CN114385611A - A precipitation prediction method and system based on artificial intelligence algorithm and knowledge graph - Google Patents

Abstract

The invention discloses a precipitation prediction method and system based on an artificial intelligence algorithm and a knowledge graph, belonging to the technical field of weather forecast and comprising the following steps: constructing a multi-mode data container, inputting different types of meteorological data into the multi-mode data container according to the structural characteristics of the multi-mode data container to obtain multi-mode data, and performing space-time alignment, data cleaning and preprocessing and data sequence segmentation on the multi-mode data; constructing a knowledge graph about precipitation prediction; and (3) creating a multi-modal precipitation prediction model based on an artificial intelligence algorithm, and intelligently correcting the predicted precipitation data. The method not only makes full use of different types of multi-modal meteorological data, improves the precision of precipitation prediction through a multi-modal precipitation prediction method, but also constructs a precipitation prediction knowledge map, and can intelligently correct the result of model prediction so as to reduce the uncertainty of an artificial intelligence algorithm and improve the accuracy and reliability of model prediction.

Description

A precipitation prediction method and system based on artificial intelligence algorithm and knowledge graph

technical field

The invention belongs to the technical field of weather forecasting, and in particular relates to a precipitation prediction method and system based on an artificial intelligence algorithm and a knowledge map.

Background technique

In recent years, precipitation has increased significantly around the world. Precipitation prediction plays an important role in urban traffic, agricultural systems, and even in water resources planning and flood control early warning. At present, the use of meteorological big data to realize the prediction of impending precipitation is a hot spot and a difficulty in disaster prevention and mitigation in China, and it has important value and social significance for impending precipitation prediction with high temporal and spatial resolution.

From the perspective of the entire development of precipitation forecasting, at the beginning, meteorologists mainly used the method of weather maps. This method has a long history and mature theory, but it is relatively subjective and is a qualitative analysis. With the continuous development of meteorology and computer technology, numerical prediction methods have gradually developed, which are mainly methods for predicting precipitation changes by solving atmospheric dynamic equations. This method is also a method used in practice in recent years, but this method has certain limitations in its application. First, the prediction accuracy of this method in the short term (especially within 1-2 hours) is often low. The main reason is that the results of short-term predictions are more dependent on the initial value of the equation system of atmospheric dynamics, and the currently obtained atmospheric information is limited, so it is difficult to accurately estimate the initial value of this equation. Secondly, it requires the cooperation of supercomputers, which are relatively bulky and require expensive computing and maintenance costs. In addition, in terms of timeliness, numerical models often take a long time. From data acquisition to assimilation to forecasting, it often takes several hours to calculate the results. For certain fields, forecasting may have lost timeliness .

In recent years, with the exponential increase in the type and quantity of meteorological data, it has become an inevitable trend to combine artificial intelligence technology with precipitation prediction. Useful information can be mined from massive multimodal meteorological big data to discover climate characteristics. and atmospheric motion, and thus have a more accurate prediction of precipitation. But at the same time, artificial intelligence algorithms have certain uncertainties and may have certain errors in the prediction of some rare scenarios.

SUMMARY OF THE INVENTION

Aiming at the deficiencies of the prior art, the purpose of the present invention is to provide a precipitation prediction method and system based on an artificial intelligence algorithm and a knowledge graph, so as to solve the problems existing in the above background technology.

The present invention is realized in this way, a kind of precipitation prediction method based on artificial intelligence algorithm and knowledge graph, described method comprises the following steps:

Build a multi-modal data container, input different types of meteorological data into the multi-modal data container according to their own structural characteristics to obtain multi-modal data, perform spatio-temporal alignment, data cleaning and preprocessing, and data sequence segmentation for the multi-modal data;

Build a knowledge graph about precipitation forecasting;

Create a multi-modal precipitation prediction model based on artificial intelligence algorithms, specifically: building a multi-modal radar echo prediction model, integrating various historical meteorological data in time and space, and using a deep neural network model to predict future radar sequence diagrams; The radar precipitation intelligent transformation model, combined with geographic location data, uses the adversarial generation network to convert radar data and precipitation to each other, and the predicted radar sequence diagram can be converted into precipitation data; based on the constructed precipitation prediction knowledge map, the neural network The network algorithm conducts knowledge reasoning and intelligently corrects the predicted precipitation data;

Classified visualization of corrected precipitation data.

