution fusion feature map is processed through a regression output layer to obtain a preliminary gridded temperature forecast field covering the target area; the temperature forecast model is a convolutional neural network with an encoder-decoder architecture.

2. The deep learning-based temperature forecasting method according to claim 1, characterized in that, The method for calculating the cold air flow intensity contribution based on the relative altitude difference and determining the urban heat island intensity contribution based on the building density of the key urban heat island area includes: The relative altitude difference is used to calculate the cold air flow intensity contribution term through the cold air flow intensity function, which represents the positive correlation between the relative altitude difference and the cold air sinking speed. Obtain building density data for key areas of the urban heat island, and use the building density data to determine the contribution of urban heat island intensity through an urban heat island intensity mapping function.

3. The deep learning-based temperature forecasting method according to claim 1, characterized in that, The method for quantifying the heat flux conduction intensity of the cold air flow intensity contribution and the urban heat island intensity contribution in microscale space through nonlinear multiplication includes: The cold air flow intensity contribution term is defined as the equivalent thermal conductivity of the cold source, and the urban heat island intensity contribution term is defined as the heat source intensity value; a nonlinear product relationship function between the equivalent thermal conductivity of the cold source and the heat source intensity value is established according to Fourier's law of heat conduction. The equivalent thermal conductivity of the cold source and the intensity value of the heat source are used to calculate the microscale heat exchange flux value through the nonlinear product relationship function; the heat flux conduction intensity is generated by integrating the microscale heat exchange flux value over a unit area, and the heat flux conduction intensity characterizes the amount of heat energy transferred through a unit area per unit time.

4. The deep learning-based temperature forecasting method according to claim 1, characterized in that, The method for establishing a distance attenuation function based on the Euclidean distance according to the thermal diffusion attenuation law and calculating the real-time attenuation coefficient includes: Based on the diffusion equation of Fourier's law of heat conduction, a distance decay function is constructed with the Euclidean distance as the independent variable, and the initial thermal diffusion decay factor is calculated through the distance decay function. The elevation gradient data of the key area of ​​the cold lake on the hillside is obtained to generate a terrain dynamic enhancement coefficient; the initial thermal diffusion attenuation factor and the terrain dynamic enhancement coefficient are dynamically weighted and fused to obtain a terrain correction attenuation factor. The real-time attenuation coefficient is obtained by correcting the heat exchange path of the terrain correction attenuation factor based on the relative azimuth angle data between the key urban heat island area and the key hillside cold lake area.

5. The deep learning-based temperature forecasting method according to claim 1, characterized in that, The method for calculating the three-dimensional shading angle from the azimuth data and elevation profile data through three-dimensional terrain analysis includes: A spatial straight line vector is constructed based on the three-dimensional coordinates of the geometric center of the key area of ​​the urban heat island and the three-dimensional coordinates of the geometric center of the key area of ​​the cold lake on the hillside; a continuous topographic elevation profile sequence along the spatial straight line vector is generated based on the altitude and spatial location information in the geographic information data; Calculate the elevation offset of each sampling point in the continuous terrain elevation profile sequence relative to the spatial straight line vector, and form a terrain shading profile curve based on the elevation offset. Identify the elevation maxima points in the terrain shading profile curve as key shading points. The three-dimensional occlusion angle is obtained by calculating the three-dimensional spatial angle between the line connecting the start and end points of the spatial straight line vector and the key occlusion point.

6. The deep learning-based temperature forecasting method according to claim 1, characterized in that, The method for obtaining the elevation correction factor based on the analysis of cold air flow kinetic energy loss according to the minimum elevation difference includes: The theoretical kinetic energy loss value of cold air ascent is calculated based on the minimum elevation difference data; the theoretical kinetic energy loss value is dimensionless to obtain the initial correction coefficient; the initial correction coefficient is corrected according to the slope of the key area of ​​the cold lake on the hillside to obtain the elevation correction factor.

