Short temporary rainfall forecasting method based on TimeUnet and phenology knowledge
Through the short-term precipitation forecast method based on TimeUnet and phenology knowledge, combined with meteorological elements and bird sound abnormal information, the problem of difficulty in predicting extreme precipitation in the existing technology is solved, and high-quality short-term precipitation prediction and heavy rain forecast are achieved.
Patent Information
- Application Number
- CN202411565389.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-05
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-11-05
AI Technical Summary
The existing short-term precipitation forecasting methods are difficult to effectively characterize the interaction between nonlinear characteristics of atmospheric motion and the spatial and temporal information of different scales, and are difficult to predict extreme values, resulting in limited forecasting performance, especially the low predictability of heavy rain forecasting.
Using a short-term precipitation forecast method based on TimeUnet and phenology knowledge, we collect meteorological element observation data and bird sound abnormal information, feature extraction and preprocessing, and build a deep learning model TimeUnet that integrates TimesNet and Unet, and design customized losses. Functions, extract information from different time and space scales to improve forecast accuracy.
It effectively improves the skills of short-term precipitation forecasting and forecast performance in extreme scenarios, especially heavy rain forecasting, providing additional forecasting factors, and improving the application value of the model in extreme scenarios.
Smart Images

Figure CN119066524A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of precipitation forecasting, and in particular to a short-term precipitation forecasting method based on TimeUnet and phenological knowledge. Background Art
[0002] Extrapolating the precipitation in the future based on the current weather background is a commonly used short-term precipitation forecasting method. The most common method is the optical flow method, which has been widely used in the business sector. However, due to the chaotic characteristics of the atmosphere, this method is difficult to fully characterize the nonlinear characteristics of atmospheric motion, so the forecasting performance of this method is greatly limited. In recent years, deep learning technology has been widely used for its high nonlinearity and strong robustness. Some scholars have also applied it to the short-term precipitation forecasting, such as ConvLSTM, Unet and other algorithms, which show higher forecasting skills than the traditional optical flow method. However, the current deep learning algorithms are still difficult to characterize the interaction of spatiotemporal information at different scales and it is difficult to forecast and warn extreme values well, which to a certain extent limits the forecasting skills of short-term precipitation (especially heavy rain). At the same time, due to the limitation of the source of predictability, even the short-term precipitation forecasting based on deep learning algorithms faces the problem of low predictability, and it is urgent to explore additional sources of predictability. Considering the response of bird activities to environmental changes (especially to heavy rain), feature extraction from abnormal bird sound information can provide an additional window of opportunity for short-term precipitation forecasting, which is still in the initial stage of development. Summary of the invention
[0003] Purpose of the invention: The purpose of the present invention is to provide a method for forecasting short-term precipitation based on TimeUnet and phenological knowledge, extract features from abnormal bird sound information, and solve the problem of short-term precipitation forecasting.
[0004] Technical solution: The method for short-term precipitation forecasting based on TimeUnet and phenological knowledge described in the present invention comprises the following steps: (1) Collect meteorological element observation data and bird sound abnormality information samples, and pre-process the data; (2) Extract features from the preprocessed meteorological element observation data and bird sound abnormal information samples to construct a feature data set; (3) Build a deep learning model TimeUnet that combines TimesNet and Unet and design a customized loss function; (4) Train the TimeUnet model and adjust the hyperparameters in the model to obtain the optimal model; (5) Generate short-term precipitation forecast products based on real-time meteorological element observation data and bird sound abnormal information samples.
[0005] Furthermore, in step (1), the meteorological element observation data includes precipitation data, pressure field data, humidity data and wind field data at the current time and in the past 2 hours; the bird sound abnormal information sample includes information at the current time and in the past 2 hours; the preprocessing is as follows: Based on the inverse distance weighted interpolation, the station data is interpolated into uniform grid data; the outliers in the grid data are removed; the calculation formula of the inverse distance weighted interpolation is as follows: ; Where n represents the total number of sites around the target grid point, represents the observation data of the ith station, represents the distance from the i-th station to the target grid point, Represents the interpolated grid data.
