Meteorological element forecast product intelligent generation method and system

Abstract

The invention relates to the technical field of meteorological element forecasting, in particular to a meteorological element forecasting product intelligent generation method and system. The method comprises the following steps: acquiring meteorological element original data, and performing data assimilation on the meteorological element original data to obtain assimilated product data; combining the meteorological element original data and assimilation product data to obtain a forecast sensitive factor and site feature data; establishing a meteorological element product data set according to the meteorological element original data, the assimilation product data, the forecast sensitive factors and the site feature data; constructing a meteorological element forecast product intelligent generation model, and training and verifying the meteorological element forecast product intelligent generation model by using the meteorological element product data set; and generating a meteorological element forecasting product through the trained meteorological element forecasting product intelligent generation model. According to the method, the meteorological element original data is utilized, and the problem of efficient, accurate and intelligent generation of the meteorological element forecast product of the meteorological station is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological element forecasting, and in particular to a method and system for intelligently generating meteorological element forecast products. Background Art

[0002] In recent years, affected by global climate change, extreme weather events have occurred frequently. High temperatures, severe cold, droughts, heavy rains and other weather conditions have continued to affect people's production and life. The country has paid more and more attention to the field of meteorology, and people's requirements for the accuracy of weather forecasts have gradually increased. The forecast of meteorological elements such as temperature, air pressure, wind speed, rainfall, visibility, etc. is one of the focuses of national and public attention, and it is also a key research topic for meteorological personnel.

[0003] Traditional weather forecasts have problems such as low forecast accuracy, short forecast time, and single forecast elements. Therefore, how to efficiently and accurately provide meteorological element forecast products for meteorological stations is an issue that urgently needs to be addressed. Summary of the invention

[0004] In view of the shortcomings of existing methods and the needs of practical applications, in order to solve the problem of efficiently and accurately providing meteorological element forecast products of meteorological stations, the present invention provides a method for intelligently generating meteorological element forecast products, comprising the following steps: Obtain the original data of meteorological elements, perform data assimilation on the original data of meteorological elements, and obtain assimilated product data; combine the original data of meteorological elements and the assimilated product data to obtain forecast sensitivity factors and site characteristic data; establish a meteorological element product data set based on the original data of meteorological elements, the assimilated product data, the forecast sensitivity factors and the site characteristic data; construct a meteorological element forecast product intelligent generation model, use the meteorological element product data set to train and verify the meteorological element forecast product intelligent generation model; generate meteorological element forecast products through the trained meteorological element forecast product intelligent generation model. The present invention uses the original data of meteorological elements to perform data assimilation, and then obtains forecast sensitivity factors and site characteristic data, thereby constructing a meteorological element product data set to train the meteorological element forecast product intelligent generation model, and finally generates forecast products through the trained model, solves complex prediction problems according to the machine learning model, and effectively solves the problem of efficiently and accurately providing meteorological element forecast products for meteorological station sites.

[0005] Optionally, the method for intelligently generating meteorological element forecast products further includes performing missing value filling processing, outlier cleaning processing and time series decomposition processing on the original meteorological element data. The present invention eliminates the problem of noise data and inconsistency between data by preprocessing the original meteorological element data, which is conducive to improving the training speed and adaptability of the intelligent generation model of meteorological element forecast products.

[0006] Optionally, the outlier cleaning process includes climate extreme value check, consistency check between meteorological elements, time consistency check and continuity check; The time series decomposition process satisfies the following formula: ,in, represents a time series signal, Represents the trend component, represents the periodic component, The present invention decomposes the data into time series, which is beneficial to improve the utilization rate of highly non-stationary data and obtain high-precision prediction results.

[0007] Optionally, performing data assimilation on the original data of the meteorological elements to obtain assimilated product data comprises the following steps: A plurality of interpolation methods are used to interpolate different meteorological factors to meteorological stations; the effect after interpolation is evaluated, and the assimilated product data is obtained according to the evaluation result. The present invention interpolates different meteorological factors to meteorological stations by using a plurality of interpolation methods, and selects the optimal interpolation method for data interpolation, which is conducive to interpolating the assimilated product data to the corresponding mobile stations and corresponding to the actual data of the mobile stations.

[0008] Optionally, combining the original data of meteorological elements and the assimilated product data to obtain forecast sensitivity factors includes combining the original data of meteorological elements and the assimilated product data to obtain possible factors of each meteorological element through multiple factor analysis methods. The present invention obtains forecast sensitivity factors through multi-level analysis using multiple factor analysis methods, which is conducive to extracting possible factors of meteorological elements, thereby improving the accuracy of meteorological element forecast products.

[0009] Optionally, combining the meteorological element raw data and the assimilated product data to obtain site characteristic data comprises the following steps: Conventional features are extracted from the raw data of meteorological elements and the assimilated product data; main features are obtained from the conventional features using a feature selection method; and feature generation is performed based on the main features to obtain the site feature data. The present invention extracts main features from conventional features, and then performs feature generation to obtain site feature data, removes redundant and irrelevant variables, and integrates historical data features, which is conducive to improving the training efficiency and accuracy of the invention.