7. A temperature forecasting system based on deep learning, characterized in that, The system includes: a data acquisition and key area marking module, a preliminary gridded temperature forecast field generation module, a final temperature difference compensation generation module, and a temperature forecast field correction and temperature forecast module, with each module connected in a sequential communication manner; The data acquisition and key area marking module is used to acquire meteorological observation data, geographic information data, and land cover characteristic data of the target area; and to identify and mark the key urban heat island area and the adjacent cold lake area on the hillside within the target area based on the geographic information data and land cover characteristic data; the geographic information data includes slope, altitude, and spatial information; the land cover characteristic data includes building density and vegetation coverage. The preliminary gridded temperature forecast field generation module is used to standardize and fuse the meteorological observation data, geographic information data and land cover characteristic data into a multidimensional feature tensor, and then use the multidimensional feature tensor to generate a preliminary gridded temperature forecast field through a temperature forecast model. The final temperature difference compensation generation module is used to obtain the relative altitude difference between the key urban heat island area and the key hillside cold lake area, calculate the cold air flow intensity contribution term based on the relative altitude difference, determine the urban heat island intensity contribution term based on the building density of the key urban heat island area, obtain the heat flux conduction intensity by coupling the cold air flow intensity contribution term and the urban heat island intensity contribution term, and generate the final temperature difference compensation term by combining the distance attenuation effect between the key urban heat island area and the key hillside cold lake area. The temperature forecast field correction and temperature forecast module is used to correct the temperature forecast values ​​corresponding to the key urban heat island area and the key hillside cold lake area in the preliminary gridded temperature forecast field according to the final temperature difference compensation amount to obtain the final gridded temperature forecast field, and to make temperature forecasts based on the final gridded temperature forecast field; The method for obtaining heat flux conduction intensity by coupling analysis of the cold air flow intensity contribution term and the urban heat island intensity contribution term, and generating the final temperature difference compensation amount by combining the distance attenuation effect between the key area of ​​the urban heat island and the key area of ​​the hillside cold lake, includes: The heat flux conduction intensity of the cold air flow intensity contribution and the urban heat island intensity contribution is quantified in microscale space by nonlinear product operation. The Euclidean distance is calculated based on the geometric center of the key area of ​​the urban heat island and the geometric center of the key area of ​​the hillside cold lake. The distance attenuation function is established based on the heat diffusion attenuation law, and the real-time attenuation coefficient is calculated. The heat flux conduction intensity and the real-time attenuation coefficient are weighted and fused to generate the basic amount of temperature difference compensation; the basic amount of temperature difference compensation is corrected for terrain shading effect based on the relative positional topological relationship between the key urban heat island area and the key hillside cold lake area, and then the final temperature difference compensation amount is output. The method for correcting the temperature difference compensation baseline based on the topological relationship between the key urban heat island area and the key hillside cold lake area, and then outputting the final temperature difference compensation amount, includes: Based on the azimuth and elevation profile data of the geometric center of the key urban heat island area and the key hillside cold lake area, the three-dimensional shading angle is calculated through three-dimensional terrain analysis, and the shading coefficient is determined based on the three-dimensional shading angle; the three-dimensional shading angle and the shading coefficient are negatively correlated. The minimum elevation difference data between the key areas of the urban heat island and the key areas of the cold lake on the hillside is obtained. The elevation correction factor is obtained by analyzing the kinetic energy loss of cold air flow based on the minimum elevation difference. The dynamic terrain shading factor is calculated based on the shading coefficient and the elevation correction factor. The final corrected temperature difference compensation amount is obtained by multiplying the dynamic terrain shading factor with the basic amount of temperature difference compensation. The method for generating a preliminary gridded temperature forecast field from the multidimensional feature tensor using a temperature forecast model includes: The multi-dimensional feature tensor is extracted by multi-layer convolution through the encoder network of the temperature forecast model to obtain a multi-scale meteorological feature map; the multi-scale meteorological feature map is then upsampled through the decoder network of the temperature forecast model to restore the spatial resolution and fused with the meteorological feature map of the corresponding scale from the encoder network to generate a grid resolution fused feature map. The grid resolution fusion feature map is processed through a regression output layer to obtain a preliminary gridded temperature forecast field covering the target area; the temperature forecast model is a convolutional neural network with an encoder-decoder architecture.

Images