Weather forecast learning system based on artificial intelligence algorithm

Abstract

The invention provides a weather forecast learning system based on an artificial intelligence algorithm, and the system comprises a data access collection module which collects a multi-source data set; the data fusion and processing module is used for carrying out data fusion and processing to obtain a multi-source fusion data set; the AI model design module is used for constructing a multi-model collaborative architecture and carrying out multi-model parallel training and optimization; the model fusion module is used for carrying out multi-model weighted fusion to obtain an AI model; the real-time prediction module is used for updating a prediction result according to the real-time data; the visualization and interpretability module is used for designing a visualization interface and carrying out interpretability verification; and the evaluation and iteration module is used for carrying out comprehensive evaluation and continuous improvement on a prediction result in combination with evaluation indexes. According to the method, accurate and reliable observation data can be obtained, massive meteorological data are efficiently processed by combining an artificial intelligence algorithm, trend analysis and prediction are automatically carried out, and a reliable prediction result is generated.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and particularly to a meteorological forecasting learning system based on artificial intelligence algorithms. Background Art

[0002] A meteorological forecasting learning system is a system for learning, analyzing, and predicting meteorological data. It collects and processes a large amount of meteorological data (such as temperature, humidity, wind speed, precipitation, etc.), and combines mathematical models, physical models, and data-driven algorithms to generate weather forecasts or climate trend predictions. This system can be used for short-term weather forecasting, medium- and long-term climate prediction, and early warning of extreme weather events.

[0003] Currently, meteorological forecasting systems mainly rely on: 1) Numerical Weather Prediction (NWP), which is based on physical models and mathematical equations (such as atmospheric dynamics equations, thermodynamics equations, etc.) to simulate the evolution process of the atmosphere; 2) Data assimilation technology, which combines observational data with numerical models to generate initial conditions and improve forecasting accuracy; 3) Statistical methods, which establish empirical models for prediction based on the statistical laws of historical meteorological data.

[0004] However, the existing technologies have the following problems: 1) In Numerical Weather Prediction (NWP), the model is very sensitive to initial conditions, and errors in observational data will lead to deviations in forecasting results. The spatial and temporal resolutions of the model are limited, making it difficult to accurately predict local weather; 2) In data assimilation technology, the computational efficiency is relatively low, and it is difficult to process large-scale data in real time; 3) In statistical methods, since they are based on historical laws, they cannot effectively handle the high nonlinearity and complexity in the meteorological system. Summary of the Invention

[0005] In order to overcome the deficiencies of the prior art, the object of the present invention is to provide a meteorological forecasting learning system based on artificial intelligence algorithms, which can obtain accurate and reliable observational data, and through the combination of artificial intelligence algorithms, efficiently process massive meteorological data, automatically perform trend analysis and prediction, generate reliable prediction results, and improve the overall intelligent level.

[0006] To achieve the above object, the present invention provides the following solution: A meteorological forecasting learning system based on artificial intelligence algorithms, comprising:

[0007] A data access and collection module, which is used to obtain various types of data from multiple sources, introduce external data sources to obtain an original data set, and then perform standardization and conversion operations on the original data set to obtain a multi-source data set;

[0008] A data fusion and processing module, which is used to utilize AI-enhanced data assimilation technology to fuse the multi-source data set and the numerical model output of traditional numerical weather forecasting to obtain an original fusion data set, and then perform data cleaning and quality improvement on the original fusion data set to obtain a multi-source fusion data set;

[0009] An AI model design module, which is used to combine multiple deep learning models to construct a multi-model collaborative architecture. Based on the multi-model collaborative architecture, use a distributed training framework for multi-model parallel training, and automatically search for the best combination of network structure and hyperparameters for model optimization to obtain an initial collaborative model;

[0010] A model fusion module, which is used to deploy a weighted fusion algorithm and a meta-model in the initial collaborative model to combine prediction results, and introduce statistical equations and Bayesian deep learning in the initial collaborative model to calculate prediction distributions to obtain an AI model;

[0011] A real-time prediction module, which is used to introduce an online training mechanism into the AI model and update the prediction results according to real-time data;

[0012] A visualization and interpretability module, which is used to design a visualization interface to output meteorological forecast results in a diversified manner, and introduce interpretability technology into the visualization interface to verify the rationality of prediction results;

[0013] An evaluation and iteration module, which is used to combine evaluation metrics to comprehensively evaluate and continuously improve prediction results for different weather elements and time scales;

[0014] Among them, the data access and acquisition module, the data fusion and processing module, the AI model design module, the model fusion module, the real-time prediction module, the visualization and interpretability module, and the evaluation and iteration module are connected to each other.