As a further scheme of the present invention: the step of performing spatiotemporal alignment on the multimodal data specifically includes:

Merge and align historical radar echo sequence data with a time error of less than 2 minutes and satellite cloud image data; perform bilinear interpolation on weather station data, and fuse the data obtained after bilinear interpolation with radar data and satellite data , to spatially match and align radar data, satellite data, weather station data and altitude data according to latitude and longitude;

The radar sequence data and geographic location big data, including latitude and longitude data and altitude data, are aligned in space and time with precipitation data.

As a further solution of the present invention: the step of constructing a knowledge map about precipitation prediction specifically includes:

Obtain precipitation-related data, including unstructured and semi-structured precipitation data from publicly available websites and research, and structured precipitation data from radar and satellites;

For unstructured precipitation data and semi-structured precipitation data, use the attention mechanism-based convolutional neural network to extract the entities, relationships and attributes of precipitation prediction to obtain the extracted precipitation information, the entities include latitude and longitude, altitude, time, temperature, Wind direction, humidity and precipitation, etc.;

Fusion of structured precipitation data and extracted precipitation information, the fusion process includes precipitation entity disambiguation and alignment;

The extracted precipitation information is further processed, and the machine learning correlation analysis algorithm is used to analyze the relationship between latitude and longitude, altitude, time, temperature, wind direction, humidity and precipitation, that is, given the specific values of meteorological factors, to obtain the area's latitude and longitude. maximum and minimum precipitation;

According to the extracted precipitation information, the precipitation prediction knowledge map is constructed, and the precipitation prediction knowledge map is stored in the map database.

As a further solution of the present invention: the step of constructing a multi-modal radar echo prediction model specifically includes:

Create a multimodal radar echo prediction model consisting of Data Integration module, CNN+LSTM module, DS+LSTM module, US+LSTM module, CNN+LSTM module and Data Generation module, in order to better predict the precipitation in the central area, The multimodal radar echo prediction model utilizes the meteorological data in the area of 512km*512km and the radar image in the prediction center area of 256km*256km; the multimodal radar echo prediction model is used to fuse historical multimodal meteorological big data , Historical multi-modal meteorological big data includes radar data, satellite cloud images, latitude and longitude data, altitude data, temperature, wind direction and humidity, using multiple neural network modules for training and learning to accurately analyze future high-resolution radar echo images The prediction of ; the loss function used by the multimodal radar echo prediction model is:

表示模型预测的雷达图数据，y_k表示真实的雷达图数据，ω_k和μ_k表示对不同强度的雷达图的加权值；in,

represents the radar map data predicted by the model, y _k represents the real radar map data, ω _k and μ _k represent the weighted values of radar maps of different intensities;

The evaluation function is used to evaluate the multimodal radar echo prediction model. The evaluation function is:

The value of the Value evaluation result is between -1 and 1. The larger the Value evaluation result value, the better the effect. Among them, M is the length of the predicted radar map sequence, and Value _i is the score of a single radar map, namely:

Among them, L is the predicted category, N is the total number of samples, ω _j is the weight of the j-th category, and p(R _i , T _j ) represents the total number of pixels with the real category j in the pixels of the predicted category i. , p(R _i ) represents the total number of pixels predicted as category i, and p(T _j ) represents the total number of pixels with true category j.

As a further scheme of the present invention: the step of constructing a radar precipitation intelligent transformation model specifically includes: constructing a transformation model comprising four modules of an R2P generator, a P2R generator, a P discriminator and an R discriminator, combined with latitude and longitude and altitude geography For location data, using historical radar data and precipitation data, the conversion model is trained and optimized on the gpu cluster to obtain the radar precipitation intelligent conversion model; the loss function used by the radar precipitation intelligent conversion model is:

Loss=Loss1+Loss2+Loss3

in:

Another object of the present invention is to provide a precipitation prediction system based on an artificial intelligence algorithm and a knowledge graph, the system comprising:

The multi-model data container is used to input different types of meteorological data into the multi-modal data container according to its own structural characteristics to obtain multi-modal data, and perform spatio-temporal alignment, data cleaning and preprocessing, and data sequence segmentation for the multi-modal data;

The precipitation prediction knowledge map building module is used to build a knowledge map about precipitation prediction;