[0006] Furthermore, step (2) includes the following steps: (21) Based on short-time Fourier transform, the bird sound characteristic spectrum containing time-frequency domain characteristic information is obtained; the formula is as follows: ; in, is the extracted bird sound feature map. is the original sound spectrum at time T, is the analysis window function, is a complex function; (22) De-dimensionalize the meteorological element data to obtain the characteristic map of meteorological elements; the formula is as follows: ; Among them, f is the original meteorological element data, is the average value of meteorological element data, is the standard deviation of meteorological element data.
[0007] Furthermore, in step (3), the TimeUnet network includes two modules, MFCM-Unet and TimesNet, and includes the following steps: (31) Inputting the meteorological element feature map obtained in step (2) into the MFCM-Unet network to extract meteorological information of different spatial scales; the MFCM-Unet network includes an encoder, a channel attention mechanism based on multiple features, and a decoder; wherein the encoder is used to extract features of different spatial scales; the channel attention mechanism based on multiple features is used to give higher weights to important forecast features to highlight important information; the decoder is used to reconstruct and generate the final forecast; at the same time, the bird sound feature map TimesNet obtained in step (2) captures abnormal bird sound information of different time scales; including: signal reading, signal decomposition and construction of a two-dimensional signal field, and signal reconstruction; (32) Design a customized loss function, the formula is as follows: ; in, , , , They represent the root mean square error of light rain, moderate rain, heavy rain and rainstorm respectively, , , , Respectively , , , The weight of is as follows: ; in, represents the weight of the i-th precipitation, Represents the number of samples of the i-th type of precipitation; precipitation includes: light rain, moderate rain, heavy rain, and rainstorm.
[0008] The present invention provides a short-term precipitation forecasting system based on TimeUnet and phenological knowledge, comprising: Collection module: used to collect meteorological element observation data and bird sound abnormal information samples, and pre-process the data; Dataset module: used to extract features from preprocessed meteorological element observation data and bird sound abnormal information samples to construct feature datasets; TimeUnet module: used to build the TimeUnet deep learning model that combines TimesNet and Unet and design a customized loss function; Training module: used to train the TimeUnet model and adjust the hyperparameters in the model to obtain the optimal model; Forecast module: used to generate short-term precipitation forecast products based on real-time meteorological element observation data and bird sound abnormal information samples.
[0009] Furthermore, in the acquisition module, the meteorological element observation data include precipitation data, pressure field data, humidity data and wind field data at the current time and in the past 2 hours; the bird sound abnormal information sample includes the information at the current time and in the past 2 hours; the preprocessing is as follows: interpolating the station data into uniform grid data based on inverse distance weighted interpolation; removing outliers in the grid data; the calculation formula of inverse distance weighted interpolation is as follows: ; Where n represents the total number of sites around the target grid point, represents the observation data of the ith station, represents the distance from the i-th station to the target grid point, Represents the interpolated grid data.
[0010] Furthermore, the dataset module includes the following steps: (21) Based on short-time Fourier transform, the bird sound characteristic spectrum containing time-frequency domain characteristic information is obtained; the formula is as follows: ; in, is the extracted bird sound feature map. is the original sound spectrum at time T, is the analysis window function, is a complex function; (22) De-dimensionalize the meteorological element data to obtain the characteristic map of meteorological elements; the formula is as follows: ; Among them, f is the original meteorological element data, is the average value of meteorological element data, is the standard deviation of meteorological element data.
[0011] Furthermore, in the TimeUnet module, the TimeUnet network includes two modules, MFCM-Unet and TimesNet, including the following steps: (31) Inputting the meteorological element feature map obtained in step (2) into the MFCM-Unet network to extract meteorological information of different spatial scales; the MFCM-Unet network includes an encoder, a channel attention mechanism based on multiple features, and a decoder; wherein the encoder is used to extract features of different spatial scales; the channel attention mechanism based on multiple features is used to give higher weights to important forecast features to highlight important information; the decoder is used to reconstruct and generate the final forecast; at the same time, the bird sound feature map TimesNet obtained in step (2) captures abnormal bird sound information of different time scales; including: signal reading, signal decomposition and construction of a two-dimensional signal field, and signal reconstruction; (32) Design a customized loss function, the formula is as follows: ; in, , , , They represent the root mean square error of light rain, moderate rain, heavy rain and rainstorm respectively, , , , Respectively , , , The weight of is as follows: ; in, represents the weight of the i-th precipitation, Represents the number of samples of the i-th type of precipitation; precipitation includes: light rain, moderate rain, heavy rain, and rainstorm.
