Weather Clock Forecasting System Based on Multi-Source Data Fusion

Abstract

The present invention relates to the technical field of weather forecasting, and specifically relates to a weather clock forecasting system based on multi-source data fusion, including a device terminal and a cloud. The cloud includes a data acquisition module, a data processing module, and a data transmission module. The data acquisition module is used to collect the location information of the device terminal and multi-source data at that location. The data processing module divides the target area range into grid cells based on the location information of the device terminal, and the data processing module is also used to map the location information of the device terminal into the grid cells to obtain the target unit grid where the device terminal is located. The present invention corrects the fine meteorological data of the target grid cell by introducing local characteristic data, so that the finally obtained meteorological correction data has a more refined spatial characterization ability, overcoming the defect of inaccurate prediction caused by traditional meteorological models ignoring local characteristic data under complex terrain or urban microclimate.

Description

Technical Field

[0001] The present invention relates to the technical field of weather forecasting, and particularly to a weather clock forecasting system based on multi-source data fusion. Background Art

[0002] Meteorological services are an important part of modern society, and their accuracy and timeliness directly affect people's production activities, social operation, and daily life. With the development of modern technology, weather forecasting has changed from traditional empirical forecasting to numerical weather forecasting as the dominant, and has gradually incorporated advanced technologies such as big data, artificial intelligence, and the Internet of Things. The weather trends within a day or a certain time period are clearly and dynamically displayed to users through a weather clock, which can provide more convenient information services for users. However, there are still problems such as low spatial resolution, time delay, and insufficient accuracy in the prediction of a single data source.

[0003] After retrieval, Chinese Patent No. CN202211010895.8 discloses an emergency short-term and imminent forecasting method based on multi-source data fusion. A sequence of multi-temporal and multi-channel satellite cloud images, a sequence of multi-temporal radar echo images, and basic meteorological forecasting factors at the same time series are packaged as data samples, preprocessed and data enhanced to obtain training samples; a network for fusing multi-source data is constructed, and the training samples are input into the network for training to obtain a trained multi-source data fusion network; a sequence of multi-temporal and multi-channel satellite cloud images, a sequence of multi-temporal radar echo images, and basic meteorological forecasting factors at the same time series within the effective time of short-term and imminent forecasting before the prediction area are input into the trained multi-source data fusion network to predict the weather in the area for several future time periods. The above solution effectively fuses multiple data sources, compensates for the missing interaction information due to different data patterns, and greatly improves the accuracy of forecasting emergency weather events and provides strong warning capabilities.

[0004] However, although the existing weather clock forecasting system improves the forecasting accuracy through historical deviation correction of grid data, there are still significant deficiencies in local and short-term accurate forecasting; local weather is significantly affected by terrain (such as mountains, valleys, and plains). For example, precipitation may increase due to terrain uplift in mountainous areas, but the existing correction methods are usually based on large-scale grid data and do not fully combine high-resolution terrain data for optimization; in addition, urban microclimate also has a significant impact on the temperature and wind field distribution in the area, but the correction of grid data fails to fully capture these local characteristics.

[0005] Therefore, a weather clock forecasting system based on multi-source data fusion is proposed to solve the above-mentioned problems. Summary of the Invention

[0006] Technical Problems to be Solved

[0007] In view of the above disadvantages of the existing technology, the present invention provides a weather clock forecasting system based on multi-source data fusion, which can effectively solve the problem that the historical deviation correction of large-scale grid data in the existing technology ignores the influence of local characteristics.

[0008] Technical solution

[0009] To achieve the above object, the present invention is realized through the following technical solutions:

[0010] The present invention provides a weather clock forecasting system based on multi-source data fusion, including a device terminal and a cloud. The cloud includes a data acquisition module, a data processing module and a data transmission module. The data acquisition module is used to collect the location information of the device terminal and multi-source data at this location. The data processing module divides the target area range into grid cells based on the location information of the device terminal, and the data processing module is also used to map the location information of the device terminal into the grid cell to obtain the target unit grid where the device terminal is located; input the target grid cell and multi-source data into the meteorological matching model, and calculate to obtain the fine meteorological data F of the target grid cell qp ; and input the fine meteorological data F qp into the pre-trained meteorological correction model, and combine with multi-source data to obtain the meteorological correction data F' of the target grid cell qp ; input the meteorological correction data F' qp into the pre-trained weather prediction model to obtain the weather prediction result at the location of the device terminal; the data transmission module is used to transmit the weather prediction result to the device terminal.

[0011] Furthermore, the multi-source data includes:

[0012] Meteorological observation data: temperature, humidity, air pressure, wind speed and direction, precipitation, ultraviolet intensity and particle concentration;

[0013] Meteorological model numerical prediction data: temperature field, air pressure field, wind field, humidity field and precipitation probability;

[0014] Historical meteorological data: daily temperature, daily precipitation, daily wind speed and direction, and periodic patterns;

[0015] Local characteristic data: micro-topographic changes, urban microclimate, water body details, vegetation type and distribution, soil conditions and local air pollution.

[0016] Furthermore, the method for dividing the target area range into grid cells includes:

[0017] Define the central coordinates (lat device , lon device) and the spatial radius R of the surrounding area region ; Convert the spatial radius R region into the difference in latitude and longitude values. The conversion formula is:

[0018] In the formula, Δlat is the distance in the latitude direction; Δlon is the distance in the longitude direction; cos(lat device ) represents the current latitude of the device terminal;

[0019] Define the target area range A region = {(lat min , lat max ), (lon min , lon max )};

[0020] In the formula, (lat min , lat max ) represents the boundary values of latitude in the north-south direction, and (lon min , lon max ) represents the boundary values of longitude in the east-west direction; lat min = lat device - Δlat, lat max = lat device + Δlat, lon min = lon device - Δlon, lon max = lon device + Δlon;

[0021] Divide the target area range into regular grid cells based on the set grid resolution (Δx, Δy); where Δx is the longitude difference and Δy is the latitude difference Δy;