A multimodal precipitation prediction and correction module based on artificial intelligence algorithms, used to create a multimodal precipitation prediction model based on artificial intelligence algorithms, including: a radar echo prediction model building unit for constructing multimodal radar echo predictions The model, which integrates various historical meteorological data in time and space, uses the deep neural network model to predict the future radar sequence diagram; the intelligent transformation model building unit is used to construct the intelligent transformation model of radar precipitation, combined with geographic location data, using the adversarial generation network, The radar data and precipitation can be converted into each other, and the predicted radar sequence diagram can be converted into precipitation data; the intelligent correction unit, based on the constructed precipitation prediction knowledge map, uses the neural network algorithm to perform knowledge inference, and analyzes the predicted precipitation data. make smart corrections; and

A regional precipitation visualization module for classifying and visualizing corrected precipitation data.

Beneficial effects: The present invention proposes a precipitation prediction method and system based on an artificial intelligence algorithm and a knowledge map, which not only makes full use of different types of multi-modal meteorological data, but also improves the accuracy of precipitation prediction through the multi-modal precipitation prediction method , solves the problems existing in the existing technology, and builds a precipitation prediction knowledge map, which can intelligently correct the results of the model prediction to reduce the uncertainty of the artificial intelligence algorithm and further improve the accuracy and reliability of the model prediction .

Description of drawings

In order to illustrate the technical solutions and embodiments of the present invention more clearly, the accompanying drawings used in the technical solutions and embodiments will be briefly introduced below.

Fig. 1 is a flow chart of a precipitation prediction method based on artificial intelligence algorithm and knowledge graph.

FIG. 2 is a flowchart of constructing a knowledge graph about precipitation prediction in an embodiment of the present invention.

FIG. 3 is a schematic structural diagram of a multi-modal radar echo prediction model in an embodiment of the present invention.

FIG. 4 is a schematic structural diagram of a radar precipitation intelligent conversion model in an embodiment of the present invention.

Figure 5 is an architecture diagram of a precipitation prediction system based on artificial intelligence algorithms and knowledge graphs.

Detailed ways

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

The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

As shown in FIG. 1 , an embodiment of the present invention provides a precipitation prediction method based on an artificial intelligence algorithm and a knowledge graph, and the method includes the following steps:

S100, constructing a multimodal data container, inputting different types of meteorological data into the multimodal data container according to their own structural characteristics to obtain multimodal data, performing spatiotemporal alignment, data cleaning and preprocessing, and data sequence on the multimodal data segmentation;

S200, build a knowledge map about precipitation prediction, use natural language processing technology to extract precipitation-related knowledge from public websites and research to obtain precipitation-related information, and fuse the precipitation-related information with the precipitation data collected by radar, satellite and weather stations , using deep learning algorithms to build a precipitation prediction knowledge map;

S300, creating a multimodal precipitation prediction model based on an artificial intelligence algorithm, specifically:

S301, construct a multi-modal radar echo prediction model, fuse various historical meteorological data in time and space, and use a deep neural network model to predict future radar sequence diagrams;

S302, build a radar precipitation intelligent transformation model, combine geographic location data, and use the confrontation generation network to mutually transform radar data and precipitation, so that the predicted radar sequence diagram can be converted into precipitation data;

S303, based on the constructed precipitation prediction knowledge map, use a neural network algorithm to perform knowledge inference, and intelligently correct the predicted precipitation data;

S400, classify and visualize the modified precipitation data, which is convenient for users to observe, compare and analyze. For more intuitive observation, the precipitation is classified into 8 categories, namely 0, (0, 2.5], (2.5, 5], (5,10], (10,25], (25,50], (50,100], (100,+∞) (unit: mm/h), then, based on the latitude and longitude data, the precipitation data is displayed on the map.

In this embodiment of the present invention, before acquiring multimodal meteorological data related to precipitation, it is necessary to acquire historical radar echo sequence data, satellite cloud image data, and weather station data (temperature, humidity, wind direction, etc.) within a period of time in the same area. etc.), geographic location big data (including latitude and longitude data, altitude data, etc.), meteorological data such as precipitation data, and save these data into multimodal data containers according to categories. Then, it is necessary to align the acquired multimodal data in space and time, so that the processed data can be used to build a multimodal radar echo prediction model and a radar precipitation intelligent transformation model; in addition, all data needs to be cleaned and preprocessed , introduce a smoothing factor to remove the historical data, then use the two-dimensional wavelet transform to remove the noise, and use bilinear interpolation to fill the default data. In addition, remove some abnormal sequences or too many missing values. sequence, to prevent these noises from interfering with the model, and finally, store all the preprocessed data into the data container for subsequent modeling.