[0015] Optionally, the data access and acquisition module includes:

[0016] A data acquisition unit, which is used to combine a breakpoint resumption mechanism to obtain various types of data from multiple sources, and configure corresponding data interface information, data transmission protocols, and file storage formats for each data source; among them, various types of data include grid data, image data, and time series data;

[0017] A data introduction unit, which is used to introduce additional geographical information, marine meteorological information, transportation information, and energy consumption information into the acquired data to obtain an original data set, and perform hash verification and integrity check on the original data set;

[0018] A standardization and transformation unit for unifying the coordinate system and time steps of the original dataset, constructing the metadata standard of the original dataset, and projecting the image data onto a grid consistent with the grid data to obtain a multi-source dataset;

[0019] A storage and management unit for storing the multi-source dataset in a distributed file system and constructing a spatial index, a time index, and a unified API interface in the distributed file system.

[0020] Optionally, the data fusion and processing module includes:

[0021] A multi-source data fusion unit for combining a deep learning network with data assimilation technology, optimizing the assimilation weights to obtain an AI-enhanced data assimilation technology, using the AI-enhanced data assimilation technology to fuse the numerical weather prediction results with the multi-source dataset to obtain raw fusion data, and then using deep learning technology to interpolate and reconstruct the missing grid points in the raw fusion data to output a complete spatial distribution map;

[0022] An anomaly detection unit for identifying the outliers in the raw fusion data by combining statistical methods and an AI anomaly detection algorithm, and marking and deleting the identified outliers; wherein, the AI anomaly detection algorithm is implemented by constructing an isolation forest model of meteorological data;

[0023] A missing value filling unit for filling the raw fusion data containing missing values from the dimensions of spatial features and time features using a spatio-temporal convolutional network;

[0024] A quality improvement unit for generating extreme weather samples using a generative adversarial network and a variational autoencoder, oversampling the extreme weather samples to generate extreme weather synthetic samples that conform to the actual distribution, downsampling the raw fusion data to balance the data distribution, and then fusing the raw fusion data that has completed data cleaning and quality improvement with the extreme weather synthetic samples to obtain multi-source fusion data.

[0025] Optionally, the AI model design module includes:

[0026] A collaborative architecture unit for processing two-dimensional image data using a CNN model to identify cloud cluster structures and precipitation distribution changes, processing time series data using an LSTM / GRU model to capture medium- and short-term weather evolution trends, and using a Transformer model to capture long-range dependencies and predict long-time series climate or large-area weather forecasts to obtain the prediction results of multiple prediction models;

[0027] A multi-task learning unit for designing a multi-task learning model to extract the general features of the multi-source fusion data;

[0028] A model training unit, which is used to utilize a distributed training framework to allocate different prediction models in a multi-model collaborative architecture to different GPUs for parallel training;

[0029] A hyperparameter optimization unit, which is used to automatically search for optimal hyperparameters based on a defined hyperparameter space by using an optimization method and an optimization tool.

[0030] Optionally, the multi-task learning model includes an input layer for inputting the multi-source fusion data, a shared feature extraction layer for extracting general features of the multi-source fusion data, a task branch layer for designing an independent branch network for each prediction task, and an output layer for outputting the prediction results of each branch network.

[0031] Optionally, the expression of the multi-model collaborative framework is:

[0032] Y final = w cnn ·Y cnn + w LSSSTM ·Y LSTM + w Transformer ·Y Transformer

[0033] where Y cnn 、Y LSTM 、Y Transformer are the prediction results of the CNN, LSTM / GRU, and Transformer models respectively, and w cnn 、w LSSSTM 、w Transformer are all model weights.

[0034] Optionally, the model fusion module includes:

[0035] A result fusion unit, which is used to assign different weights to different prediction models and then combine the prediction results to obtain a comprehensive result;

[0036] An ensemble prediction unit, which is used to calculate the mean and variance of the comprehensive result by combining the Bayesian deep learning method to obtain a prediction distribution.