[0010] Optionally, establishing a meteorological element product data set according to the meteorological element original data, the assimilated product data, the forecast sensitivity factor and the site characteristic data comprises the following steps: According to the time series, the sliding window method is used to perform time slicing; based on the time characteristic variables, the meteorological element original data, the assimilated product data, the forecast sensitivity factor and the site characteristic data are combined to obtain the meteorological element change characteristics; the variables of the meteorological element change characteristics are standardized using the maximum and minimum value standardization method, and the time characteristic variables are sine-cosine encoded; the meteorological element product data set is established by combining the standardization processing results and the encoding results. The present invention can obtain the periodic characteristics of time by performing sine-cosine encoding on the time characteristic variables, which is further conducive to improving the training accuracy of the present invention.

[0011] Optionally, the sine-cosine encoding of the time characteristic variable satisfies the following formula:

[0012]

[0013]

[0014] in, Represents the monthly sine-cosine code, Indicates the month, Represents the daily sine and cosine encoding, Indicates 24-hour time. Represents the sine-cosine encoding of the circumference, Indicates a 7-day week.

[0015] Optionally, the meteorological element forecast product intelligent generation model is an improved Transformer model that introduces an autocorrelation mechanism. The meteorological element forecast product intelligent generation model provided by the present invention can be trained on small-scale data, and the model setting is simple, which is conducive to rapid training of the meteorological element forecast product intelligent generation model corresponding to each meteorological station.

[0016] In the second aspect, in order to efficiently execute the method for intelligently generating a meteorological element forecast product provided by the present invention, the present invention further provides a system for intelligently generating a meteorological element forecast product, including a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute the method for intelligently generating a meteorological element forecast product as described in the first aspect of the present invention. The system for intelligently generating a meteorological element forecast product of the present invention has a compact structure and stable performance, and can stably execute the method for intelligently generating a meteorological element forecast product provided by the present invention, further improving the overall applicability and practical application capability of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flow chart of a method for intelligently generating meteorological element forecast products provided by an embodiment of the present invention; Figure 2 A framework diagram of a meteorological element forecast product intelligent generation system provided by an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a meteorological element forecast product intelligent generation device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The specific embodiments of the present invention will be described in detail below. It should be noted that the embodiments described herein are only for illustration and are not intended to limit the present invention. In the following description, a large number of specific details are set forth in order to provide a thorough understanding of the present invention. However, it is obvious to those of ordinary skill in the art that these specific details do not need to be adopted to implement the present invention. In other examples, in order to avoid confusing the present invention, known circuits, software or methods are not specifically described.

[0019] Throughout the specification, references to "one embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in conjunction with the embodiment or example is included in at least one embodiment of the present invention. Therefore, the phrases "in one embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily all refer to the same embodiment or example. In addition, particular features, structures, or characteristics may be combined in one or more embodiments or examples in any suitable combination and / or subcombination. In addition, it should be understood by those of ordinary skill in the art that the figures provided herein are for illustrative purposes and that the figures are not necessarily drawn to scale.

[0020] See also Figure 1 In order to solve the problem of providing meteorological element forecast products of meteorological stations efficiently and accurately. The present invention provides a method for intelligently generating meteorological element forecast products, such as Figure 1 As shown, the method comprises the following steps: S1. Acquire original data of meteorological elements, perform data assimilation on the original data of meteorological elements, and obtain assimilated product data.

[0021] In the embodiment, the obtained raw data of meteorological elements usually comes from multiple channels, including but not limited to: Meteorological station observation: Meteorological stations are the main source of real-time and historical meteorological data. They regularly record data on meteorological elements such as temperature, humidity, air pressure, wind speed, wind direction, precipitation, etc. These data are highly accurate and reliable and are the basis for meteorological research and forecasting. Meteorological satellite observation: Meteorological satellites can cover a wider geographical area and provide meteorological data observed from space. These data include cloud maps, land surface temperature, sea surface temperature, precipitation distribution, etc., which are of great significance for the monitoring and forecasting of large-scale weather systems. Radar observation: Radar can detect the echo signals of meteorological targets such as precipitation particles and clouds, thereby providing real-time data such as precipitation intensity, precipitation type, and cloud movement direction. These data play an important role in the early warning and forecast of extreme weather such as short-term heavy rainfall and hail.

[0022] Furthermore, in order to improve the credibility of the data, the acquired raw data of meteorological elements also need to be preprocessed, including missing value filling, outlier cleaning and time series decomposition of the raw data of meteorological elements.

[0023] Missing data are usually filled by deletion and interpolation. Deletion refers to directly deleting the entire record with missing values; interpolation is to supplement the missing data with possible values, including mean interpolation, regression interpolation, maximum likelihood estimation, etc. Since the deletion method loses the sample volume and causes the incoherence of the time series, which in turn affects the output of the model, it is best not to use the deletion method to process missing data. In the case where only individual time data are missing in the original data set of continuous data, the mean interpolation method is used to interpolate the adjacent time data to obtain the current time data, thereby completing the supplement of the missing data.

[0024] The outlier cleaning process includes climate extreme value check, consistency check between meteorological elements, time consistency check and continuity check.

[0025] The climate extreme value check includes wind speed, temperature, altitude, ground pressure, sea level pressure and humidity. The check rule is: when the checked data exceeds the corresponding value range of the climate extreme value check of its element, the erroneous data will no longer participate in other subsequent check steps.