[0022] Generate the latitude index lat region and longitude index lon q for each grid cell based on the target area range A p : The generation formula is:

[0023] lat q = lat min + (q - 1)·Δy, q = 1, 2,..., n lat ; In the formula, lat q represents the minimum latitude of the qth grid cell; n lat represents the total number of grid cells in the latitude direction, and the calculation formula is

[0024] lon p = lon min+(p - 1)·Δx, p = 1, 2, ..., m lon ; lon p represents the minimum longitude of the p-th grid cell; m lon represents the total number of grid cells in the latitude direction, and the calculation formula is;

[0025]

[0026] Based on the latitude index lat q and longitude index lon p of each grid cell, calculate the center point coordinates (lat center , lon center ) of the grid cell and generate a unique identifier ID for each grid cell qp , where q and p correspond to the latitude index and longitude index of the grid cell respectively,

[0027] Furthermore, the method of mapping the location information of the device terminal to the grid cell includes:

[0028] Calculate the grid number q of the device terminal in latitude device and the grid number p in longitude device , to obtain the target grid cell G where the device terminal is located qp and its center point coordinates The calculation formula is:

[0029]

[0030] Furthermore, the calculation method of the fine meteorological data F qp is:

[0031] Input the center point coordinates of the target grid cell and the meteorological model numerical prediction data into the meteorological matching model, and calculate the distance between the center point coordinates of the target grid cell and the center point coordinates (lat k , lon k ) of all meteorological model grids. The calculation formula is:

[0032]

[0033] In the formula, k is the index of the meteorological model grid;

[0034] Find the meteorological model grid with the minimum distance D and use it as the object aligned with the target grid cell; extract its meteorological data F k ={T k , P k , U k , Rk , V k , D k , C k} as the initial meteorological data of the target grid cell G qp ; where, T k represents the air temperature in the corresponding meteorological model grid; P k represents the probability of rainfall in the corresponding meteorological model grid; U k represents the humidity in the corresponding meteorological model grid; R k represents the ultraviolet intensity in the corresponding meteorological model grid; V k represents the average wind speed in the corresponding meteorological model grid; D k represents the direction in the corresponding meteorological model grid; C k represents the particulate matter concentration in the corresponding meteorological model grid;

[0035] Compare the grid resolution (Δx, Δy) of the target grid cell and the grid resolution (Δx w , Δy w ) of the meteorological model grid. If the grid resolution of the meteorological model grid is coarser than that of the grid cell, update the initial meteorological data through the interpolation algorithm, and the update formula is:

[0036] In the formula, α k is the interpolation weight, and its calculation formula is: m = 1, 2, 3, 4 represents the numbers of the four adjacent meteorological model grids of the target grid cell.

[0037] Further, the calculation formula of the meteorological correction data F′ qp is:

[0038] F′ qp = F qp + ΔF terrain + ΔF urban + ΔF water + ΔF vegetation + ΔF soil + ΔF pollution + ΔF ;

[0039] In the formula, ΔF terrain is the micro-topography change correction amount; ΔF urban is the urban microclimate correction amount; ΔF water is the water body detail correction amount; ΔF vegetation is the vegetation type and distribution correction amount; ΔF soil is the soil condition correction amount; ΔF pollution is the local air pollution correction amount;

[0040] And the fine meteorological data Fqp Input it into the pre-trained meteorological correction model, and combine multi-source data to obtain the meteorological correction data F′ of the target grid cell qp ; Input the meteorological correction data F′ qp into the pre-trained weather prediction model to obtain the weather prediction result at the location of the device terminal.

[0041] Furthermore, the calculation formula of the micro-topography change correction amount ΔF terrain is as follows:

[0042] ΔF terrain =δ H (H qp -H0)+δ S ·S qp +δ A ·f(A qp ); In the formula, H qp is the altitude of the target grid cell; H0 is the reference altitude; S qp and A qp are the slope and aspect values of the target grid cell respectively; δ H , δ S , δ A are the altitude weight coefficient, slope weight coefficient and aspect weight coefficient respectively, f(A qp ) is the aspect function, f(A qp ) = cos(A qp -A ref ), A ref is the reference aspect;

[0043] The calculation formula of the urban microclimate correction amount ΔF urban is as follows:

[0044] ΔF urban =γ UHI ·U qp +γ D ·D u,qp ; In the formula, U qp is the urban heat island effect intensity value of the target grid cell, D u,qp is the building density quantization value of the target grid cell, γ UHI and γ D are the urban heat island effect weight coefficient and building density weight coefficient respectively;

[0045] The calculation formula of the water body detail correction amount ΔF water is as follows:

[0046] ΔF water =λ W ·W qp +λ E ·Ew,qp ; where, W qp is the water area ratio of the target grid cell, E w,qp is the evaporation intensity of the target grid cell, λ W and λ E are the water area ratio weight coefficient and the water evaporation intensity weight coefficient respectively;

[0047] The vegetation type and distribution correction amount ΔF vegetation is calculated as follows:

[0048] ΔF vegetation = ρ V ·V qp + ρ T ·f(T v,qp ); where, V qp is the vegetation coverage rate of the target grid cell, f(T v,qp ) is the vegetation type function, ρ V and ρ T are the vegetation coverage rate weight coefficient and the vegetation type weight coefficient respectively;

[0049] The soil condition correction amount ΔF soil is calculated as follows:

[0050] ΔF soil = υ T ·f(T s,qp ) + υ L ·L s,qp ; where, f(T s,qp ) is the soil type function, T s,qp is the soil type of the target grid cell, L s,qp is the soil utilization rate of the target grid cell, υ T and υ L are the soil type weight coefficient and the soil utilization rate weight coefficient respectively;

[0051] The local air pollution correction amount ΔF pollution is calculated as follows:

[0052] ΔF pollution = ζ PM ·PM qp + ζ E ·E p,qp ; where, PM qp is the particulate matter concentration of the target grid cell, E p,qp is the pollution source intensity of the target grid cell, ζ PM and ζ E are the particulate matter concentration weight coefficient and the pollution source intensity weight coefficient respectively.