[0012] An electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, any one of the methods for forecasting short-term precipitation based on TimeUnet and phenological knowledge is implemented.
[0013] A storage medium described in the present invention stores a computer program, and when the computer program is executed by a processor, any one of the short-term precipitation forecasting methods based on TimeUnet and phenological knowledge is implemented.
[0014] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: based on phenological knowledge, by extracting abnormal information of bird sounds, additional forecasting factors are provided for short-term precipitation forecasting, thereby effectively improving the forecasting skills of short-term precipitation; the TimeUnet neural network based on the combination of TimesNet and MFCM-Unet can effectively extract effective information at different time and space scales, thereby achieving high-quality short-term precipitation prediction; the introduction of a customized loss function effectively improves the forecasting skills of heavy rain, thereby improving the application value of the model in extreme scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 is a flow chart of the present invention; Figure 2 It is a schematic diagram of the model structure of the present invention. DETAILED DESCRIPTION
[0016] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.
[0017] like Figure 1 As shown, an embodiment of the present invention provides a short-term precipitation forecasting method based on TimeUnet and phenological knowledge, comprising the following steps: (1) Collect meteorological element observation data and bird sound abnormal information samples, and pre-process the data; meteorological element observation data include precipitation data, pressure field data, humidity data and wind field data at the current time and in the past 2 hours; bird sound abnormal information samples include information at the current time and in the past 2 hours; the pre-processing is as follows: Based on the inverse distance weighted interpolation, the station data is interpolated into uniform grid data; the outliers in the grid data are removed; the calculation formula of the inverse distance weighted interpolation is as follows: Based on the inverse distance weighted interpolation, the station data is interpolated into uniform grid data; the outliers in the grid data are removed; the calculation formula of the inverse distance weighted interpolation is as follows: ; Where n represents the total number of sites around the target grid point, represents the observation data of the ith station, represents the distance from the i-th station to the target grid point, Represents the interpolated grid data.
[0018] (2) Extract features from the preprocessed meteorological element observation data and bird sound abnormal information samples to construct a feature data set; the following steps are included: (21) Based on short-time Fourier transform, the bird sound characteristic spectrum containing time-frequency domain characteristic information is obtained; the formula is as follows: ; in, is the extracted bird sound feature map. is the original sound spectrum at time T, is the analysis window function, is a complex function; (22) De-dimensionalize the meteorological element data to obtain the characteristic map of meteorological elements; the formula is as follows: ; Among them, f is the original meteorological element data, is the average value of meteorological element data, is the standard deviation of meteorological element data.
[0019] (3) Build a deep learning model TimeUnet that combines TimesNet and Unet and design a customized loss function; the TimeUnet network includes two modules, MFCM-Unet and TimesNet, such as Figure 2 As shown, the following steps are included: (31) Inputting the meteorological element feature map obtained in step (2) into the MFCM-Unet network to extract meteorological information of different spatial scales; the MFCM-Unet network includes an encoder, a channel attention mechanism based on multiple features, and a decoder; wherein the encoder is used to extract features of different spatial scales; the channel attention mechanism based on multiple features is used to give higher weights to important forecast features to highlight important information; the decoder is used to reconstruct and generate the final forecast; at the same time, the bird sound feature map TimesNet obtained in step (2) captures abnormal bird sound information of different time scales; including: signal reading, signal decomposition and construction of a two-dimensional signal field, and signal reconstruction; (32) Design a customized loss function, the formula is as follows: ; in, , , , They represent the root mean square error of light rain, moderate rain, heavy rain and rainstorm respectively, , , , Respectively , , , The weight of is as follows: ; in, represents the weight of the i-th precipitation, Represents the number of samples of the i-th type of precipitation; precipitation includes: light rain, moderate rain, heavy rain, and rainstorm.