[0026] The consistency check between meteorological elements includes wind direction DD and wind speed FF. The following cases will be excluded: Wind direction DD is missing, wind speed Not missing a test; Wind direction DD is not missing, wind speed Missing test; Wind direction DD<0 or wind direction DD>360; Wind direction DD=0, wind speed ≠0; Wind direction DD≠0, wind speed =0; Wind direction DD is when the wind direction is uncertain (when the code is 620); Wind speed ≥ ; Also includes temperature And dew point TdTd check, the following cases will be eliminated: Temperature on land <Dew point TdTd or —TdTd>50°; Ship air temperature <TdTd-1℃or —TdTd>30℃.

[0027] The temporal consistency check is to examine the observation information or observation elements from the perspective of the temporal rate of change and identify undesirable sudden changes. Specifically, the current observation information is compared with the observation information at adjacent times. If their difference exceeds a given allowable range, the current observation information is considered suspicious. The inspection criteria for the temporal consistency check are related to the temporal resolution of the sample.

[0028] The time consistency check criteria are met: ,in, Indicates the maximum difference between the current observation information and the observation information at the adjacent time. represents the rate of change of the observed element over time, Indicates the time difference between two observations. In the embodiment, the maximum change rate of the observation element with altitude, latitude and time satisfies: temperature is 5°C / h, and air pressure is 3hPa / h.

[0029] The main function of the continuity check is to check whether the changes in observation parameters caused by instrument failure (such as icing, communication problems) are too small or unchanged. Compared with the time consistency check, the continuity check checks the rationality of the time change rate of observation elements from another perspective. The continuity check uses the data difference method of adjacent observation times. It requires that the maximum value of the data difference of all adjacent observation times within a given time range is greater than a given criterion. The check result is applied to all data within a given time window. The continuity check criteria are 0.1. If the maximum value of the observation data difference of all adjacent observation times within a given time range is less than a given criterion, it is considered that the continuity check has not passed.

[0030] The time series decomposition process satisfies the following formula: ,in, represents a time series signal, Represents the trend component, represents the periodic component, represents the random noise component.

[0031] It should be understood that the traditional time series prediction method is only applicable to the prediction of stationary time series. It is impossible to obtain high-precision prediction results for highly non-stationary data. The original data is decomposed into multiple groups of time series with different characteristics. The low-frequency trend information eliminates the interference of random data and reflects the inherent change trend of the original data. The high-frequency information is the random disturbance, which reflects the details of the change.

[0032] In the embodiment, the temperature curve of a certain station for a period of time is decomposed by wavelet transform, the wave kernel function used is db wavelet, the decomposition level is 2, and the time series composed of the historical temperature curve can be decomposed into multiple additive components for analysis. The trend term represents the overall trend of temperature change, and its value is not easy to change suddenly in a short time, so it can be used as the basic value of temperature prediction; the periodic term represents the fluctuation of temperature in recent years, and has a certain statistical change law.

[0033] Furthermore, the acquired data is also standardized. Data standardization is an important basic work of data preprocessing. It can be understood as a kind of scaling, that is, adjusting the proportion of values ​​within a specific boundary. Common standardization intervals include [0, 1] and [-1, 1].

[0034] The normalized interval range [0,1] satisfies: , the standardized interval range [-1,1] satisfies: . Where x represents the original data, x* represents the standardized data, and max and min are the maximum and minimum values ​​in the data set, respectively.

[0035] Normalization refers to processing new data to conform to the normal distribution. The commonly used method is Z-score standardization, which is a standard normal distribution with a mean of 0 and a standard deviation of 1: , where μ is the mean and σ is the standard deviation of a given time series.

[0036] It is understandable that the attributes of some samples are composed of character data, which takes up more computer resources during model training and is not convenient to read. Therefore, integer data is usually used to replace the meaning of character data, that is, "semantic conversion" is performed, and the most common one is the "0, 1" conversion of "yes, no".

[0037] In the embodiment, an adjacency matrix can also be constructed for the effective meteorological station information data after cleaning and screening. The meteorological observation station is modeled on the graph, that is, the observation station is used as a node on the graph, and then the weight between each two nodes is calculated using the distance between the stations. Finally, the graph structure is represented by an adjacency matrix. For meteorological observation stations with known longitude and latitude, the formula for calculating the distance between two stations is as follows:

[0038] in, Represents the actual distance from node i to node j; represents the radius of the Earth; , Respectively represent the latitude and longitude of node i; , Respectively represent the difference between the longitude and latitude of two nodes.

[0039] use represents the weighted adjacency matrix, then The calculation is shown in the formula:

[0040] in, Indicates the use distance The calculated edge weights; and Used to control the weighted adjacency matrix distribution and sparsity.

[0041] Furthermore, a high-resolution land surface data assimilation system is used to assimilate the raw data of meteorological elements to obtain assimilated product data.

[0042] It should be understood that the assimilated product data is based on the international time standard and is grid data, while the real-time data is based on Beijing time and is station data. Therefore, the assimilated product data time needs to be processed accordingly first. Beijing time is 8 hours later than the world time, so it can be processed uniformly according to Beijing time.

[0043] The step of performing data assimilation on the original data of meteorological elements to obtain assimilated product data comprises the following steps: A variety of interpolation methods are used to interpolate different meteorological factors to meteorological stations. Specifically, the various interpolation methods used include inverse distance interpolation, Kriging interpolation, and spline interpolation.