[0037] Optionally, the expression of the comprehensive result is:

[0038]

[0039] where Y fused is the comprehensive prediction result, Y i is the prediction result of the i-th model, w i is the weight of the i-th model, and N is the number of models participating in the fusion;

[0040] The expression of the predicted distribution is as follows:

[0041]

[0042]

[0043] where, μ pred is the mean of the predicted distribution, σ 2 pred is the variance of the predicted distribution, T is the number of sampling times, is the comprehensive prediction result of the t-th sampling.

[0044] Optionally, the visualization interface includes a spatial distribution map, a wind field animation, extreme weather warnings, and a time axis control for setting the prediction time step. The interpretability techniques include SHAP and Attention visualization. The evaluation metrics include mean square error, mean absolute error, correlation coefficient, and skill score.

[0045] The present invention discloses the following technical effects by providing a meteorological forecast learning system based on an artificial intelligence algorithm:

[0046] 1. Observation data is accurate and reliable: 1) By integrating multi-source data, multi-type data, and extended external data, the diversity of data is ensured; 2) Through AI-enhanced data assimilation technology, the performance of traditional methods is optimized, and missing grid points can also be reconstructed through deep learning technology to improve data integrity; 3) By performing standardization and conversion operations on the collected data, data source differences are eliminated, providing high-quality training data for model construction.

[0047] 2. High prediction accuracy: 1) By combining CNN models, LSTM / GRU models, Transformer models, and multi-task learning models to form a multi-model collaborative framework, it can accurately predict various meteorological elements and improve prediction accuracy by making full use of the characteristics of data types (such as image data, time series data) and prediction requirements (such as short-term weather forecasts, long-term climate predictions); 2) Through distributed training and hyperparameter optimization of the model, the operation efficiency of the model is improved; 3) Through techniques such as weighted fusion, the fusion effect of the model is enhanced; 4) By constructing extreme weather samples and specialized sub-models, the prediction accuracy and early warning ability for extreme weather events (typhoons, heavy rains, tornadoes, etc.) are effectively improved.

[0048] 3. Efficiently handle complex non - linear problems: By introducing AI algorithms (such as deep learning techniques), which are good at dealing with complex non - linear relationships, it can better capture the complex dynamics in the meteorological system, achieve automated data processing, trend analysis and prediction, reduce manual intervention, and improve the intelligence level of the system. And compared with the dependence of numerical models on supercomputers, AI models can run on ordinary computing devices, reducing hardware costs.

[0049] The technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings required for use in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0051] Figure 1 Schematic diagram of the system architecture provided for the embodiments of the present invention;

[0052] Figure 2 Schematic diagram of the process of multi - source data processing provided for the embodiments of the present invention;

[0053] Figure 3 Schematic diagram of the AI model construction process provided for the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0055] To make the above - mentioned objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] As Figure 1 shown, the present invention provides a meteorological forecast learning system based on artificial intelligence algorithms, including a data access and acquisition module, a data fusion and processing module, an AI model design module, a model fusion module, a real - time prediction module, a visualization and interpretability module, and an evaluation and iteration module that are interconnected.

[0057] 1. Data access and acquisition module

[0058] AsFigure 2 As shown, it is used to obtain various types of data from multiple sources, introduce external data sources to obtain an original data set, and then perform standardization and transformation operations on the original data set to obtain a multi-source data set.

[0059] The data access and collection module includes:

[0060] 1.1 Data collection unit

[0061] It is used to combine the breakpoint resumption mechanism to obtain various types of data from multiple sources, and configure corresponding data interface information (API address, data format, access rights, etc.), data transfer protocols (such as HTTP, FTP, MQTT) and file storage formats (such as NetCDF, CSV, HDF5) for each data source.

[0062] 1) Multiple sources include: regularly collecting multi-source meteorological data from various meteorological agencies (such as the National Meteorological Administration, satellite centers), weather stations, radars, and sources such as the ocean and Internet of Things sensors.

[0063] 2) Various types of data include:

[0064] Grid data, such as temperature, humidity, wind speed, etc., is generated through a gridded meteorological model output or observational data (such as temperature, humidity);

[0065] Image data, such as satellite cloud images and radar echoes, is obtained through remote sensing technology;

[0066] Time series data, such as observational values (temperature, humidity, precipitation, etc.) of weather stations, records the trend of weather elements changing over time.