In the embodiment of the present invention, the step of performing spatiotemporal alignment on the multimodal data specifically includes:

S101, merge and align the historical radar echo sequence data with a time error of less than 2 minutes and the satellite cloud image data; perform bilinear interpolation on the weather station data, and combine the data obtained after the bilinear interpolation with the radar data and satellite data For fusion, the radar data, satellite data, weather station data and altitude data are spatially matched and aligned according to latitude and longitude; it should be noted that in terms of time, the time interval of historical radar echo sequence data is generally 6 minutes, and the satellite cloud image data The time interval is generally 5 minutes. Based on the radar map time, the data with the closest time distance (with an error of less than 2 minutes) are integrated; secondly, the time interval of weather station data (temperature, humidity, wind direction, etc.) is generally 1 In the space, according to the latitude and longitude, the radar data, satellite data, weather station data (temperature, humidity, wind direction, etc.) and altitude data can be spatially fused by bilinear interpolation. After alignment, the fused data can be used to build a multimodal radar echo prediction model.

S102 , align the radar sequence data and the geographic location big data with the precipitation data in space and time, where the geographic location big data includes latitude and longitude data and altitude data. Similarly, the time interval of currently obtained precipitation data is generally 1 hour. Based on this data, the radar sequence data and geographic location big data (including latitude and longitude data, altitude data, etc.) and precipitation data are aligned in time and space. The fused data can be used to build a radar precipitation intelligent transformation model.

As shown in FIG. 2, in the embodiment of the present invention, the step of constructing a knowledge graph about precipitation prediction specifically includes:

S201, obtain precipitation-related data, including unstructured precipitation data and semi-structured precipitation data obtained from public websites and research, and structured precipitation data obtained from radar and satellites;

S202, for the unstructured precipitation data and the semi-structured precipitation data, use the attention mechanism-based convolutional neural network to extract the entities, relationships and attributes of the precipitation prediction to obtain the extracted precipitation information, where the entities include latitude and longitude, altitude, time, temperature, wind direction, humidity and precipitation, etc.;

S203, fuse the structured precipitation data and the extracted precipitation information, and the fusion process includes precipitation entity disambiguation and alignment.

S204, further processing the extracted precipitation information, using a machine learning correlation analysis algorithm to analyze the relationship between longitude and latitude, altitude, time, temperature, wind direction, humidity and precipitation, that is, given the specific values of meteorological factors, to obtain the Maximum and minimum precipitation in the area;

S205 , constructing a precipitation prediction knowledge map according to the extracted precipitation information, and storing the precipitation prediction knowledge map in a map database.

In this embodiment of the present invention, unstructured data and semi-structured data are first obtained from public web pages and papers, through technologies such as crawler, and by parsing the pages and data, for example: "July 20, 2021, afternoon From 4:00 to 5:00, the accumulated precipitation in Zhengzhou Station reached 201.9 mm in one hour, setting a single-hour precipitation record for large, medium and small cities in the world. To extract knowledge for unstructured and semi-structured precipitation data, first extract entities, such as longitude, latitude, temperature, altitude, wind direction, humidity, precipitation and other entities related to precipitation. Due to the limited data obtained, the fine-tuning is mainly based on the pre-trained transfer model in natural language processing, and the fields with rich training expectations are used to help complete entity extraction. Second, extract the relationship and attributes of these entities. Here, the convolutional neural network based on the attention mechanism is mainly used to extract the relationship and attributes in units of combined sentence vectors. For example, the knowledge of an area extracted is, longitude: 109.809000, latitude: 40.657000, altitude: 1069m, time: June 1, 2021, temperature: 10 degrees Celsius, wind direction: level 3, precipitation: 20mm/h (note, This is just for a brief description of the extraction results, the data may be biased).

It is also necessary to use structured data obtained from products such as radar satellites, combined with previously extracted knowledge and the experience of meteorological experts, to integrate knowledge, including entity disambiguation and alignment, such as "precipitation", "rainfall", "rainfall" , "rainfall", "precipitation", etc. are unified as "precipitation". The biLSTM neural network is mainly used here, the entity alignment is regarded as the similarity comparison between strings, and finally the cosine similarity is calculated as the matching probability of the entity. Finally, the extracted knowledge is further processed. Use the machine learning correlation analysis algorithm to analyze the relationship between latitude and longitude, altitude, time, temperature, wind direction, humidity and other meteorological factors and precipitation, that is, given the specific values of meteorological factors, the maximum precipitation in the area can be obtained. and minimum precipitation. Construct the knowledge graph of precipitation prediction, use the inference engine based on the RDF model, construct the extracted data into an ontology in RDF format, and import it into the graph database to complete the construction and storage of the knowledge graph.