[0053] Furthermore, the training method of the meteorological correction model is as follows:

[0054] Collect historical meteorological data of different target grid cells And process it to obtain the historical refined meteorological data of different target grid cells, which is used as the training input of the meteorological correction model;

[0055] Initialize the learnable parameters of the meteorological correction model, including the weight coefficient matrix of each influencing factor in the local feature data;

[0056] Define the loss function of the meteorological correction model

[0057] In the formula, is the predicted value after correction; is the corresponding true observed value; n is the number of samples;

[0058] Define the regularization loss function of the meteorological correction model

[0059] In the formula, ω is the regularization coefficient, θ u is the learnable parameter in the meteorological correction model, u is the index label of the weight coefficient, is the regularization term of the sum of squares of all weight coefficients;

[0060] Use the gradient descent method to update the learnable parameter θ u , to reduce the value of the loss function L reg ; Repeat the training on the historical refined meteorological data until the meteorological correction model converges or reaches the pre-set number of iterations in advance, that is, the training of the meteorological correction model is completed.

[0061] Furthermore, the training method of the weather prediction model includes:

[0062] Define the long short-term memory network model as the basic structure of the weather prediction model, and the basic structure includes an input layer, a hidden layer, and an output layer;

[0063] Introduce the parameter vector time step d, and combine the vector time step d, the meteorological correction data F′ qp and the historical meteorological data to obtain the fused feature vector where s is the number of vector time steps d;

[0064] Collect the historical meteorological data of the target grid cell within the past fixed time And correspondingly construct it into a fused feature vector, denoted as the historical fused feature vector Use the historical fused feature vector Expand according to the vector time step to form sequence data as the training input of the weather prediction model. Correspondingly, label the sequence data according to the vector time step d, and the label is the historical weather result.

[0065] Initialize the network parameters of the weather prediction model; the network parameters include the weight matrix W from the input layer to the hidden layer. x and the recurrent weight matrix W of the hidden layer. h and the bias vector b of the hidden layer. h and the weight matrix W from the hidden layer to the output layer. o and the bias vector b of the output layer. o ;

[0066] Define the loss function of the weather prediction model.

[0067] In the formula, S is the length of the sequence data; Y true,d is the label annotated at the d-th vector time step, that is, the historical weather result. Y pred,d is the label annotated at the d-th vector time step, that is, the predicted weather prediction result.

[0068] Pass the sequence data to the input layer, then pass through the hidden layer and the output layer in sequence, and output the predicted weather prediction result. And calculate the corresponding weather prediction result. Calculate the value of the loss function of the weather prediction result and backpropagate the error gradient from the output layer to the input layer; according to the error gradient, use the optimization algorithm to update the network parameters to reduce the value of the loss function; repeat the training of the sequence data until the weather prediction model converges or reaches the preset number of iterations in advance, which means the training of the weather prediction model is completed.

[0069] Furthermore, the input layer is used to receive the fused feature vector X d as the input; the hidden layer consists of y state vectors, and the calculation formula for each state vector is: h d = f(W h ·h d-1 + W x ·X d + b h ); In the formula, h d is the hidden state at the current vector time step d, f(...) is the activation function, W x is the weight matrix from the input layer to the hidden layer, W h is the recurrent weight matrix of the hidden layer, and b h is the bias vector of the hidden layer;

[0070] The output layer calculates the output result of the weather prediction model based on the output of the hidden layer, that is, the weather prediction result Its calculation formula is: where η(...) is the activation function of the output layer, and W o is the weight matrix from the hidden layer to the output layer, and b o is the bias vector of the output layer.

[0071] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are as follows:

[0072] In this solution, the grid subdivision module takes the longitude and latitude coordinates of the device terminal as the center, divides the target area range of the device terminal into grids, obtains standard grid cells, and maps the location information of the device terminal to the corresponding grid cells; compared with the traditional weather that usually uses regularly gridded large areas (such as urban or national geographical areas), this solution can dynamically generate local grids centered on the device terminal, perform refined analysis of multi-source data in a smaller spatial range, better reflect small-scale weather differences within the area, such as local rainfall and wind speed changes; and compare the weather conditions (such as precipitation, wind force, temperature, etc.) around the target area range with a larger area, so as to more accurately identify weather anomaly data related to the location of the device terminal; and by dynamically generating grid cells centered on the location of the device terminal, it can reduce meteorological calculations for non-essential areas, optimize calculation performance, and improve the response speed of local weather forecasts;

[0073] In this solution, local feature data is introduced to correct the fine meteorological data of the target grid cell, so that the finally obtained meteorological correction data has a more refined spatial characterization ability, overcoming the defect of inaccurate prediction caused by traditional meteorological models ignoring local feature data in complex terrain or urban microclimate; and coping with the challenges of environmental diversity, it can adapt to diverse landform environments such as mountains, plains, and cities, enhance the sensitivity of data to extreme meteorological phenomena by dynamically adjusting the meteorological correction model, and at the same time make up for the deficiency of traditional numerical weather prediction models in responding to small-scale and large-gradient change characteristics, improving the early warning ability of extreme weather. Brief Description of the Drawings

[0074] Figure 1 It is a schematic diagram of the structural composition of the weather clock forecasting system in the embodiment of the present invention. Detailed Embodiments

[0075] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0076] Embodiment 1:

[0077] Reference Figure 1 , this case proposes a weather clock forecasting system based on multi-source data fusion, including a device terminal and a cloud. The device terminal and the cloud are connected through a wireless signal. The cloud includes a data acquisition module, a data processing module, and a data transmission module. The data acquisition module is used to obtain the location information of the device terminal and obtain multi-source data of the location based on the location information of the device terminal. The data processing module divides the target area range into grid cells based on the location information of the device terminal, and obtains grid cells. Then, the location information of the device terminal is mapped into the grid cells to obtain the target unit grid where the device terminal is located. The target grid cell and the multi-source data are input into the meteorological matching model, and the fine meteorological data F of the target grid cell is calculated qp ; The fine meteorological data F of the target grid cell qp is input into the pre-trained meteorological correction model, and the meteorological correction data F' of the target grid cell is obtained by combining multi-source data q ′ p ; Finally, the meteorological correction data F' q ′ p is input into the pre-trained weather prediction model to obtain the weather prediction result at the location of the device terminal. The data transmission module is used to send the weather prediction result to the device terminal for dynamic display

[0078] Specifically, in this solution, the data acquisition module obtains the location information of the device terminal based on the Location-Based Service (LBS) technology: LBS(t) = {ID i , (lat i , lon i ), R accuracy}; In the formula, ID i is the unique identifier of the device terminal (such as device number, user ID), which is used to distinguish different device terminals and provide end-to-end services; (lat i , lon i ) is the latitude and longitude coordinates of the current location of the device terminal; R accuracy is the error radius of the location information of the device terminal; i is the unique identifier index of the device terminal; t is the timestamp, which represents the acquisition time of the location information of the device terminal, that is, the time dimension of the location information data

[0079] The data acquisition module uses the location information LBS(t) of the device terminal to obtain multi-source data at that location, including: meteorological observation data: temperature, humidity, air pressure, wind speed and direction, precipitation, ultraviolet intensity, and particle concentration; meteorological model numerical prediction data: temperature field, air pressure field, wind field, humidity field, and precipitation probability; historical meteorological data: daily temperature, daily precipitation, daily wind speed and direction, and periodic patterns; local feature data: micro-topographic changes, urban microclimate, water body details, vegetation type and distribution, soil conditions, and local air pollution.

[0080] More specifically, in this case, the collected location information LBS(t) and multi-source data are input into the data processing module; the data processing module cleans and calibrates the collected multi-source data, removes outliers from the data and fills in missing values, and then performs feature normalization on the data to map the value range to a certain range; then the location information LBS(t) is input into the grid subdivision module, and the grid subdivision module uses the longitude and latitude coordinates (lat i , lon i ) of the device terminal as the center to divide the surrounding area of the device terminal into grids (i.e., the target area range), and the division method is as follows:

[0081] Define the central coordinates (lat device , lon device ) with the longitude and latitude coordinates of the device terminal as the center point, and the spatial radius R region of the surrounding area; convert the spatial radius R region into longitude and latitude value differences, and the conversion formula is: and In the formula, Δlat is the distance in the latitude direction, representing the change value of latitude within the target area range; Δlon is the distance in the longitude direction, representing the change value of longitude within the target area range; the denominator of the conversion formula is 111, indicating that 1° of latitude or 1° of longitude is approximately equal to 111 kilometers (a fixed constant on the Earth's surface); cos(lat device ) represents the current latitude of the device terminal, that is, the cosine value of the latitude lat device with the device terminal as the center point to adjust the longitude direction distance;

[0082] Define the target area range A region = {(lat min , lat max ), (lon min , lon max )}; in the formula, (lat min , lat max ) represents the boundary values of latitude in the north-south direction, and (lon min , lon max ) represents the boundary values of longitude in the east-west direction; latmin = lat device - Δlat, lat max = lat device + Δlat, lon min = lon device - Δlon, lon max = lon device + Δlon;

[0083] The target area range is divided into regular grid cells. The division of the grid cells is based on the set grid resolution (Δx, Δy), including the longitude difference Δx and the latitude difference Δy; the longitude difference Δx represents the width of each grid cell in the longitude direction, and the latitude difference Δy represents the width of each grid cell in the latitude direction;

[0084] Based on defining the target area range A region and the grid resolution (Δx, Δy), generate the latitude index lat q and the longitude index lon p : The generation formula is:

[0085] lat q = lat min +(q - 1)·Δy, q = 1, 2,..., n lat ; In the formula, lat q represents the minimum latitude of the qth grid cell, that is, the latitude value of the southern boundary of the grid cell; (q - 1) represents that starting from the first grid cell, the current grid cell is the qth grid starting from lat min ; n lat represents the total number of grid cells in the latitude direction, and the calculation formula is

[0086] lon p = lon min +(p - 1)·Δx, p = 1, 2,..., m lon ; lon p represents the minimum longitude of the pth grid cell, that is, the longitude value of the western boundary of the grid cell; (p - 1) represents that starting from the first grid cell, the current grid cell is the pth grid starting from lon min ; m lon represents the total number of grid cells in the latitude direction, and the calculation formula is

[0087] Based on the latitude index lat q and the longitude index lon p of each grid cell, calculate the center point coordinates (lat center , loncenter ) and generate a unique identification ID for each grid cell qp , where q and p respectively correspond to the latitude index and longitude index of the grid cell

[0088] The data processing module maps the location information LBS(t) of the device terminal into the corresponding grid cell, and obtains the grid number q of the device terminal in latitude device and the grid number p in longitude device , and the mapping formula is: In the formula, q device represents the row number corresponding to the device terminal location in the grid division, and p device represents the column number corresponding to the device terminal location in the grid division represents the floor operation of the calculation result, and the grid number starts from 1, so 1 needs to be added to the rounded value; furthermore, the grid number q of the device terminal in latitude device and the grid number p in longitude device correspond to the latitude index lat q and longitude index lon p of the grid cell, that is, q device = lat q and p device = lon p , and obtain the target grid cell G where the device terminal is located qp and its center point coordinates

