High-temporal-spatial-resolution short-term imminent forecasting method for atmosphere temperature and humidity based on machine learning

Abstract

The invention provides a machine learning-based atmosphere temperature and humidity high-temporal-spatial-resolution short-term and imminent forecasting method, and belongs to the technical field of weather forecasting. The method comprises the following steps: inputting a low-spatial-resolution coarse-grained weather data sample; inputting the samples into a prediction framework, and changing time and space resolutions; a channel attention mechanism module CAM is introduced, and sample features under different time representations are learned; carrying out inverse sampling on output samples of the CAM and shaping the output samples to a specific channel number and a specific spatial resolution; splicing an output sample on the channel dimension; performing convolution operation on the spliced samples for a plurality of times to generate a high-spatial-resolution forecasting result; the method is based on 3D convolution and a channel attention mechanism, and is used for comprehensively utilizing meteorological data features under different spatial-temporal scales so as to jointly improve the quality of weather forecast generated by the model from two aspects of error indexes and spatial-temporal resolution.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting technology, specifically, it relates to a high spatiotemporal resolution short-term forecasting method for atmospheric temperature and humidity based on machine learning. Background Technology

[0002] Atmospheric temperature and humidity are essential parameters for describing the thermal and dynamic state of the atmosphere. Predicting atmospheric temperature and relative humidity through technological means plays an important role in many fields such as scientific research, production and daily life. To date, the world's mainstream weather forecasting models are mainly divided into two types: (1) numerical weather prediction model methods based on physical states; (2) weather forecasting methods based on machine learning modeling.

[0003] Numerical weather prediction models based on physical states simulate the future state of the atmosphere, breaking down possible meteorological characteristic data through simulation to generate weather forecasts. However, running and maintaining such models typically requires extremely high-performance computing systems. Some complex simulations require hundreds or thousands of computing resources and can take several hours to obtain relatively accurate results, resulting in significant time and space overhead. Numerical weather prediction models simulate the entire Earth's atmosphere, whose state is complex and variable, placing extremely high demands on human and material resources. Therefore, many researchers hope to find methods to reduce resource consumption in order to generate weather forecasts more efficiently.

[0004] For weather forecasting methods based on machine learning modeling, significant breakthroughs have been made in this century, leading to its successful introduction into meteorological fields. Improvements in machine learning models and rapid iterations by numerous scholars have made them a current focus. Consequently, increasing efforts have been devoted to replacing traditional numerical weather prediction models with machine learning models in recent years. However, machine learning models still have certain limitations: due to differences in observation methods and the spatiotemporal limitations of the equipment itself, weather forecasts generated by machine learning models suffer from insufficient spatiotemporal resolution. This results in insufficient accuracy across the spatiotemporal span of the forecasts, limiting their practical applicability. Furthermore, research has found that machine learning models still exhibit some performance issues when performing long-term forecasting.

[0005] Research reveals that the problems with existing machine learning-based weather forecasting models primarily stem from two issues: spatiotemporal resolution and long-term forecasting. Firstly, regarding spatiotemporal resolution, current models focus on improving forecast accuracy using meteorological data at a specific scale. However, due to insufficient spatiotemporal resolution, even highly accurate forecasts offer limited real-world relevance. Secondly, the complex and highly chaotic nature of the Earth's atmospheric environment inevitably leads to error accumulation in long-term forecasts. This results in high-quality short-term forecasts but a decline in accuracy for long-term forecasts. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a high spatiotemporal resolution short-term weather forecasting method based on machine learning for atmospheric temperature and humidity. This method utilizes 3D convolution and channel attention mechanisms to comprehensively leverage meteorological data features at different spatiotemporal scales, thereby improving the quality of model-generated weather forecasts from both error index and spatiotemporal resolution perspectives.

[0007] This invention is achieved through the following technical solution: A high spatiotemporal resolution short-term forecasting method for atmospheric temperature and humidity based on machine learning: The method specifically includes the following steps: Step 1: Input low spatial resolution coarse-grained meteorological data samples; Step 2: Input the samples from Step 1 into the prediction framework and change the temporal and spatial resolution. Step 3: Introduce the Channel Attention (CAM) module to learn the sample features under different time representations; Step 4: Inverse sample the output samples of the CAM and reshape them to a specific number of channels and a specific spatial resolution; Step 5: Concatenate the output samples along the channel dimension; Step 6: Perform several convolution operations on the stitched samples to generate high spatial resolution prediction results.