[0044] Evaluate the effect after interpolation, and obtain the assimilated product data according to the evaluation result. Specifically, the effect after interpolation can be evaluated in the following two ways. The first is to analyze the correlation coefficient between each meteorological factor after interpolation and the actual situation. The correlation coefficient is calculated by multiplication and difference. The deviation between different variables and their own average values ​​is multiplied to analyze the correlation between different variables. The smaller the absolute value of the correlation coefficient R is from 1, the closer the relationship between the variables is, and the smaller the difference from 0, the weaker the relationship between the variables is.

[0045] The second method is to calculate the root mean square error (RMSE), absolute error (MAE) and forecast accuracy (ACC) between the interpolated model forecast temperature and the actual station temperature. The forecast accuracy is the ratio of the number of stations with absolute errors within the specified range to the total number of stations. Since these three indicators are not necessarily proportional, it is necessary to make an overall evaluation of the error between the forecast and the actual. The forecast error assessment comprehensive index AMM satisfies: .

[0046] S2. Combining the original data of meteorological elements with the assimilated product data, obtaining forecast sensitivity factors and site characteristic data.

[0047] In the embodiment, in step S2, the original data of the meteorological elements and the assimilated product data are combined to obtain the forecast sensitivity factor, including combining the original data of the meteorological elements and the assimilated product data to obtain the possible factors of each meteorological element through multiple factor analysis methods.

[0048] Specifically, various factor analysis methods include conditional probability analysis, range analysis, correlation analysis, and EOF analysis.

[0049] Conditional probability analysis refers to the probability of event B occurring under the condition that event A occurs, given two events A and B. In weather forecasting, the probability of occurrence of meteorological element B can be analyzed under the condition that forecast factor A occurs, thereby screening out the sensitive factors of the forecast element.

[0050] The range analysis method includes two steps: calculation and judgment, that is, calculating the test index and , for The average value of The size of can determine the optimal level of factor j and the level combination of each factor, that is, the optimal combination.

[0051] is the range of the j-th factor, that is, the difference between the maximum and minimum values ​​of the average index value at each level of the j-th factor:

[0052] It reflects the change range of the test index when the level of the j-th column factor changes. The larger the factor is, the greater its influence on the test index is, and therefore the more important it is. By measuring the size of the factors, we can determine the importance of the factors and screen out the factors that are sensitive to element forecasting.

[0053] The correlation analysis method is to analyze the signs that are indeed related in the population. Its main body is to analyze the signs of causal relationship in the population. It is a process of describing the closeness of the relationship between objective things and expressing it with appropriate statistical indicators. In order to determine the relationship between related variables, some data should be collected first. These data should be in pairs, and then these points are described on the rectangular coordinate system. This set of points is called a "scatter plot". According to the scatter plot, when the independent variable takes a certain value, the dependent variable corresponds to a probability distribution. If the probability distribution of all independent variable values ​​is the same, it means that the dependent variable and the independent variable are not correlated; on the contrary, if the values ​​of the independent variable are different, the distribution of the dependent variable is also different, which means that there is a correlation between the two.

[0054] EOF analysis method gives the observation data of a climate variable field in matrix form: , m is the spatial point, i.e., the observation station or grid point, and n is the length of the time series, i.e., the number of observations; the natural orthogonal expansion of the meteorological field is to decompose X into two parts, the time function Z and the space function V, i.e. ,or, , the t-th observation value at the i-th grid point in the field can be regarded as a linear combination of p spatial functions vtk and time functions zki; in, , , is the jth typical field, which is just a function of space. The tth spatial field can be expressed as: ,or , the above formula shows that the tth field can be expressed as m spatial typical fields, which are linearly superimposed according to different weights. Each column of V represents a spatial typical field, because this field is determined by actual data.

[0055] In the embodiment, in step S2, combining the meteorological element raw data and the assimilated product data to obtain the site characteristic data includes the following steps: S21. Extracting conventional features from the meteorological element raw data and the assimilated product data.

[0056] In the embodiment, the extracted conventional features include the site number, air pressure, sea level air pressure, maximum air pressure, minimum air pressure, maximum wind speed, maximum wind speed, wind direction (angle) of maximum wind speed, 2-minute average wind direction (angle), 2-minute average wind speed, wind direction (angle) of maximum wind speed, temperature / air temperature, maximum temperature, minimum temperature, relative humidity, minimum relative humidity, precipitation, wind force, and perceived temperature.

[0057] S22. Obtaining main features from the conventional features using a feature selection method.

[0058] Specifically, variance selection method and / or Pearson correlation coefficient method are used to find out the main features.

[0059] The variance selection method belongs to the filtering feature selection method. Its principle is that if the variance of two features is small, it means that the values ​​of most samples of these two features are similar. In this case, one of the features is considered redundant and only one feature is selected. On the contrary, if the variance of two features is large, it means that the values ​​of many samples of the features are different. The specific method is to first calculate the variance between each two features, and then compare it with the threshold. If the calculated variance is greater than the threshold, the two features are retained. Otherwise, only one feature is retained.