As shown in FIG. 3 , in the embodiment of the present invention, the step of constructing a multi-modal radar echo prediction model specifically includes:

S3011, create a multi-modal radar echo prediction model consisting of Data Integration module, CNN+LSTM module, DS+LSTM module, US+LSTM module, CNN+LSTM module and Data Generation module. Since the atmospheric clouds are constantly moving In order to better predict the precipitation in the central area, the model increases the receptive field of the observation data, that is, the meteorological data in the 512km*512km area and the radar image in the 256km*256km area of the prediction center are used; multimodal radar echo prediction The model is used to fuse historical multi-modal meteorological big data. Historical multi-modal meteorological big data includes radar data, satellite cloud images, latitude and longitude data, altitude data, temperature, wind direction and humidity. Multiple neural network modules are used for training and learning. It is used for more accurate prediction of future high-resolution radar echo images; the loss function used by the multi-modal radar echo prediction model is:

表示模型预测的雷达图数据，y_k表示真实的雷达图数据，ω_k和μ_k表示对不同强度的雷达图的加权值。in,

represents the radar map data predicted by the model, y _k represents the real radar map data, and ω _k and μ _k represent the weighted values of radar maps of different intensities.

S3012, use the evaluation function to evaluate the multi-modal radar echo prediction model, and the evaluation function is:

The value of the Value evaluation result is between -1 and 1. The larger the Value evaluation result value, the better the effect. Among them, M is the length of the predicted radar map sequence, and Value _i is the score of a single radar map, namely:

Among them, L is the predicted category, N is the total number of samples, ω _j is the weight of the j-th category, and p(R _i , T _j ) represents the total number of pixels with the real category j in the pixels of the predicted category i. , p(R _i ) represents the total number of pixels predicted as category i, p(T _j ) represents the total number of pixels with real category j;

In the embodiment of the present invention, the sequence data processed in the data container is divided according to the requirements, that is, the radar sequence data of the next 3 hours is predicted according to the meteorological data of the first 3 hours, and the data is divided according to 7:1: The ratio of 2 is divided into training set, validation set and test set. First, the input data is the multimodal meteorological big data of the first 3 hours, and the time interval is 6 minutes, namely -180min, -174min, ..., 0min. The multimodal meteorological big data includes radar data, satellite data, temperature, wind direction, Longitude, latitude, altitude, and input them into each module. The Data Integration module mainly uses data fusion technology in deep learning to integrate radar data, satellite cloud images, latitude and longitude data, altitude data, temperature, wind direction, and humidity. First, in order to prevent the subsequent gradient explosion or gradient disappearance of the model, the hyperbolic tangent function is used to normalize the data of each dimension:

Then, data fusion is performed using technologies such as Space-to-depth, Mean-pooling, Center-crop, Conv (convolution), and Max-pooling. The CNN+LSTM module is an encoder-decoder framework as a whole. Here, time and space are mainly integrated. The convolution calculation is introduced into the calculation of LSTM inside the network, and the spatiotemporal correlation of meteorological data is extracted. If the feature extraction is more sufficient and the sample size is sufficient, the CNN+LSTM module can stack several layers of the network, but due to the limited samples we obtain, in order to prevent the sample from overfitting, the number of layers M can be set to 1. The DS+LSTM module mainly uses down-sampling technology to fully extract the deep features of historical meteorological data of various dimensions. The US+LSTM module mainly uses the up-sampling technology to predict the future radar sequence data characteristics by using the extracted atmospheric motion characteristics. The CNN+LSTM module+DataGeneration module mainly generates the radar sequence data of the original area as accurately as possible according to the features of the extracted radar sequence data. Finally, through the Data Generation module, the radar sequence diagram of the next 3 hours is output, and the time interval is 30 minutes, that is, the forecast is once every half an hour.