[0067] 1.2 Data introduction unit

[0068] It is used to introduce additional geographical information, ocean meteorological information, transportation information, and energy consumption information into the acquired data to obtain an original data set, and perform hash verification and integrity check on the original data set.

[0069] Geographical information: such as terrain height and vegetation coverage rate, for local weather feature analysis;

[0070] Ocean meteorology: such as ocean temperature and ocean current, for ocean weather and typhoon path prediction;

[0071] Transportation and energy consumption: such as traffic flow and energy demand, supporting multi-field collaborative prediction;

[0072] Data packet verification: During the collection process, perform hash verification (such as MD5 or SHA256) on each data packet to ensure that the data has not been tampered with. If the verification fails, automatically trigger the retransmission logic;

[0073] Data integrity check: Check the timestamp, number of records, field integrity, etc. of the data file. For missing time periods or fields, log the information and send an alert.

[0074] 1.3 Standardization and Conversion Unit

[0075] Used to unify the coordinate system and time step of the original dataset, construct the metadata standard of the original dataset (such as unified data source, spatial resolution, time resolution, variable name, etc.), and project the image data onto a grid consistent with the grid data to obtain a multi-source dataset.

[0076] 1) Coordinate system unification: Project multi-source data onto a unified geographic coordinate system (such as WGS84) to ensure the consistent spatial distribution of different data sources. Use GIS tools (such as GDAL, ArcGIS) or Python libraries (such as `pyproj`) to complete the projection conversion.

[0077] 2) Time step unification: Convert the timestamps of all data to standard UTC time to eliminate differences in different time standards (such as local time, UTC time). For data with different time resolutions, perform interpolation or aggregation processing; for high-frequency data (such as minute-level), perform downsampling, and for low-frequency data (such as daily-level), perform time interpolation (such as linear interpolation or cubic spline interpolation).

[0078] 3) Projection conversion: Perform grid processing on satellite observation data and radar image data, project them onto a grid consistent with the grid data, and use bilinear interpolation or nearest neighbor interpolation algorithms to complete data resampling.

[0079] 1.4 Storage and Management Unit

[0080] Used to store the multi-source dataset in a distributed file system (such as HDFS, which stores large-scale meteorological data and supports parallel read and write operations), and construct a spatial index, time index (such as R-tree, timestamp index) and unified API interface in the distributed file system to support data retrieval by time, location, and variable type.

[0081] 2. Data Fusion and Processing Module

[0082] As Figure 2 shown, it is used to utilize AI-enhanced data assimilation technology to fuse the multi-source dataset and the output of the numerical model of traditional numerical weather prediction (NWP) to obtain an original fusion dataset, and then perform data cleaning and quality improvement on the original fusion dataset to obtain a multi-source fusion dataset. The data fusion and processing module includes:

[0083] 2.1 Multi-source Data Fusion Unit

[0084] It is used to combine deep learning networks (such as neural networks, Transformers) with data assimilation techniques, optimize the assimilation weights, obtain AI-enhanced data assimilation techniques, and use the AI-enhanced data assimilation techniques to fuse numerical weather prediction results with the multi-source dataset to obtain the original fused data;

[0085] Then, deep learning technology is used to interpolate and reconstruct the missing grid points in the original fused data to output a complete spatial distribution map; for example, a spatio-temporal convolutional network (ST-CNN) or a generative adversarial network (GAN) is used to reconstruct the missing area, and the observed data and historical data around the missing grid points are input to output a complete spatial distribution map.

[0086] 2.2 Anomaly Detection Unit

[0087] It is used to combine statistical methods and AI anomaly detection algorithms to identify the outliers in the original fused data, and mark and delete the identified outliers; among them, the AI anomaly detection algorithm is implemented by constructing an isolation forest model for meteorological data; the isolation forest detects isolated points (outliers) by randomly dividing the data space, and outliers usually require fewer division steps to be isolated.

[0088] For example, construct an isolation forest model, input meteorological data (such as temperature, humidity, wind speed), calculate the anomaly score of each data point, set a threshold (such as 0.7), and mark the data with an anomaly score higher than the threshold as an anomaly.