[0060] The Pearson correlation coefficient method also belongs to the filtering feature selection method. Its principle is to reflect the statistical indicator of the closeness of the correlation between variables. More precisely, it measures the correlation between features and labels. Its calculation formula satisfies: ,in, is the correlation coefficient, when When , it indicates that the two variables are positively correlated. When , it indicates that the two variables are negatively correlated. When , it indicates that the two variables are completely correlated. When , it indicates that there is no correlation between the two variables; n is the number of samples, x and y represent the feature variable and label variable respectively.

[0061] S23: Generate features according to the main features to obtain the site feature data.

[0062] Feature generation refers to the creation of new features. Feature generation is a key step in feature engineering, because good features can significantly improve the effect of the model. In addition, business understanding can be integrated into the design of features to increase the interpretability of the model. Feature generation methods can be roughly divided into two categories. The first category is grouping and aggregation, which includes calculating the maximum value, minimum value, average value, count, etc. The second category is conversion methods, including addition, subtraction, multiplication, division, sorting and encoding of numerical types, truncation and statistical frequency of string types, date intervals, statistical months, cycles, holidays, etc. of date types.

[0063] Since meteorological data is inseparable in low dimensions, it is necessary to make a high-dimensional mapping of the data to make the data more separable, while obtaining higher-dimensional features and terms of mutual relationships; polynomial cross expansion can be considered as a transformation of existing features. By mapping the data to a higher-dimensional space, subsequent models can fit a wider range of data.

[0064] Polynomials are used to expand features. Assuming the original features are , the second expansion is In the same way, the existing features are used to multiply them n times to obtain new features, thus achieving n expansions. In view of the problem that the input feature dimension is small and there may be underfitting, the polynomial expansion method is used to analyze the input features composed of the latitude and longitude of the forecast data and the actual site respectively. At the same time, considering the problem of large number of features after expansion, the forecast data is expanded twice, three times and four times respectively to mine deeper information, and the correlation between the expanded features and the actual situation is analyzed. The required features are selected according to the correlation coefficient for subsequent modeling. The amount of data after high-order expansion is huge, which is easy to cause dimensionality explosion, and the information between the expanded features may be repeated.

[0065] Considering that meteorological data belongs to time series data, the date is converted to obtain new features. The purpose of this is to provide the model with better time features so that the model can more easily identify the relationship between the target value and time hidden in the data. The specific approach is as follows: group by site dimension, and then sort each group from small to large according to the time attribute; convert the time attribute of each sample into season, month, date (day of the year and day of the month), week, and hour; splice the generated features onto the original data set to complete the feature generation operation.

[0066] S3. Establish a meteorological element product data set based on the meteorological element original data, the assimilated product data, the forecast sensitivity factor and the site characteristic data.

[0067] Specifically, step S3, establishing a meteorological element product data set according to the meteorological element original data, the assimilated product data, the forecast sensitivity factor and the site characteristic data, comprises the following steps: According to the time series, time slicing is performed using the sliding window method; based on the time characteristic variables, the original data of the meteorological elements, the assimilated product data, the forecast sensitivity factor and the site characteristic data are combined to obtain the change characteristics of the meteorological elements; the variables of the change characteristics of the meteorological elements are standardized using the maximum and minimum value normalization method, and the time characteristic variables are sine-cosine encoded; combining the standardization results and the encoding results, the meteorological element product data set is established.

[0068] The data set is divided into three parts, namely, training set, validation set and test set in a ratio of 7:2:1; the data sets are all continuous time series, so they need to be constructed into time-continuous input-output sample pairs. Assume that the initial forecast time is t, the historical step length is T, and the forecast step length is , represents the historical feature vector, , including meteorological variables and time variables, represents the prediction factor, then the prediction modeling can be expressed as: ,in, is a nonlinear fitting function.

[0069] As meteorological elements have obvious periodic change characteristics, such as daily change and seasonal change, three time characteristic variables, month, hour, and corresponding day of the week, are introduced to establish the change characteristics of meteorological elements. Among them, the month is represented by 1-12 to represent the twelve months of the year, 0-23 to represent the 24 hours of the day, and 1-7 to represent the seven days of the week. Since the above variables cannot represent information such as the proximity of 23 hours to 0 hours, the time characteristics are processed by sine and cosine to obtain the periodic characteristics of time.

[0070] The sine-cosine encoding of the time characteristic variable satisfies the following formula:

[0071] in, Represents the monthly sine-cosine code, Indicates the month, Represents the daily sine and cosine encoding, Indicates 24-hour time. Represents the sine-cosine encoding of the circumference, Indicates a 7-day week.

[0072] The data in the dataset is sorted according to the time index. In order to adapt to the input of the model, the data needs to be converted into data that can be recognized by the neural network, that is, the time series prediction problem needs to be converted into a supervised learning problem; time series refers to a sequence of numbers sorted by time index, which can be understood as a list of ordered values, and the supervised learning problem contains a set of paired input sets and output sets. The algorithm can learn how to predict the output set based on the input set; the key to converting time series data into a supervised learning problem is to create a copy that can fill N value rows forward or backward. The overall translation of the copy column can create a lagged observation value column, which together with the predicted observation value column can form the format of the supervised learning dataset.