[0089] Based on the above content, in this solution, the grid subdivision module takes the longitude and latitude coordinates (lat i , lon i ) of the device terminal as the center, divides the target area range of the device terminal into grids, obtains standard grid cells, and maps the location information LBS(t) of the device terminal into the corresponding grid cells; compared with the traditional weather which usually uses regular grid division for large-scale areas (such as urban or national geographical areas), this solution can dynamically generate local grids centered on the device terminal, can perform refined analysis of multi-source data in a smaller spatial range, can better reflect small-scale weather differences within the region, such as local rainfall, wind speed changes, etc.; and compare the weather conditions (such as precipitation, wind force, temperature, etc.) around the target area range with a larger range, can more accurately identify weather anomaly data related to the location of the device terminal; and by dynamically generating grid cells centered on the location of the device terminal, it can reduce meteorological calculations for non-essential areas, optimize calculation performance, and improve the response speed of local weather forecasts

[0090] Further, in this case, the data processing module inputs the central point coordinates of the target grid cell where the device terminal is located and the numerical weather prediction data of the meteorological model into the meteorological matching model, and corresponds the central point of the target grid cell to the central point coordinates (lat k , lon k ) of the meteorological model grid, and finds the nearest meteorological model grid;

[0091] The corresponding method between the target grid cell and the meteorological model grid is: calculate the distance between the central point coordinates of the target grid cell and the central point coordinates of all meteorological model grids, and the calculation formula is: where k is the index of the meteorological model grid; then find the meteorological model grid with the smallest distance D, and use it as the object aligned with the target grid cell; extract its meteorological data as the initial meteorological data of the target grid cell G qp ; extract the meteorological data F k = {T k , P k , U k , R k , V k , D k , C k}; where T k represents the air temperature in the corresponding meteorological model grid; P k represents the probability of rainfall in the corresponding meteorological model grid; U k represents the humidity in the corresponding meteorological model grid; R k represents the ultraviolet intensity in the corresponding meteorological model grid; V k represents the average wind speed in the corresponding meteorological model grid; D k represents the direction in the corresponding meteorological model grid; C k represents the particulate matter concentration in the corresponding meteorological model grid;

[0092] Compare the grid resolution (Δx, Δy) of the target grid cell and the grid resolution (Δx w , Δy w ) of the meteorological model grid. If the grid resolution of the meteorological model grid is coarser than that of the grid cell, update the initial meteorological data through the difference algorithm, generate finer meteorological data for the target grid cell and represent it as a vector variable, that is, the fine meteorological data F qp of the target grid cell, and its calculation formula is: In the formula, α kis the interpolation weight, which represents the relative distance between the target grid cell and each neighboring point (i.e., the adjacent meteorological model grid), and is used to determine the fine meteorological data F of each neighboring point to the target grid cell. qp The contribution degree is calculated as follows: m=1, 2, 3, 4 represents the numbers of the four meteorological model grids adjacent to the target grid cell. When the meteorological model grid has a lower resolution, the target grid cell requires more refined meteorological data, and the meteorological information of the target grid cell is uniformly adjusted through interpolation.

[0093] Furthermore, in this case, the data processing module is based on the local characteristic data and the fine meteorological data F of the target grid unit. qp Build and train the weather correction model to obtain the weather correction data F′ of the target grid cell qp , improve the accuracy of subsequent weather forecast results;

[0094] Specifically, the construction method of the meteorological correction model includes:

[0095] Based on the linear superposition method, the contribution of each influencing factor to the meteorological variables in the local characteristic data is corrected to define the meteorological correction data F′ qp The calculation formula is:

[0096] F′ qp =F qp +ΔF terrain +ΔF urban +ΔF water +ΔF vegetation +ΔF soil +ΔF pollution ;

[0097] Where, ΔF terrain ΔF is the correction for microtopography changes, including the effects of changes in altitude, slope, and aspect parameters. Temperature generally decreases with increasing altitude, typically by about 0.65°C / 100m. Slope has a significant impact on precipitation collection and surface runoff, indirectly regulating humidity and precipitation. Aspect affects the distribution of solar radiation, which in turn affects temperature (e.g., sunny slopes in the Northern Hemisphere are hotter). The calculation formula is: terrain =δ H (H qp -H0)+δ S ·S qp +δ A ·f(A qp ); where H qp is the altitude of the target grid cell; H0 is the reference altitude; S qp and A qp are the slope and aspect values of the target grid cell respectively; δ H , δS , δ A are the altitude weight coefficient, slope weight coefficient, and aspect weight coefficient respectively, determined based on the actual experience of the region; f(A qp ) is the aspect function, describing the degree of correction of the specific aspect to meteorological variables. f(A qp ) = cos(A qp - A ref ), where A ref is the reference aspect;

[0098] ΔF urban is the urban microclimate correction amount, including the effects of the urban heat island effect and building height and density. The temperature in urbanized areas is generally higher than that in surrounding areas, and the specific intensity is related to the urban form and area. High-density buildings change the wind field distribution, affecting temperature and wind speed; the calculation formula is: ΔF urban = γ UHI ·U qp + γ D ·D u,qp ; where U qp is the urban heat island effect intensity value of the target grid cell, D u,qp is the building density quantization value of the target grid cell, and γ UHI and γ D are the urban heat island effect weight coefficient and building density weight coefficient respectively;

[0099] ΔF water is the water body detail correction amount, including the effects of the water body area ratio and water body evaporation intensity. The water body has a regulatory effect on the surrounding climate, significantly affecting the temperature distribution (cooling in summer and warming in winter), and the water body evaporation intensity affects the local humidity and precipitation distribution; the calculation formula is: ΔF water = λ W ·W qp + λ E ·E w,qp ; where W qp is the water body area ratio of the target grid cell, E w,qp is the evaporation intensity of the target grid cell, and λ W and λ E are the water body area ratio weight coefficient and water body evaporation intensity weight coefficient respectively;

[0100] ΔF vegetation is the vegetation type and distribution correction amount, including the effects of vegetation coverage and vegetation type. The temperature fluctuates less and the humidity is higher in areas with high vegetation coverage. Different types of vegetation have different regulatory abilities for the regional climate through shading and transpiration; the calculation formula is: ΔF vegetation = ρ V ·V qp + ρ T ·f(Tv,qp ); where, V qp is the vegetation coverage rate of the target grid cell, and f(T v,qp ) is the vegetation type function, which is used to describe the influence degree of the vegetation type T v,qp of the target grid cell on meteorological variables and is obtained through empirical values. ρ V and ρ T are the vegetation coverage rate weight coefficient and the vegetation type weight coefficient respectively;