[0008] Furthermore, in step 1, the input is three coarse-grained data samples with a spatiotemporal resolution of 2 hours each, representing a low spatial resolution.

[0009] Furthermore, in step 2, the prediction framework consists of four pathways with identical modules but different scales: two high-scale pathways and two low-scale pathways; each pathway comprises a Cov-3D module, an upsampling module, and a downsampling module. The Cov-3D module changes the number of channels in the data sample, which in turn changes the temporal resolution of the meteorological data sample itself. The upsampling module uses a trilinear interpolation method to improve spatial resolution; The downsampling module uses strided convolution with a stride of 2 to reduce spatial resolution.

[0010] Furthermore, in step 2, the high-scale path and low-scale path sampling modules are used in different ways, and the number of sampling modules used is different. By combining upsampling and downsampling modules, the spatial resolution of the sample is gradually changed to achieve hierarchical feature extraction. The Cov-3D module changes the temporal resolution to extract the spatiotemporal features contained in the original data samples; before the channel attention mechanism module, it continuously increases the number of channels of the samples through 3D convolution.

[0011] Furthermore, in step 3, the Channel Attention Mechanism (CAM) module performs average pooling and max pooling on the intermediate samples, inputs the pooled samples into the MLP for learning and outputs a weight matrix, and multiplies the weight matrix with the original samples to obtain new samples.

[0012] Furthermore, in step 5, the number of channel dimensions is the same as that of the prediction framework in step 2, which is 4. After splicing, intermediate data samples from different pathways are combined, and sample features from different spatiotemporal representations are integrated, so that the final weather forecast is closer to the real situation in both the overall and local aspects.

[0013] Furthermore, in step 6, the output is a high spatial resolution sample with a spatiotemporal resolution of 6 samples and a temporal resolution of 1 hour.

[0014] A high spatiotemporal resolution short-term forecasting system for ambient temperature and humidity based on machine learning: The forecasting system includes an input module, a multi-path processing module, a channel attention module, a feature fusion module, and a forecasting module. The input module is used to input low spatial resolution coarse-grained meteorological data samples; The multi-path processing module inputs the samples from the input module into the prediction framework and modifies the temporal and spatial resolution. A channel attention mechanism module (CAM) is introduced to learn sample features under different temporal representations. The feature fusion module inverse samples the output samples of CAM and reshapes them to a specific number of channels and a specific spatial resolution; then it stitches the output samples together along the channel dimension. The forecasting module performs several convolution operations on the stitched samples to generate high spatial resolution forecasting results.

[0015] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.

[0016] A computer-readable storage medium for storing computer instructions that, when executed by a processor, implement the steps of the above-described method.

[0017] Beneficial effects of the invention This invention focuses on improving spatiotemporal resolution in its modeling approach. The advantage of this approach is that even if the spatiotemporal resolution of actual observation equipment is insufficient, technical means can still be used to enable the model to generate high-quality forecasts with higher spatiotemporal resolution. The original spatiotemporal resolution samples are improved at the spatiotemporal level to provide the weather forecast data generated by the machine learning model with a more detailed spatiotemporal distribution, thereby enhancing its practical value.

[0018] This invention addresses the problems arising from long-series forecasting by employing short-term forecasting to minimize the error accumulation issues associated with long-series forecasts. It also establishes a robust data acquisition mechanism to ensure timely forecasting. Furthermore, it fully utilizes and combines data sample characteristics from different spatiotemporal scales, considering both overall and local data sample features, thereby improving the accuracy of the final weather forecast and making it more consistent with reality. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the external input and output of the model of the present invention.

[0020] Figure 2 It is an internal component of the Channel Attention Mechanism (CAM) module.

[0021] Figure 3 This is a framework diagram of a short-term forecasting model based on 3D convolution and channel attention mechanisms. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Unless otherwise specified, the experimental methods used in the following examples are conventional methods. Unless otherwise specified, the materials, reagents, methods, and instruments used are all conventional materials, reagents, methods, and instruments in the art, and can be obtained commercially by those skilled in the art.

[0024] This invention proposes a high spatiotemporal resolution short-term forecasting method for atmospheric temperature and humidity based on machine learning, employing a short-term prediction framework based on 3D convolution and channel attention mechanisms. This scheme focuses on short-term (0-6 hour) forecasts, such as... Figure 1 As shown, the model input consists of three low-spatial-resolution coarse-grained data samples with a temporal resolution of 2 hours each, and the output consists of six high-spatial-resolution samples with a temporal resolution of 1 hour each. Standardizing the model's input and output to this level can improve the spatiotemporal resolution of weather forecasts.