In this embodiment of the present invention, the loss function used by the entire model is:

表示模型预测的雷达图数据，y_k表示真实的雷达图数据，ω_k和μ_k表示对不同强度的雷达图的加权值。利用前3个小时的多模态气象大数据，在gpu集群上，对这个改进后的模型进行训练，直到模型达到最优。利用如下评估函数，对上述最优模型进行评估：in,

represents the radar map data predicted by the model, y _k represents the real radar map data, and ω _k and μ _k represent the weighted values of radar maps of different intensities. Using the multi-modal meteorological big data of the first 3 hours, the improved model is trained on the GPU cluster until the model reaches the optimum. The optimal model above is evaluated using the following evaluation function:

The value evaluation result ranges between -1 and 1, and the larger the value, the better the effect.

As shown in Fig. 4, the present invention proposes an artificial intelligence transformation model, which can regard the radar map and the precipitation map as two image domains, based on the confrontation generation network, including R2P generator, P2R generator, P discriminator and R The four modules of the discriminator, combined with geographic location data such as latitude, longitude, and altitude, use historical radar data and precipitation data to train and optimize the model on the gpu cluster, and obtain a model that can be intelligently transformed between radar and precipitation.

In the embodiment of the present invention, first, obtain radar data, geographic location big data (including latitude and longitude data, altitude data, etc.) and precipitation data, set as {(X ₁ , X ₁ , Y ₁ ), (X ₂ , Z ₂ , Y ₂ ), ..., (X _n , Z _n , Y _n )}), the data is divided into training set, validation set and test set according to the ratio of 7:1:2. Then, a deep generative network is constructed, which can be regarded as composed of two adversarial networks as a whole, including four modules: R2P generator: mainly using the encoder-decoder framework, using real radar data, combined with geographic location data such as latitude, longitude and altitude , encode its features through the encoder encoder, and then use the decoder decoder to generate the corresponding precipitation data. P discriminator: It is mainly used to distinguish whether it is the fake precipitation data of the R2P generator or the real precipitation data. P2R generator: Similar in structure to the R2P generator, but its main function is to combine the real precipitation data with geographic location data such as latitude, longitude and altitude, encode its features through the encoder encoder, and then use the decoder decoder to generate the corresponding radar. data. R discriminator: It is mainly used to distinguish whether it is the fake radar data of the P2R generator or the real radar data. The model first uses the real radar data, combined with the latitude and longitude altitude data (X _i , Z _i ), _i =1, 2, ..., n, and input it to the R2P generator to generate the corresponding precipitation data Yi ', i=1,2,...,n. Then, use the P _{discriminator} to discriminate between the generated _Yi ' and the real Yi. In the same way, using real precipitation data, combined with latitude and longitude altitude data (Y _i , Z _i ), _i =1, 2, . =1,2,...,n. Then, use the R _{discriminator} to discriminate between the generated _Xi and the real Xi.

In this embodiment of the present invention, the loss function used by the entire model is:

Loss=Loss1+Loss2+Loss3

in:

The adversarial training optimization is carried out on the gpu cluster, and finally the R2P generator and the P2R generator obtain a set of optimal parameters, and the radar data and precipitation data can be flexibly transformed in combination with the latitude and longitude altitude. The resulting R2P generator is the near precipitation conversion model, which can convert radar data into precipitation data.

In the embodiment of the present invention, based on the constructed precipitation prediction knowledge map, a deep learning algorithm is used to perform knowledge inference, and the obtained precipitation data is intelligently corrected. Using the data and relationships obtained in the knowledge graph obtained by S200, the deep learning prediction model can be used to predict the maximum and minimum values of regional precipitation according to meteorological factors such as latitude and longitude, altitude, time, temperature, wind direction, and humidity. Furthermore, the precipitation data obtained by S32 is intelligently corrected according to the precipitation threshold predicted by the algorithm, so as to further improve the accuracy and reliability of precipitation prediction.

In the embodiment of the present invention, the visualization of precipitation data is for more intuitive observation, and the predicted and corrected precipitation is classified into 8 categories, namely 0, (0, 2.5], (2.5, 5] , (5,10], (10,25], (25,50], (50,100], (100,+∞) (unit: mm/h). Then, based on the latitude and longitude data, the precipitation data is displayed on the map. display for easy observation and analysis.