[0101] ΔF soil is the soil condition correction amount, including the influence of soil type and soil utilization rate. The soil thermodynamic properties determine the surface heat exchange efficiency, and utilization methods such as farming and construction change the soil albedo and drainage capacity; the calculation formula is: ΔF soil = υ T ·f(T s,qp ) + υ L ·L s,qp ; where, f(T s,qp ) is the soil type function, which is used to describe the influence degree of the soil type T s,qp of the target grid cell on meteorological variables, L s,qp is the soil utilization rate of the target grid cell, and υ T and υ L are the soil type weight coefficient and the soil utilization rate weight coefficient respectively;

[0102] ΔF pollution is the local air pollution correction amount, including the influence of atmospheric particulate matter concentration and pollution source intensity. The temperature in the high-concentration PM area drops, and industrial emissions, traffic pollution, etc. affect the local microclimate; the calculation formula is: ΔF pollution = ζ PM ·PM qp + ζ E ·E p,qp ; where, PM qp is the particulate matter concentration of the target grid cell, E p,qp is the pollution source intensity of the target grid cell, and ζ PM and ζ E are the particulate matter concentration weight coefficient and the pollution source intensity weight coefficient respectively.

[0103] The training method of the meteorological correction model is as follows: collect the historical meteorological data of different target grid cells and process them to obtain the historical fine meteorological data of different target grid cells, which are used as the training input of the meteorological correction model;

[0104] Initialize the learnable parameters of the meteorological correction model, including the weight coefficient matrix of each influencing factor in the local characteristic data;

[0105] Define the loss function of the meteorological correction model

[0106] In the formula, is the corrected predicted value, that is, the output after the meteorological correction model corrects the historical fine meteorological data of the target grid cell with latitude index q and longitude index p; is the corresponding true observed value; n is the number of samples;

[0107] Define the regularization loss function of the meteorological correction model

[0108] In the formula, ω is the regularization coefficient, and θ u are the learnable parameters in the meteorological correction model (i.e., different weight coefficients, such as the water area ratio weight coefficient λ UHI and the urban heat island effect weight coefficient γ UHI etc.), u is the index label of the weight coefficient, is the regularization term of the sum of squares of all weight coefficients, used to constrain the weight values of the model;

[0109] Use the gradient descent method to update the learnable parameter θ u , to reduce the value of the loss function L reg ; Repeat the training on the historical fine meteorological data until the meteorological correction model converges (i.e., the value of the loss function no longer changes) or reaches the pre-set number of iterations in advance, that is, the training of the meteorological correction model is completed.

[0110] Introduce local feature data to correct the fine meteorological data F qp of the target grid cell, so that the finally obtained meteorological correction data F′ qp has a finer spatial characterization ability, overcoming the defect of inaccurate prediction caused by the traditional meteorological model ignoring local feature data under complex terrain or urban microclimate; and coping with the challenge of environmental diversity, being able to adapt to diverse geomorphic environments such as mountains, plains, and cities, enhancing the sensitivity of data to extreme meteorological phenomena by dynamically adjusting the meteorological correction model, and at the same time making up for the deficiency of the traditional numerical weather prediction model's slow response to small-scale and large-gradient change characteristics, improving the early warning ability of extreme weather.

[0111] It should be noted that in this case, the training method of the weather prediction model includes:

[0112] Define the long short-term memory network model as the basic structure of the weather prediction model, and the basic structure includes an input layer, a hidden layer, and an output layer;

[0113] Introduce the parameter vector time step d, and combine the vector time step d, the meteorological correction data F′ qp and the historical meteorological data Combined to obtain a fused feature vector where s is the number of vector time steps d;

[0114] The input layer is used to receive the fused feature vector X d as input; The hidden layer consists of y state vectors, and the calculation formula for each state vector is: h d = f(W h ·h d-1 + W x ·X d + b h ); In the formula, h d is the hidden state of the current vector time step d, f(...) is the activation function, W x is the weight matrix from the input layer to the hidden layer, W h is the recurrent weight matrix of the hidden layer, b h is the bias vector of the hidden layer;

[0115] The output layer is based on the output of the hidden layer, that is, the hidden state h d of the current vector time step d, and calculates the output result of the weather prediction model, that is, the weather prediction result The calculation formula is: where η(...) is the activation function of the output layer, W o is the weight matrix from the hidden layer to the output layer, b o is the bias vector of the output layer;

[0116] Collect the historical meteorological data of the target grid cell within the past fixed time and correspondingly construct it into a fused feature vector, denoted as the historical fused feature vector The historical fused feature vector is unfolded according to the vector time step to form sequence data as the training input of the weather prediction model, and correspondingly the sequence data is labeled according to the vector time step d, and the label is the historical weather result

[0117] Initialize the network parameters of the weather prediction model; The network parameters include the weight matrix W x from the input layer to the hidden layer, the recurrent weight matrix W h of the hidden layer, the bias vector b h of the hidden layer, the weight matrix W o from the hidden layer to the output layer, and the bias vector b o of the output layer;

[0118] Define the loss function of the weather prediction model

[0119] In the formula, S is the length of the sequence data; Ytrue,d The label annotated for the d-th vector time step, i.e., the historical weather result Y pred,d is the label annotated for the d-th vector time step, i.e., the predicted weather forecast result

[0120] Pass the sequence data to the input layer, then successively through the hidden layer and the output layer, and output the predicted weather forecast result and calculate the corresponding weather forecast result Calculate the value of the loss function of [], and backpropagate the error gradient from the output layer to the input layer; according to the error gradient, use an optimization algorithm (such as gradient descent, Adam, etc.) to update the network parameters to reduce the value of the loss function; repeat the training of the sequence data until the weather forecast model converges (i.e., the value of the loss function no longer changes) or reaches the pre-set number of iterations in advance, which indicates that the training of the weather forecast model is completed.