[0073] All variables in the time series can be automatically converted into a new format data set with supervised learning by translation, generating a sequence of multivariate input and output values. By setting different parameters, not only can single-step prediction be performed, that is, using the single variable observation value at the lag t-1 moment as the input variable to predict the value of the variable at the current time t, but also multi-step sequence prediction and multivariate prediction can be performed; multi-step sequence prediction refers to using the single variable observation value input sequence of the past m moments to construct a prediction problem to predict the observation value of the next n moments, and multivariate prediction refers to observing multiple different variables, and multi-step sequence prediction can be performed at the same time.

[0074] S4. Construct a meteorological element forecast product intelligent generation model, and use the meteorological element product dataset to train and verify the meteorological element forecast product intelligent generation model.

[0075] In the embodiment, the meteorological element forecast product intelligent generation model is an improved Transformer model that introduces an autocorrelation mechanism.

[0076] The Transformer model is a sequence-to-sequence model consisting of an encoder and a decoder, both of which are stacks of identical blocks. Each encoder block is mainly composed of a multi-head self-attention module and a position feedforward network. In order to build a deeper model, residual connections are used around each module, followed by a layer normalization module. Compared to the encoder block, the decoder block inserts an additional cross-attention module between the multi-head self-attention module and the position feedforward network. In addition, the multi-head self-attention module in the decoder is used to prevent each position from affecting subsequent positions.

[0077] The encoder is composed of 6 identical stacked block structures, each of which is further divided into two sub-layers: a multi-head self-attention mechanism and a feedforward network fully connected layer. After each sub-layer in the block, a normalization layer is added.

[0078] The decoder is also composed of 6 identical stacked block structures. Each block structure adds a third sublayer on the basis of the two sublayers of the encoder, that is, a multi-head self-attention sublayer is added; similar to the encoder, a normalization layer is added after each sublayer in the block; on the decoder side, the self-attention sublayer in the decoder stack is modified to prevent the position encoding from being correlated with the subsequent position encoding. Through this masking, it is ensured that the prediction of position i can only depend on the known output of positions less than i.

[0079] The attention function maps the triple Q(Query), K(Key), V(Value) to an output, where the triple Q(Query), K(Key), V(Value) and the output are all vectors, and the output is the weighted sum of V(Value), where the weight is the value calculated from the corresponding combination of Q(Query) and K(Key). The Transformer model uses a multi-head attention mechanism in three places: in the encoder-decoder attention layer, the Q value comes from the previous decoder layer, and the K value and V value come from the output of the encoder, so that each position information of the decoder is related to the position information of the input sequence. This architecture imitates the encoder-decoder attention mechanism of the sequence-to-sequence model. The encoder includes a self-attention layer. In the self-attention layer, the Q value, K value, and V value all come from the output of the previous encoder layer, and the position information in the encoder participates in the position encoding of the previous layer; similarly, the self-attention mechanism in the decoder allows the position information in the decoder to participate in the decoding of all position information.

[0080] In each block of the Transformer encoder-decoder architecture, in addition to the multi-head attention mechanism, it also contains a fully connected feedforward network. The fully connected feedforward network layer contains two linear transformations of the ReLU activation function. Similar to other sequence-to-sequence models, the Transformer model uses word embedding technology to convert input tags and output tags into vectors of dimension , and uses trainable linear transformations and Softmax functions to transform the output of the decoder into the probability of the next tag to be predicted. In the Transformer model, the weight matrix is ​​shared between the two embedding layers and the Softmax layer.

[0081] The autocorrelation mechanism is mainly composed of three parts: sequence decomposition module, autocorrelation mechanism, encoder and decoder; the encoding and decoding architecture is composed of three modules: sequence correlation, sequence decomposition, and position feedforward network in each layer, and a residual structure is used at the same time.

[0082] Based on this, the series can be decomposed into two parts: trend term and seasonal term. The understanding of these two parts is that one can reflect short-term fluctuations and the other can reflect long-term seasonality.

[0083] The input of the encoder is a sequence of known historical I time points, and the input of the decoder includes seasonal terms and periodic terms, and the subsequent O-length sequence is padded with 0 and the mean respectively; d is the number of time series, and O is the future time length; the encoder focuses on seasonal modeling; the output is past seasonal information, which will be used as mutual information to help the decoder adjust the prediction results. The decoder consists of two parts: the cumulative structure of the periodic trend component and the superimposed autocorrelation mechanism of the seasonal component; each decoding layer contains an internal autocorrelation block and an encoder-decoder autocorrelation block, which respectively refine the prediction and utilize past seasonal information.

[0084] Autocorrelation discovers period-based correlation by calculating the autocorrelation of the sequence, and aggregates similar subsequences through time-delay aggregation. The same phase position between periods naturally provides similar sub-processes. In order to achieve sequence-level connection, similar subsequence information also needs to be aggregated.

[0085] Finally, the meteorological element product dataset is used to train and verify the meteorological element forecast product intelligent generation model.

[0086] Specifically, the input information of the time series prediction model is decomposed seasonally (periodicly) and trend-wise, and the original input is decomposed into seasonal trend as input parameters. A multi-attention mechanism is used to split the original channel into three and then piece them together to get QKV, where K represents expansion from S to L, Q is convolution, and QK is transformed into a Fourier transform. Q and K are separated, and the results of the Fourier transform are multiplied and then an inverse Fourier transform is performed to get the original value, which is then pieced together to get the multi-headed attention value, that is, the data after attention calculation.