[0025] The prediction framework proposed in this invention consists of four pathways with identical modules but different scales: two high-scale pathways and two low-scale pathways. Each pathway itself comprises three modules: a Cov-3D module, an Upsampling module, and a Subsampling module. The Upsampling module uses trilinear interpolation, while the Subsampling module uses strided convolution with a stride of 2. The Cov3D module is responsible for changing the number of channels in the data sample (the temporal resolution of the meteorological data sample itself), while the Upsampling and Subsampling modules are responsible for changing the spatial resolution of the sample data. Through the cooperation of the three modules mentioned above, the scale of the data sample can be changed step by step, and the sample features represented at different spatiotemporal scales can be extracted layer by layer and scale by scale to take into account both global and local meteorological data features.

[0026] Simultaneously, the predictive architecture of this invention incorporates a channel attention mechanism (CAM) module during the path design to enhance the extraction of temporal features from data samples. The specific channel attention mechanism module is as follows: Figure 2 As shown, the intermediate samples are first subjected to average pooling and max pooling respectively. Then, the two pooled samples are used as input to the MLP, which learns from the data samples and outputs a weight matrix. Finally, the weight matrix is ​​multiplied by the original samples to obtain new samples. It should be noted that the channel attention mechanism module designed in this invention is adapted to three-dimensional meteorological samples, making the location information of meteorological data more three-dimensional, and can take into account the three-dimensional information of longitude, latitude, and altitude.

[0027] The method proposed in this invention is illustrated in the embodiments as follows: Figure 3 As shown, the input raw data is a 3-hour low spatiotemporal resolution data sample. The sample features are extracted through 4 channels. The difference between the 4 channels lies in the different module usage orientations and the number of modules used for sampling methods (upsampling and downsampling), thereby continuously changing the spatial resolution of the sample data.

[0028] Simultaneously, by incorporating 3D convolution to alter the temporal resolution, this modular approach to integer data allows for the full extraction of the spatiotemporal features inherent in the original data samples. Prior to the channel attention mechanism module, 3D convolution continuously increases the number of channels in the samples.

[0029] Subsequently, after increasing the number of sample channels to a certain extent, a channel attention mechanism module is introduced into the processing of intermediate data samples. The MLP learns from the intermediate samples and outputs a weight matrix. After obtaining the weight of each feature channel, the weight is applied to each original feature channel. Based on the spatiotemporal feature enhancement task of this invention, the importance of sample features under different channels, i.e., different time representations, can be learned.

[0030] Then, inverse sampling is performed, and the data samples are rounded to a specific number of channels and a specific spatial resolution, such as 64 channels in the figure.

[0031] Finally, the splicing operation and final convolution are performed. The splicing operation splices the samples along the channel dimension. The purpose of this step is to combine intermediate data samples from different channels so that the sample features under different spatiotemporal representations can be organically integrated, which helps the final weather forecast to be as close as possible to the real situation in both the overall and local aspects.

[0032] Finally, the stitched data samples are subjected to several convolution operations until a weather forecast with the expected spatiotemporal resolution is generated.

[0033] At the same time, it needs to be explained that, Figure 3 The constants mentioned are examples. In practical applications, the constants should be flexibly set based on the specific task size, such as the spatiotemporal resolution requirements of weather forecasts, the dimensionality of input samples, and the actual forecast results obtained through experiments. This invention presents an explanation of the method; the actual implementation of the model framework still requires specific analysis based on the specific circumstances.

[0034] A high spatiotemporal resolution short-term forecasting system for ambient temperature and humidity based on machine learning: The forecasting system includes an input module, a multi-path processing module, a channel attention module, a feature fusion module, and a forecasting module. The input module is used to input low spatial resolution coarse-grained meteorological data samples; The multi-path processing module inputs the samples from the input module into the prediction framework and modifies the temporal and spatial resolution. A channel attention mechanism module (CAM) is introduced to learn sample features under different temporal representations. The feature fusion module inverse samples the output samples of CAM and reshapes them to a specific number of channels and a specific spatial resolution; then it stitches the output samples together along the channel dimension. The forecasting module performs several convolution operations on the stitched samples to generate high spatial resolution forecasting results.