[0121] Based on historical meteorological data, the long-term trends and periodic changes of the weather conditions in the target grid cell can be captured. Combining these characteristics can effectively improve the accuracy of short-term and medium- to long-term weather forecasts, and provide a basis for continuous optimization of the system. The model can perform online learning by continuously introducing new historical data to enhance its adaptability to different meteorological conditions.

[0122] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A weather clock forecast system based on multi-source data fusion, comprising a device terminal and a cloud. The cloud comprises a data acquisition module, a data processing module, and a data transmission module. The data acquisition module is used to collect the location information of the device terminal and the multi-source data at the location. The system is characterized in that: The data processing module divides the target area into grids based on the location information of the device terminal to obtain grid units. The data processing module is also used to map the location information of the device terminal into the grid units to obtain the target unit grid where the device terminal is located; Input the target grid cell and multi-source data into the meteorological matching model to calculate the fine meteorological data F of the target grid cell. qp ; and the detailed meteorological data F qp Input into the pre-trained meteorological correction model, and combine multi-source data to obtain the meteorological correction data F′ of the target grid cell qp ; The weather correction data F' qp Input into the pre-trained weather forecast model to obtain the weather forecast result for the location of the device terminal; The data transmission module is used to transmit the weather forecast results to the device terminal; The center point coordinates of the target grid cell The numerical forecast data of the meteorological model is input into the meteorological matching model to calculate the center point coordinates of the target grid cell and the center point coordinates of all meteorological model grids (lat k ,lon k ), find the meteorological model grid with the smallest distance D and use it as the target grid cell to align; extract its meteorological data F k ={T k ,P k ,U k ,R k ,V k ,D k ,C k } as the target grid cell G qp Initial meteorological data of k represents the temperature in the corresponding meteorological model grid; P k represents the probability of rainfall in the corresponding meteorological model grid; U k Represents the humidity in the corresponding meteorological model grid; R k Indicates the ultraviolet intensity in the corresponding meteorological model grid; V k represents the average wind speed in the corresponding meteorological model grid; D k Indicates the direction in the corresponding weather model grid; C k represents the particle concentration in the corresponding meteorological model grid; Weather correction data F′ qp The calculation formula is: F′ qp =F qp +ΔF terrain +ΔF urban +ΔF water +ΔF vegetation +ΔF soil +ΔF pollution ; Where, ΔF terrain is the correction value of micro-topography change; ΔF urban is the urban microclimate correction; ΔF water is the water detail correction; ΔF vegetation is the vegetation type and distribution correction; ΔF soil is the soil condition correction; ΔF pollution is the local air pollution correction amount; And the detailed meteorological data F qp Input into the pre-trained meteorological correction model, and combine multi-source data to obtain the meteorological correction data F′ of the target grid cell qp ; The weather correction data F' qp Input into the pre-trained weather forecast model to obtain the weather forecast result for the location of the device terminal; Micro-topography change correction ΔF terrain The calculation formula is: ΔF terrain =δ H (H qp -H0)+δ S ·S qp +δ A ·f(A qp );where H qp is the altitude of the target grid cell; H0 is the reference altitude; S qp and A qp are the slope and aspect values of the target grid cell respectively; δ H , δ S , δ A are the altitude weight coefficient, slope weight coefficient and aspect weight coefficient respectively, f(A qp ) is the slope function, f(A qp )=cos(A qp -A ref ), A ref is the reference slope aspect; The urban microclimate correction value ΔF urban The calculation formula is: ΔF urban =γ UHI ·U qp +γ D ·D u,qp Where U qp is the urban heat island effect intensity value of the target grid cell, D u,qp is the quantitative value of the building density of the target grid cell, γ UHI and γ D are the urban heat island effect weight coefficient and the building density weight coefficient respectively; Water detail correction ΔF water The calculation formula is: ΔF water =λ W W qp +λ E ·E w,qp Where W qp is the water area ratio of the target grid cell, E w,qp is the evaporation intensity of the target grid cell, λ W and λ E are the water body area ratio weight coefficient and the water body evaporation intensity weight coefficient respectively; Vegetation type and distribution correction ΔF vegetation The calculation formula is: ΔF vegetation =ρ V ·V qp +ρ T ·f(T v,qp );where V qp is the vegetation coverage of the target grid cell, f(T v,qp ) is the vegetation type function, ρ V and ρ T are the vegetation coverage weight coefficient and vegetation type weight coefficient respectively; Soil condition correction ΔF soil The calculation formula is: ΔF soil =υ T ·f(T s,qp )+υ L ·L s,qp ; Among them, f(T s,qp ) is the soil type function, T s,qp is the soil type of the target grid cell, L s,qp is the soil utilization rate of the target grid cell, υ T and υ L are soil type weight coefficient and soil utilization rate weight coefficient respectively; Local air pollution correction ΔF pollution The calculation formula is: ΔF pollution =ζ PM PM qp +ζ E ·E p,qp Where, PM qp is the particle concentration of the target grid cell, E p,qp is the pollution source intensity of the target grid unit, ζ PM and ζ E are the particulate matter concentration weight coefficient and the pollution source intensity weight coefficient respectively.

2. The weather clock forecast system based on multi-source data fusion according to claim 1 is characterized in that: The multi-source data includes: Meteorological observation data: temperature, humidity, air pressure, wind speed, wind direction, precipitation, UV intensity, and particle concentration; Meteorological model numerical forecast data: temperature field, pressure field, wind field, humidity field and precipitation probability; Historical meteorological data: daily temperature, daily precipitation, daily wind speed and direction, and periodic patterns; Local characteristic data: microtopography changes, urban microclimate, water body details, vegetation types and distribution, soil conditions and local air pollution.