[0087] By training and outputting the weight parameters of the intelligent generation model of meteorological element forecast products of the corresponding station, the intelligent generation model of meteorological element forecast products of the corresponding meteorological station can be obtained.

[0088] S5. Generate meteorological element forecast products through the trained meteorological element forecast product intelligent generation model.

[0089] In an embodiment, after the training of the intelligent generation model of the meteorological element forecast product is completed, the network prediction output is used to calculate new input data that is different from the training data. The input data P of the network used for prediction has the same structure as the training, and the network calculates output data Y with the same structure as the training output data Y. Using any number of data samples, the same number of output samples are obtained. A neural network with a delayed cyclic connection between the output layer and the input layer and a delayed input, the network requires a calculation time delay output of the time step d. When the network input data P is used, for the input data at time zero, there is no previous input and output, so all unknown previous inputs and outputs are set to zero. At this time, the output of the hidden layer does not have the value of the previous time. If the network has an internal loop, the previous output of the hidden layer is also set to zero.

[0090] From the perspective of time series, prediction is a time series modeling problem. It takes the past time series as the model input, and then outputs a fixed future time series through the trained model. The moment values ​​of the known observed time series and the sequence values ​​of related variables, together with the past target observation values, constitute the training data. The deep learning prediction model is input to train it, and then the future moment values, i.e., the corresponding sequence values ​​of related variables, are input. Single-step prediction is performed through the prediction model, and the prediction interval remains consistent with the observed input data. Among them, the known input refers to the historical observation data at a fixed interval, and the step size of the interval is determined by the degree of correlation of the input sequence itself.

[0091] Furthermore, multivariate time series means the observation sequence of multiple variables obtained by observing the same process at different measurement sites or measurement methods at the same time. Correspondingly, for multivariate time series data, the sequence values ​​of p observed variables at the past m moments are used to predict the sequence values ​​of q observed variables at the next n moments. This is called multivariate multi-step prediction. Unlike single-variable single-step prediction, multivariate multi-step prediction is more difficult because the prediction model needs to deal with the complex correlation between variables when making predictions. Multi-step prediction can be recursively calculated by iterating the single-step prediction model, or it can be directly calculated using a specific set model for each time step.

[0092] Determine the quantitative analysis indicators. The object of quantitative analysis is the specific input variables and their impact on the model prediction results. Therefore, the evaluation indicators of the model prediction results, such as prediction error, model training time, model training loss, etc., can be considered as indicators for sensitivity quantitative analysis; calculate the target value of the model prediction result. Generally, the value of the model prediction result evaluation indicator under the normal parameter setting state is used as the target value. When conducting sensitivity quantitative analysis, it is not necessary to consider and calculate all uncertain factors. Instead, several factors with large possible changes and a greater impact on the target value of the model prediction result should be selected according to the specific situation of the prediction model.

[0093] See also Figure 2 In an embodiment, in order to efficiently execute the meteorological element forecast product intelligent generation method provided by the present invention, the present invention further provides a meteorological element forecast product intelligent generation system, including: an input device, an output device, a processor, and a memory, wherein the input device, the output device, the processor, and the memory are interconnected, and the memory contains program instructions, and the program instructions are used for the steps of the meteorological element forecast product intelligent generation method. The meteorological element forecast product intelligent generation system of the present invention has a compact structure and stable performance, and can stably execute the meteorological element forecast product intelligent generation method of the present invention, further improving the overall applicability and practical application ability of the present invention.

[0094] In an embodiment, the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. An input device may be used to obtain data information. An output device may be used to output the results obtained by storing program instructions contained in a computer program in a memory provided by the present invention. The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory.

[0095] In yet another alternative embodiment, see Figure 3In order to efficiently execute the method for intelligently generating meteorological element forecast products provided by the present invention, this embodiment also provides a device for intelligently generating meteorological element forecast products, such as Figure 3 As shown, including: The memory 10 is used to store computer programs; the processor 20 is used to execute the computer programs to implement the above-mentioned intelligent generation method of meteorological element forecast products. The memory 10, the processor 20, the communication interface 31 and the communication bus 32. The memory 10, the processor 20 and the communication interface 31 all communicate with each other through the communication bus 32.

[0096] In an embodiment, the memory 10 is used to store one or more program instructions. The memory 10 may store program instructions for implementing the following functions: Acquire raw data of meteorological elements, perform data assimilation on the raw data of meteorological elements, and obtain assimilated product data; combine the raw data of meteorological elements and the assimilated product data to obtain forecast sensitivity factors and site characteristic data; establish a meteorological element product data set based on the raw data of meteorological elements, the assimilated product data, the forecast sensitivity factors and the site characteristic data; construct a meteorological element forecast product intelligent generation model, and use the meteorological element product data set to train and verify the meteorological element forecast product intelligent generation model; generate meteorological element forecast products through the trained meteorological element forecast product intelligent generation model.

[0097] In one possible implementation, the memory 10 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, and at least one application required for a function, etc.; the data storage area may store data created during use. In addition, the memory 10 may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include NVRAM. The memory stores an operating system and operating instructions, executable modules or data structures, or a subset thereof, or an extended set thereof, wherein the operating instructions may include various operating instructions for implementing various operations. The operating system may include various system programs for implementing various basic tasks and processing hardware-based tasks.