[0035] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.

[0036] A computer-readable storage medium for storing computer instructions that, when executed by a processor, implement the steps of the above-described method.

[0037] The memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory of the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0038] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means such as coaxial cable, optical fiber, digital subscriber line, DSL, or wireless means such as infrared, wireless, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium such as a floppy disk, hard disk, magnetic tape; an optical medium such as a high-density digital video disc, DVD; or a semiconductor medium such as a solid-state disk, SSD, etc.

[0039] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this invention can be directly manifested as execution by a hardware processor, or as a combination of hardware and software modules within the processor. The software modules can reside in mature storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0040] It should be noted that the processor in the embodiments of the present invention can be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method embodiments can be completed by the integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of the present invention can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0041] The above provides a detailed description of the high spatiotemporal resolution short-term forecasting method for atmospheric temperature and humidity based on machine learning proposed in this invention. The principles and implementation methods of this invention have been explained. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A high spatiotemporal resolution short-term forecasting method for atmospheric temperature and humidity based on machine learning, characterized in that: The method specifically includes the following steps: Step 1: Input low spatial resolution coarse-grained meteorological data samples; Step 2: Input the samples from Step 1 into the prediction framework and change the temporal and spatial resolution. Step 3: Introduce the Channel Attention (CAM) module to learn the sample features under different time representations; Step 4: Inverse sample the output samples of the CAM and reshape them to a specific number of channels and a specific spatial resolution; Step 5: Concatenate the output samples along the channel dimension; Step 6: Perform several convolution operations on the stitched samples to generate high spatial resolution prediction results.

2. The forecasting method according to claim 2, characterized in that: In step 1, the input is three coarse-grained data samples with a spatiotemporal resolution of 2 hours each, representing a low spatial resolution.

3. The forecasting method according to claim 2, characterized in that: In step 2, the prediction framework consists of four pathways with the same modules but different scales: two high-scale pathways and two low-scale pathways. Each pathway comprises a Cov-3D module, an upsampling module, and a downsampling module. The Cov-3D module changes the number of channels in the data sample, which in turn changes the temporal resolution of the meteorological data sample itself. The upsampling module uses a trilinear interpolation method to improve spatial resolution; The downsampling module uses strided convolution with a stride of 2 to reduce spatial resolution.

4. The forecasting method according to claim 3, characterized in that: In step 2, the high-scale and low-scale path sampling modules are used differently in terms of usage orientation and the number of sampling modules used. By combining upsampling and downsampling modules, the spatial resolution of the sample is gradually changed to achieve hierarchical feature extraction. The Cov-3D module changes the temporal resolution to extract the spatiotemporal features contained in the original data samples; before the channel attention mechanism module, it continuously increases the number of channels of the samples through 3D convolution.

5. The forecasting method according to claim 4, characterized in that: In step 3, the Channel Attention Mechanism (CAM) module performs average pooling and max pooling on the intermediate samples, inputs the pooled samples into the MLP for learning and outputs a weight matrix, and multiplies the weight matrix with the original samples to obtain new samples.

6. The forecasting method according to claim 5, characterized in that: In step 5, the number of channel dimensions is the same as that in the prediction framework in step 2, which is 4. After splicing, intermediate data samples from different pathways are combined, and sample features from different spatiotemporal representations are integrated, so that the final weather forecast is closer to the real situation in both the overall and local aspects.

7. The forecasting method according to claim 6, characterized in that: In step 6, the output is a high spatial resolution sample with a spatiotemporal resolution of 6 samples and a temporal resolution of 1 hour.

8. A forecasting system for executing the high spatiotemporal resolution short-term forecasting method for atmospheric temperature and humidity based on machine learning as described in any one of claims 1 to 7, characterized in that: The forecasting system includes an input module, a multi-path processing module, a channel attention module, a feature fusion module, and a forecasting module. The input module is used to input low spatial resolution coarse-grained meteorological data samples; The multi-path processing module inputs the samples from the input module into the prediction framework and modifies the temporal and spatial resolution. A channel attention mechanism module (CAM) is introduced to learn sample features under different temporal representations. The feature fusion module inverse samples the output samples of CAM and reshapes them to a specific number of channels and a specific spatial resolution; then it stitches the output samples together along the channel dimension. The forecasting module performs several convolution operations on the stitched samples to generate high spatial resolution forecasting results.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.

Images