3. The weather clock forecast system based on multi-source data fusion according to claim 2 is characterized in that: The method of gridding the target area includes: Define the center coordinates (lat device ,lon device ) and the spatial radius R of the surrounding area region ; Set the space radius R region Converted into longitude and latitude value difference, the conversion formula is: and Where Δlat is the distance in the latitude direction; Δlon is the distance in the longitude direction; cos(lat device ) represents the current latitude of the device terminal; Define target area range A region ={(lat min ,lat max ), (lon min ,lon max )}; Where, (lat min ,lat max ) represents the boundary value of latitude in the north-south direction, (lon min ,lon max ) represents the boundary value of longitude in the east-west direction; lat min =lat device -Δlat,lat max =lat device +Δlat,lon min =lon device -Δlon, lon max =lon device +Δlon; The target area is divided into regular grid cells based on the set grid resolution (Δx, Δy); where Δx is the longitude difference and Δy is the latitude difference; Based on target area range A region and the grid resolution (Δx, Δy) to generate the latitude index lat for each grid cell q and longitude index lon p : The generation formula is: lat q =lat min +(q-1)·Δy,q=1,2,...,n lat ; In the formula, lat q Indicates the minimum latitude of the qth grid cell; n lat Represents the total number of grid cells in the latitude direction, and the calculation formula is lon p =lon min +(p-1)·Δx,p=1,2,...,m lon lon p Indicates the minimum longitude of the p-th grid cell; m lon Represents the total number of grid cells in the latitude direction, and the calculation formula is Based on the latitude index lat for each grid cell q and longitude index lon p , calculate the center point coordinates of the grid cell (lat center ,lon center ) and generates a unique identification ID for each grid cell qp , where q and p correspond to the latitude and longitude indices of the grid cell, respectively.

4. The weather clock forecast system based on multi-source data fusion according to claim 3 is characterized in that: The method of mapping the location information of the device terminal to the grid unit includes: Calculate the grid number q of the device terminal at latitude device and the grid number p in longitude device , get the target grid unit G where the device terminal is located qp and its center point coordinates The calculation formula is:

5. The weather clock forecast system based on multi-source data fusion according to claim 1 is characterized in that: The detailed meteorological data F qp The calculation method is: Compare the grid resolution (Δx, Δy) of the target grid cell with the grid resolution (Δx w ,Δy w ), if the grid resolution of the meteorological model grid is coarser than the grid resolution of the grid cell, the initial meteorological data is updated by the difference algorithm, and the update formula is: Where, α k is the interpolation weight, and its calculation formula is: Indicates the numbers of the four meteorological model grid cells adjacent to the target grid cell.

6. The weather clock forecast system based on multi-source data fusion according to claim 1 is characterized in that: The training method of the weather correction model is: Collect historical meteorological data for different target grid cells The historical fine meteorological data of different target grid cells are processed and used as training input for the meteorological correction model; Initialize the learnable parameters of the meteorological correction model, including the weight coefficient matrix of each influencing factor in the local characteristic data; Define the loss function of the weather correction model Where, is the revised forecast value; is the corresponding true observation value; n is the number of samples; Define the regularization loss function of the weather correction model Where ω is the regularization coefficient, θ u is the learnable parameter in the weather correction model, u is the index label of the weight coefficient, is the regularization term of the sum of squares of all weight coefficients; Update the learnable parameters θ using gradient descent u , to reduce the loss function L reg The training of the historical detailed meteorological data is repeated until the meteorological correction model converges or reaches the preset number of iterations, that is, the training of the meteorological correction model is completed.

7. The weather clock forecast system based on multi-source data fusion according to claim 6 is characterized in that: The training method of the weather prediction model includes: Define the long short-term memory network model as the basic structure of the weather forecast model, which includes an input layer, a hidden layer, and an output layer; Introduce the parameter vector time step d, and transform the vector time step d and the meteorological correction data F′ qp and historical weather data Combined to obtain the fusion feature vector Where s is the number of vector time steps d; Collect historical meteorological data of the target grid cell within a fixed period of time in the past And the corresponding fusion feature vector is constructed and recorded as the historical fusion feature vector Fusion of history feature vector Expand by vector time step to form sequence data as the training input of the weather forecast model. Correspondingly, the sequence data is labeled according to the vector time step d, and the label is the historical weather result Initialize the network parameters of the weather forecast model; the network parameters include the weight matrix W from the input layer to the hidden layer x , the cyclic weight matrix W of the hidden layer h , the bias vector b of the hidden layer h , the weight matrix W from the hidden layer to the output layer o and the bias vector b of the output layer o ; Define the loss function of the weather prediction model Where S is the length of the sequence data; Y true,d The label for the dth vector time step, i.e. the historical weather results Y pred,d is the label of the dth vector time step, that is, the predicted weather forecast result The sequence data is passed to the input layer, and then passes through the hidden layer and output layer in sequence to output the predicted weather forecast results And calculate the corresponding weather forecast results The value of the loss function is obtained, and the error gradient is back-propagated along the output layer to the input layer; according to the error gradient, the network parameters are updated using the optimization algorithm to reduce the value of the loss function; the sequence data is trained repeatedly until the weather forecast model converges or reaches the preset number of iterations, which indicates that the weather forecast model training is completed.

8. The weather clock forecast system based on multi-source data fusion according to claim 7 is characterized in that: The input layer is used to receive the fused feature vector X d As input; the hidden layer consists of y state vectors, and the calculation formula for each state vector is: d =f(W h ·h d-1 +W x ·X d +b h ); where h d is the hidden state of the current vector time step d, f(...) is the activation function, W x is the weight matrix from the input layer to the hidden layer, W h is the cyclic weight matrix of the hidden layer, b h is the bias vector of the hidden layer; The output layer calculates the output of the weather forecast model based on the output of the hidden layer, that is, the weather forecast result The calculation formula is: Among them, η(...) is the activation function of the output layer, W o is the weight matrix from the hidden layer to the output layer, b o is the bias vector of the output layer.

Images