[0098] The processor 20 may be a central processing unit (CPU), an application specific integrated circuit, a digital signal processor, a field programmable gate array or other programmable logic device, and the processor 20 may be a microprocessor or any conventional processor, etc. The processor 20 may call a program stored in the memory 10. The communication interface 31 may be an interface of a communication module, and is used to connect with other devices or systems.

[0099] Of course, it should be noted that Figure 3 The structure shown does not constitute a limitation on the intelligent generation device for meteorological element forecast products in this embodiment. In actual applications, the intelligent generation device for meteorological element forecast products may include Figure 3 More or fewer components than shown, or combinations of certain components.

[0100] The embodiment also provides a storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the above-mentioned method for intelligently generating meteorological element forecast products are implemented.

[0101] The storage medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0102] In summary, the present invention uses the original data of meteorological elements for data assimilation, and then obtains the forecast sensitivity factor and site characteristic data, thereby constructing a meteorological element product data set to train the meteorological element forecast product intelligent generation model, and finally generates a forecast product through the trained model, solves complex prediction problems according to the machine learning model, and effectively solves the problem of efficiently and accurately providing meteorological element forecast products for meteorological stations. Therefore, the present invention effectively overcomes the various shortcomings of the prior art and has a high industrial utilization value.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope recorded in the present invention.

Claims

1. A method for intelligently generating meteorological element forecast products, characterized in that: The method for intelligently generating meteorological element forecast products comprises the following steps: Acquire raw data of meteorological elements, perform data assimilation on the raw data of meteorological elements, and obtain assimilated product data; Combining the raw data of meteorological elements with the assimilated product data to obtain forecast sensitivity factors and site characteristic data; Establishing a meteorological element product data set according to the meteorological element original data, the assimilated product data, the forecast sensitivity factor and the site characteristic data; Constructing a meteorological element forecast product intelligent generation model, and using the meteorological element product data set to train and verify the meteorological element forecast product intelligent generation model; Generate meteorological element forecast products through the trained meteorological element forecast product intelligent generation model.

2. The method for intelligently generating meteorological element forecast products according to claim 1, characterized in that: It also includes missing value filling processing, abnormal value cleaning processing and time series decomposition processing on the original data of the meteorological elements.

3. The method for intelligently generating meteorological element forecast products according to claim 2, characterized in that: The outlier cleaning process includes climate extreme value check, consistency check between meteorological elements, time consistency check and continuity check; The time series decomposition process satisfies the following formula: ， in, represents a time series signal, Represents the trend component, represents the periodic component, represents the random noise component.

4. The method for intelligently generating meteorological element forecast products according to claim 1, characterized in that: The step of performing data assimilation on the original data of meteorological elements to obtain assimilated product data comprises the following steps: Various interpolation methods are used to interpolate different meteorological factors to meteorological stations; The interpolation effect is evaluated, and the assimilated product data is obtained according to the evaluation result.

5. The method for intelligently generating meteorological element forecast products according to claim 1, characterized in that: Combining the original data of the meteorological elements with the assimilated product data to obtain a forecast sensitivity factor includes combining the original data of the meteorological elements with the assimilated product data to obtain possible factors of each meteorological element through multiple factor analysis methods.

6. The method for intelligently generating meteorological element forecast products according to claim 1, characterized in that: Combining the meteorological element raw data and the assimilated product data to obtain site characteristic data includes the following steps: Extracting conventional features from the meteorological element raw data and the assimilated product data; Using a feature selection method, obtaining main features from the conventional features; Feature generation is performed according to the main features to obtain the site feature data.

7. The method for intelligently generating meteorological element forecast products according to claim 1, characterized in that: The step of establishing a meteorological element product data set according to the meteorological element raw data, the assimilated product data, the forecast sensitivity factor and the site characteristic data comprises the following steps: According to the time series, time slicing is performed using the sliding window method; Based on the time characteristic variables, the meteorological element raw data, the assimilated product data, the forecast sensitivity factor and the site characteristic data are combined to obtain the meteorological element change characteristics; The variables of the meteorological element change characteristics are standardized by using the maximum and minimum value standardization method, and the time characteristic variables are sine-cosine encoded; The meteorological element product data set is established by combining the standardization processing results and the coding results.

8. The method for intelligently generating meteorological element forecast products according to claim 7, characterized in that: The sine-cosine encoding of the time characteristic variable satisfies the following formula: ， ， ,in, Represents the monthly sine-cosine code, Indicates the month, Represents the daily sine and cosine encoding, Indicates 24-hour time. Represents the sine-cosine encoding of the circumference, Indicates a 7-day week.

9. The method for intelligently generating meteorological element forecast products according to claim 1, characterized in that: The intelligent generation model of meteorological element forecast products is an improved Transformer model that introduces an autocorrelation mechanism.

10. A meteorological element forecast product intelligent generation system, characterized in that: The meteorological element forecast product intelligent generation system includes: an input device, an output device, a processor, and a memory. The input device, output device, processor, and memory are interconnected. The memory includes program instructions, and the program instructions are used to execute the meteorological element forecast product intelligent generation method described in any one of claims 1-9.

Images