[0089] 2.3 Missing Value Filling Unit

[0090] It is used to fill the original fused data containing missing values from the dimensions of spatial features and time features by using a spatio-temporal convolutional network. The spatio-temporal convolutional network extracts the spatial and time features of the data through convolutional operations for missing value filling.

[0091] For example, input meteorological data containing missing values (such as temperature, humidity grids), extract spatial features through the spatial convolutional layer, extract time features through the time convolutional layer, and the output layer predicts the missing values. The spatio-temporal convolutional network can train the model using complete data to minimize the error between the predicted value and the true value.

[0092] 2.4 Quality Improvement Unit

[0093] It is used to use a generative adversarial network and a variational autoencoder to generate extreme weather samples, oversample the extreme weather samples (using SMOTE), generate extreme weather synthetic samples that conform to the actual distribution, and undersample the original fused data to balance the data distribution. Then, the original fused data that has completed data cleaning and quality improvement is fused with the extreme weather synthetic samples to obtain multi-source fused data.

[0094] For example, a Generative Adversarial Network (GAN) takes normal weather data as input and outputs simulated extreme weather data such as thunderstorms and hurricanes; a Variational Autoencoder (VAE) maps the input data to a latent space through an encoder and generates new samples (extreme weather samples) from the latent space through a decoder.

[0095] 3. AI Model Design Module

[0096] Such as Figure 3 As shown, it is used to combine multiple deep learning models to construct a multi-model collaborative architecture. Based on the multi-model collaborative architecture, a distributed training framework is used for multi-model parallel training, and the best combination of network structure and hyperparameters is automatically searched for to optimize the model and obtain an initial collaborative model. The AI model design module includes:

[0097] 3.1 Collaborative Architecture Unit

[0098] 3.11 CNN Model Architecture

[0099] Processes two-dimensional image data such as satellite cloud images and radar echoes, with a resolution of H×W×C (height, width, number of channels), and identifies changes in cloud cluster structure and precipitation distribution.

[0100] Convolutional layer: Uses multiple convolutional operations to extract spatial features. The convolutional kernel size is 3×3 or 5×5, and the stride is 1. Extracting multi-scale features through convolutional kernels of different sizes can capture the local and global structures of cloud clusters.

[0101] Pooling layer: Uses max pooling or average pooling to reduce the resolution of the feature map and reduce the computational amount.

[0102] Fully connected layer: Maps the features extracted by the convolutional layer to specific prediction tasks (such as precipitation probability, cloud cluster movement direction).

[0103] Output layer: Outputs two-dimensional prediction results (such as a precipitation distribution map).

[0104] 3.12 LSTM / GRU Model Architecture

[0105] Processes time series data and captures the medium- and short-term weather evolution trends.

[0106] Input data: Time series data from ground meteorological stations (such as temperature, humidity, wind speed), with a shape of T×F (time step, number of features).

[0107] LSTM / GRU unit: Uses LSTM or GRU units to capture long-term dependencies in time series. Bidirectional LSTM / GRU can capture both forward and backward dependencies in time series.

[0108] Multi - layer stacking: Stack multiple LSTM / GRU units to extract higher - level temporal features.

[0109] Fully - connected layer: Map the temporal features to specific prediction tasks (such as temperature changes in the next 6 hours).

[0110] Output layer: Output the time - series prediction results.

[0111] 3.13 Transformer model architecture

[0112] Capture long - range dependencies, predict long - term climate or large - area weather forecasts, and obtain the prediction results of multiple prediction models.

[0113] Input data: Long - term meteorological data (such as hourly temperature, humidity, wind speed), with a shape of \(T×F\).

[0114] Encoder: Extract the features of the input sequence.

[0115] Decoder: Generate the prediction sequence.

[0116] Self - attention mechanism: Use multi - head self - attention to capture long - range dependencies in the time series.

[0117] Positional encoding: Add positional encoding to the input sequence to retain the order information of time steps.

[0118] The expression of the multi - model collaborative framework described in 33.14 is:

[0119] Y final = w cnn ·Y cnn + w LSSSTM ·Y LSTM + w Transformer ·Y Transformer

[0120] Where Y cnn 、Y LSTM 、Y Transformer are the prediction results of the CNN, LSTM / GRU, and Transformer models respectively, and w cnn 、w LSSSTM 、w Transformer are all model weights.