As shown in FIG. 5 , an embodiment of the present invention also discloses a precipitation prediction system based on an artificial intelligence algorithm and a knowledge graph, and the system includes:

The multi-model data container is used to input different types of meteorological data into the multi-modal data container according to its own structural characteristics to obtain multi-modal data, and perform spatio-temporal alignment, data cleaning and preprocessing, and data sequence segmentation for the multi-modal data;

The precipitation prediction knowledge graph building module is used to build a knowledge graph about precipitation prediction.

A multimodal precipitation prediction and correction module based on artificial intelligence algorithms, used to create a multimodal precipitation prediction model based on artificial intelligence algorithms, including: a radar echo prediction model building unit for constructing multimodal radar echo predictions The model, which integrates various historical meteorological data in time and space, uses the deep neural network model to predict the future radar sequence diagram; the intelligent transformation model building unit is used to construct the intelligent transformation model of radar precipitation, combined with geographic location data, using the adversarial generation network, The radar data and precipitation can be converted into each other, and the predicted radar sequence diagram can be converted into precipitation data; the intelligent correction unit, based on the constructed precipitation prediction knowledge map, uses the neural network algorithm to perform knowledge inference, and analyzes the predicted precipitation data. make smart corrections; and

A regional precipitation visualization module for classifying and visualizing corrected precipitation data.

In the process of using the system, it is first necessary to obtain real-time data of various dimensions such as radar, satellites, and weather stations in a certain area A, and then divide this area into n small areas of 500*500 km, and then use the intelligent precipitation prediction system. After each module, predict the future precipitation of these n small areas, and finally combine and visualize these small areas to observe and analyze the future precipitation.

The above only describes the preferred embodiments of the present invention in detail, and is not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention should be included in the protection of the present invention. within the range.

It should be understood that although the steps in the flowcharts of the embodiments of the present invention are sequentially displayed in accordance with the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, the execution of these steps is not strictly limited to the order, and these steps may be performed in other orders. Moreover, at least a part of the steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed and completed at the same time, but may be executed at different times. The order of execution is also not necessarily sequential, but may be performed alternately or alternately with other steps or at least a portion of sub-steps or stages of other steps.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium , when the program is executed, it may include the flow of the above-mentioned method embodiments. Wherein, any reference to memory, storage, database or other medium used in the various embodiments provided in this application may include non-volatile and/or volatile memory. Nonvolatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Road (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

Other embodiments of the present disclosure will readily occur to those skilled in the art upon consideration of the disclosure at the specification and examples. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or techniques in the technical field not disclosed by the present disclosure . The specification and examples are to be regarded as exemplary only, with the true scope and spirit of the disclosure being indicated by the claims.

Claims (6)

1. a precipitation prediction method based on artificial intelligence algorithm and knowledge map, is characterized in that, described method may further comprise the steps: Build a multi-modal data container, input different types of meteorological data into the multi-modal data container according to their own structural characteristics to obtain multi-modal data, perform spatio-temporal alignment, data cleaning and preprocessing, and data sequence segmentation for the multi-modal data; Build a knowledge graph about precipitation forecasting; Create a multi-modal precipitation prediction model based on artificial intelligence algorithms, specifically: building a multi-modal radar echo prediction model, integrating various historical meteorological data in time and space, and using a deep neural network model to predict future radar sequence diagrams; The radar precipitation intelligent transformation model, combined with geographic location data, uses the adversarial generation network to convert radar data and precipitation to each other, and the predicted radar sequence diagram can be converted into precipitation data; based on the constructed precipitation prediction knowledge map, the neural network The network algorithm conducts knowledge reasoning and intelligently corrects the predicted precipitation data; Classified visualization of corrected precipitation data. 2. a kind of precipitation prediction method based on artificial intelligence algorithm and knowledge map according to claim 1, is characterized in that, the described step of carrying out space-time alignment to multimodal data, specifically comprises: Merge and align historical radar echo sequence data with a time error of less than 2 minutes and satellite cloud image data; perform bilinear interpolation on weather station data, and fuse the data obtained after bilinear interpolation with radar data and satellite data , to spatially match and align radar data, satellite data, weather station data and altitude data according to latitude and longitude; The radar sequence data and geographic location big data, including latitude and longitude data and altitude data, are aligned in space and time with precipitation data. 3. a kind of precipitation prediction method based on artificial intelligence algorithm and knowledge map according to claim 1, is characterized in that, the described step of building the knowledge map of precipitation prediction, specifically comprises: Obtain precipitation-related data, including unstructured and semi-structured precipitation data from publicly available websites and research, and structured precipitation data from radar and satellites; For unstructured precipitation data and semi-structured precipitation data, use the attention mechanism-based convolutional neural network to extract the entities, relationships and attributes of precipitation prediction to obtain the extracted precipitation information, the entities include latitude and longitude, altitude, time, temperature, Wind direction, humidity and precipitation, etc.; Fusion of structured precipitation data and extracted precipitation information, the fusion process includes precipitation entity disambiguation and alignment; The extracted precipitation information is further processed, and the machine learning correlation analysis algorithm is used to analyze the relationship between latitude and longitude, altitude, time, temperature, wind direction, humidity and precipitation, that is, given the specific values of meteorological factors, to obtain the area's latitude and longitude. maximum and minimum precipitation; According to the extracted precipitation information, the precipitation prediction knowledge map is constructed, and the precipitation prediction knowledge map is stored in the map database. 4. a kind of precipitation prediction method based on artificial intelligence algorithm and knowledge map according to claim 1, is characterized in that, the described step of building multimodal radar echo prediction model, specifically comprises: Create a multimodal radar echo prediction model consisting of Data Integration module, CNN+LSTM module, DS+LSTM module, US+LSTM module, CNN+LSTM module and Data Generation module, in order to better predict the precipitation in the central area, The multimodal radar echo prediction model utilizes the meteorological data in the area of 512km*512km and the radar image in the prediction center area of 256km*256km; the multimodal radar echo prediction model is used to fuse historical multimodal meteorological big data , Historical multi-modal meteorological big data includes radar data, satellite cloud images, latitude and longitude data, altitude data, temperature, wind direction and humidity, using multiple neural network modules for training and learning to accurately analyze future high-resolution radar echo images The prediction of ; the loss function used by the multimodal radar echo prediction model is:

in,

represents the radar map data predicted by the model, y _k represents the real radar map data, ω _k and μ _k represent the weighted values of radar maps of different intensities; The evaluation function is used to evaluate the multimodal radar echo prediction model. The evaluation function is:

The value of the Value evaluation result is between -1 and 1. The larger the Value evaluation result value, the better the effect. Among them, M is the length of the predicted radar map sequence, and Value _i is the score of a single radar map, namely:

Among them, L is the predicted category, N is the total number of samples, ω _j is the weight of the j-th category, and p(R _i , T _j ) represents the total number of pixels with the real category j in the pixels of the predicted category i. , p(R _i ) represents the total number of pixels predicted as category i, and p(T _j ) represents the total number of pixels with true category j. 5. a kind of precipitation prediction method based on artificial intelligence algorithm and knowledge map according to claim 4, is characterized in that, the described step of building radar precipitation intelligent transformation model, specifically comprises: building comprises R2P generator, P2R generator, The transformation model of the four modules of the P discriminator and the R discriminator, combined with the latitude and longitude and altitude geographic location data, uses the historical radar data and precipitation data to train and optimize the transformation model on the gpu cluster, and obtain the radar precipitation intelligent transformation model; The loss function used by the radar precipitation intelligent transformation model is: Loss==Loss1+Loss2+Loss3 in:

6. A precipitation prediction system based on an artificial intelligence algorithm and a knowledge map, wherein the system comprises: The multi-model data container is used to input different types of meteorological data into the multi-modal data container according to its own structural characteristics to obtain multi-modal data, and perform spatio-temporal alignment, data cleaning and preprocessing, and data sequence segmentation for the multi-modal data; The precipitation prediction knowledge map building module is used to build a knowledge map about precipitation prediction; A multimodal precipitation prediction and correction module based on artificial intelligence algorithms, used to create a multimodal precipitation prediction model based on artificial intelligence algorithms, including: a radar echo prediction model building unit for constructing multimodal radar echo predictions The model, which integrates various historical meteorological data in time and space, uses the deep neural network model to predict the future radar sequence diagram; the intelligent transformation model building unit is used to construct the intelligent transformation model of radar precipitation, combined with geographic location data, using the adversarial generation network, The radar data and precipitation can be converted into each other, and the predicted radar sequence diagram can be converted into precipitation data; the intelligent correction unit, based on the constructed precipitation prediction knowledge map, uses the neural network algorithm to perform knowledge inference, and analyzes the predicted precipitation data. make smart corrections; and A regional precipitation visualization module for classifying and visualizing corrected precipitation data.

Images