[0121] 3.2 Multi - task learning unit

[0122] Used to design a multi - task learning model to extract the general features of the multi - source fusion data. The multi - task learning model includes:

[0123] Input layer: Input the multi-source fusion data;

[0124] Shared feature extraction layer: Use CNN, LSTM / GRU or Transformer to extract the general features of the multi-source fusion data;

[0125] Task branch layer: Design an independent branch network for each prediction task (such as temperature, precipitation, wind speed);

[0126] Output layer: Output the prediction results corresponding to each branch network.

[0127] 3.3 Model training unit

[0128] Used to use a distributed training framework (such as TensorFlow Distributed, PyTorch Distributed) to allocate different prediction models in the multi-model collaborative architecture to different GPUs for parallel training.

[0129] Framework configuration:

[0130] TensorFlow Distributed: Use `tf.distribute.MirroredStrategy` for synchronous training on multiple GPUs; use `tf.data.Dataset` for data loading and distribution.

[0131] PyTorch Distributed: Use the `torch.distributed` module to initialize the distributed training environment; use `DistributedDataParallel` to wrap the model to achieve efficient gradient synchronization.

[0132] 3.4 Hyperparameter optimization unit

[0133] Used to automatically search for the optimal hyperparameters based on the defined hyperparameter space, using optimization methods (such as Bayesian optimization) and optimization tools (such as Optuna, Ray Tune, Hyperopt).

[0134] Define hyperparameter space: Learning rate: [10 -5 , 10 -2 , Network depth: [2, 10], Convolution kernel size: [3, 7].

[0135] 4. Model fusion module

[0136] Such as Figure 3As shown in the figure, it is used to deploy a weighted fusion algorithm and a meta-model in the initial collaborative model to combine prediction results, and introduce a statistical equation and Bayesian deep learning in the initial collaborative model to calculate the prediction distribution, resulting in an AI model. The model fusion module includes:

[0137] 4.1 Result Fusion Unit

[0138] It is used to assign different weights to different prediction models, and then combine the prediction results to obtain a comprehensive result. The expression of the comprehensive result is:

[0139]

[0140] where, Y fused is the comprehensive prediction result, Y i is the prediction result of the i-th model, w i is the weight of the i-th model, and N is the number of models participating in the fusion.

[0141] 4.2 Ensemble Prediction Unit

[0142] It is used to calculate the mean and variance of the comprehensive result by combining the Bayesian deep learning method to obtain the prediction distribution. The expression of the prediction distribution is:

[0143]

[0144] where, μ pred is the mean of the prediction distribution, σ 2 pred is the variance of the prediction distribution, T is the number of sampling times, is the comprehensive prediction result of the t-th sampling.

[0145] 5. Real-time Prediction Module

[0146] As Figure 3 shown, it is used to introduce an online training mechanism in the AI model and update the prediction result according to real-time data.

[0147] For example:

[0148] Data stream access: Real-time receive newly arrived observation data (such as weather station data, satellite observation data), and merge the new data with historical data to form a dynamically updated training set.

[0149] Mini-batch training: Divide the new data into small batches (mini-batch), gradually update the model parameters, and optimize the model using incremental training methods (such as SGD, Adam) to avoid retraining the entire model.

[0150] Online learning framework: Use deep learning frameworks that support online learning (such as PyTorch Lightning, TensorFlow Serving). Only update some of the model's parameters through the `partial_fit` or `warm_start` mechanism.

[0151] 6. Visualization and interpretability module

[0152] Used to design a visualization interface to output meteorological forecast results diversely, and introduce interpretability techniques in the visualization interface to verify the rationality of prediction results.

[0153] The visualization interface includes a spatial distribution map, a wind field animation, extreme weather warnings, and a time axis control for setting the prediction time step. The interpretability techniques include SHAP and Attention visualization. For example:

[0154] Dynamic animation: Update the wind field streamline animation in real time to show the dynamic changes in wind speed and direction. Use gradient colors to display continuous variables such as precipitation and temperature distribution.

[0155] Multi-time scale display: Short-term forecast, hourly forecast results (such as temperature and precipitation probability in the next 6 hours). Medium-term forecast, daily forecast results (such as weather trends in the next 7 days). Long-term forecast, weekly to monthly forecast results (such as monthly average temperature and total precipitation).

[0156] Extreme weather warning: Highlight extreme weather events (such as typhoon paths, heavy rain areas, extreme high temperature areas). Add warning marks (such as color coding, warning symbols).

[0157] Time axis control: Support users to select different time steps (such as the next 1 hour, 6 hours, 12 hours). Dynamically load the prediction data corresponding to the selected time and update the map and charts.

[0158] SHAP (Shapley Additive Explanations): Explain how the model output is affected by different input variables by assigning importance values to features.

[0159] Attention visualization: For models using the attention mechanism (such as Transformer), visualize the attention weights to show the model's attention to different time steps or spatial regions.

[0160] 7. Evaluation and iteration module

[0161] Used to combine evaluation metrics to comprehensively evaluate and continuously improve prediction results for different weather elements and time scales. The evaluation metrics include:

[0162] Mean Squared Error (MSE), which is suitable for measuring the prediction deviation of continuous variables (such as temperature, precipitation). It is sensitive to large errors and suitable for evaluating the accuracy of extreme weather predictions.

[0163] Mean Absolute Error (MAE), which more intuitively reflects the average level of prediction error and is insensitive to outliers. It is suitable for evaluating the prediction accuracy of regular weather.

[0164] Correlation Coefficient (CC), which measures the correlation between the predicted value and the true value, ranging from [-1, 1]. It is suitable for evaluating the prediction performance of time series data (such as long-term climate trends).

[0165] Skill Score (TS), which is specifically used for binary classification problems (such as whether precipitation occurs). It is suitable for evaluating the forecast accuracy of extreme weather events (such as heavy rain, typhoon).

[0166] Therefore, by providing a meteorological forecast learning system based on artificial intelligence algorithms, the present invention can obtain accurate and reliable observation data, and through the combination of artificial intelligence algorithms, efficiently process massive meteorological data, automatically conduct trend analysis and prediction, generate reliable prediction results, and improve the overall intelligent level.

[0167] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other.

[0168] Specific examples are used in this article to elaborate on the principles and implementation methods of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, based on the idea of the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A weather forecast learning system based on artificial intelligence algorithm, characterized in that: include: The data access and acquisition module is used to obtain various types of data from various sources and introduce external data sources to obtain original data sets, and then standardize and transform the original data sets to obtain multi-source data sets; A data fusion and processing module, which is used to fuse the multi-source data set with the numerical model output of the traditional numerical weather forecast using AI-enhanced data assimilation technology to obtain an original fused data set, and then perform data cleaning and quality improvement on the original fused data set to obtain a multi-source fused data set; The AI ​​model design module is used to combine multiple deep learning models to build a multi-model collaborative architecture. Based on the multi-model collaborative architecture, a distributed training framework is used to perform multi-model parallel training, and the optimal combination of network structure and hyperparameters is automatically found to optimize the model and obtain an initial collaborative model. A model fusion module, used to deploy a weighted fusion algorithm and a meta-model in the initial collaborative model to combine prediction results, and introduce statistical equations and Bayesian deep learning into the initial collaborative model to perform prediction distribution calculation to obtain an AI model; A real-time prediction module, used to introduce an online training mechanism into the AI ​​model and update the prediction results according to real-time data; The visualization and interpretability module is used to design a visualization interface, output the weather forecast results in a diversified manner, and introduce interpretability technology into the visualization interface to verify the rationality of the forecast results; Evaluation and iteration module, which is used to combine evaluation indicators to conduct comprehensive evaluation and continuous improvement of forecast results for different weather elements and time scales; Among them, the data access and collection module, the data fusion and processing module, the AI ​​model design module, the model fusion module, the real-time prediction module, the visualization and explainable module and the evaluation and iteration module are interconnected.

2. A weather forecast learning system based on artificial intelligence algorithm according to claim 1, characterized in that: The data access and collection module includes: The data acquisition unit is used to acquire various types of data from various sources in combination with the breakpoint-resume transmission mechanism, and configure corresponding data interface information, data transmission protocol and file storage format for each data source; wherein the various types of data include grid data, image data and time series data; A data introduction unit, used to introduce additional geographic information, marine meteorological information, transportation information and energy consumption information into the acquired data to obtain an original data set, and perform hash verification and integrity check on the original data set; A standardization and conversion unit, used to unify the coordinate system and time step of the original data set, construct a metadata standard for the original data set, and project the image data onto a grid with consistent grid point data to obtain a multi-source data set; The storage and management unit is used to store the multi-source data set in a distributed file system, and to construct a spatial index, a time index and a unified API interface in the distributed file system.

3. A weather forecast learning system based on artificial intelligence algorithm according to claim 2, characterized in that: The data fusion and processing module includes: A multi-source data fusion unit is used to combine a deep learning network with a data assimilation technology, optimize the assimilation weight, obtain an AI-enhanced data assimilation technology, use the AI-enhanced data assimilation technology to fuse the numerical weather forecast results with the multi-source data set to obtain original fused data, and then use the deep learning technology to interpolate and reconstruct the missing grid points in the original fused data to output a complete spatial distribution map; An anomaly detection unit, used to identify the outliers of the original fused data by combining statistical methods and AI anomaly detection algorithms, and mark and delete the identified outliers; wherein the AI ​​anomaly detection algorithm is implemented by constructing an isolation forest model of meteorological data; A missing value filling unit is used to fill the original fused data containing missing values ​​from the spatial feature and temporal feature dimensions by using a spatiotemporal convolutional network; The quality improvement unit is used to generate extreme weather samples by using a generative adversarial network and a variational autoencoder, oversample the extreme weather samples to generate extreme weather synthetic samples that conform to the actual distribution, downsample the original fused data to balance the data distribution, and then fuse the original fused data that has completed data cleaning and quality improvement with the extreme weather synthetic samples to obtain multi-source fused data.

4. A weather forecast learning system based on artificial intelligence algorithm according to claim 3, characterized in that: The AI ​​model design module includes: The collaborative architecture unit is used to process two-dimensional image data using the CNN model, identify changes in cloud structure and precipitation distribution, process time series data using the LSTM / GRU model, capture medium- and short-term weather evolution trends, and use the Transformer model to capture long-distance dependencies, predict long-term climate or large-area weather forecasts, and obtain prediction results from multiple prediction models; A multi-task learning unit, used for designing a multi-task learning model and extracting common features of the multi-source fusion data; The model training unit is used to use the distributed training framework to distribute different prediction models in the multi-model collaborative architecture to different GPUs for parallel training; The hyperparameter optimization unit is used to automatically search for the optimal hyperparameters based on the defined hyperparameter space using optimization methods and optimization tools.

5. A weather forecast learning system based on artificial intelligence algorithm according to claim 4, characterized in that: The multi-task learning model includes an input layer for inputting the multi-source fusion data, a shared feature extraction layer for extracting common features of the multi-source fusion data, a task branching layer for designing an independent branch network for each prediction task, and an output layer for outputting the prediction result of each branch network.

6. A weather forecast learning system based on artificial intelligence algorithm according to claim 5, characterized in that: The expression of the multi-model collaborative framework is: Y final =w cnn ·Y cnn +w LSSSTM ·Y LSTM +w Transformer ·Y Transformer Among them, Y cnn , Y LSTM , Y Transformer are the prediction results of CNN, LSTM / GRU and Transformer models respectively, w cnn 、w LSSSTM 、w Transformer are model weights.

7. A weather forecast learning system based on artificial intelligence algorithm according to claim 6, characterized in that: The model fusion module includes: The result fusion unit is used to assign different weights to different prediction models, and then combine the prediction results to obtain a comprehensive result; The ensemble prediction unit is used to calculate the mean and variance of the comprehensive result in combination with the Bayesian deep learning method to obtain the predicted distribution.

8. A weather forecast learning system based on artificial intelligence algorithm according to claim 7, characterized in that: The expression of the comprehensive result is: Among them, Y fused is the comprehensive prediction result, Y i is the prediction result of the i-th model, w i is the weight of the i-th model, and N is the number of models involved in the fusion; The expression of the predicted distribution is: Among them, μ pred is the mean of the prediction distribution, σ 2 pred is the variance of the predicted distribution, T is the number of samplings, is the comprehensive prediction result of the t-th sampling.

9. A weather forecast learning system based on artificial intelligence algorithm according to claim 8, characterized in that: The visualization interface includes a spatial distribution map, a wind field animation, an extreme weather warning, and a timeline control for setting the prediction time step. The interpretability techniques include SHAP and Attention visualization. The evaluation indicators include mean square error, mean absolute error, correlation coefficient, and skill score.

Images