[0020] (4) Train the TimeUnet model and adjust the hyperparameters in the model to obtain the optimal model; specifically, the hyperparameters that can be adjusted in the model mainly include: learning rate, number of network layers, number of convolution kernels, etc., which can be adjusted according to the actual training results. Generally, a smaller learning rate, a larger number of network layers, and a larger number of convolution kernels means a more complex model, which usually leads to higher prediction skills, but also increases the computational requirements of the model. When the complexity of the model reaches a certain level, the skill improvement brought by a more complex model is very limited. Therefore, it is necessary to set the optimal hyperparameters according to the actual training results to ensure the performance and computational requirements of the model.
[0021] (5) Generate short-term precipitation forecast products based on real-time meteorological element observation data and bird sound abnormal information samples.
[0022] The embodiment of the present invention further provides a short-term precipitation forecasting system based on TimeUnet and phenological knowledge, including: Acquisition module: used to collect meteorological element observation data and bird sound abnormal information samples, and pre-process the data; meteorological element observation data include precipitation data, pressure field data, humidity data and wind field data at the current time and in the past 2 hours; bird sound abnormal information samples include information at the current time and in the past 2 hours; pre-processing is as follows: interpolate the station data into uniform grid data based on inverse distance weighted interpolation; remove outliers in the grid data; the calculation formula of inverse distance weighted interpolation is as follows: ; Where n represents the total number of sites around the target grid point, represents the observation data of the ith station, represents the distance from the i-th station to the target grid point, Represents the interpolated grid data.
[0023] Dataset module: used to extract features from the preprocessed meteorological element observation data and bird sound abnormal information samples and construct a feature dataset; it includes the following steps: (21) Based on short-time Fourier transform, the bird sound characteristic spectrum containing time-frequency domain characteristic information is obtained; the formula is as follows: ; in, is the extracted bird sound feature map. is the original sound spectrum at time T, is the analysis window function, is a complex function; (22) De-dimensionalize the meteorological element data to obtain the characteristic map of meteorological elements; the formula is as follows: ; Among them, f is the original meteorological element data, is the average value of meteorological element data, is the standard deviation of meteorological element data.
[0024] TimeUnet module: used to build the TimeUnet deep learning model that combines TimesNet and Unet and design a customized loss function; in the TimeUnet module, the TimeUnet network includes two modules, MFCM-Unet and TimesNet, including the following steps: (31) Inputting the meteorological element feature map obtained in step (2) into the MFCM-Unet network to extract meteorological information of different spatial scales; the MFCM-Unet network includes an encoder, a channel attention mechanism based on multiple features, and a decoder; wherein the encoder is used to extract features of different spatial scales; the channel attention mechanism based on multiple features is used to give higher weights to important forecast features to highlight important information; the decoder is used to reconstruct and generate the final forecast; at the same time, the bird sound feature map TimesNet obtained in step (2) captures abnormal bird sound information of different time scales; including: signal reading, signal decomposition and construction of a two-dimensional signal field, and signal reconstruction; (32) Design a customized loss function, the formula is as follows: ; in, , , , They represent the root mean square error of light rain, moderate rain, heavy rain and rainstorm respectively, , , , Respectively , , , The weight of is as follows: ; in, represents the weight of the i-th precipitation, Represents the number of samples of the i-th type of precipitation; precipitation includes: light rain, moderate rain, heavy rain, and rainstorm.
[0025] Training module: used to train the TimeUnet model and adjust the hyperparameters in the model to obtain the optimal model; Forecast module: used to generate short-term precipitation forecast products based on real-time meteorological element observation data and bird sound abnormal information samples.
[0026] An embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is loaded into the processor, any one of the methods for forecasting short-term precipitation based on TimeUnet and phenological knowledge is implemented.
[0027] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program, when executed by a processor, implements any one of the methods for forecasting short-term precipitation based on TimeUnet and phenological knowledge.
Claims
1. A short-term precipitation forecasting method based on TimeUnet and phenological knowledge, characterized in that: The following steps are involved: (1) Collect meteorological element observation data and bird sound abnormality information samples, and pre-process the data; (2) Extract features from the preprocessed meteorological element observation data and bird sound abnormal information samples to construct a feature data set; (3) Build a deep learning model TimeUnet that combines TimesNet and Unet and design a customized loss function; (4) Train the TimeUnet model and adjust the hyperparameters in the model to obtain the optimal model; (5) Generate short-term precipitation forecast products based on real-time meteorological element observation data and bird sound abnormal information samples.
2. A method for short-term precipitation forecasting based on TimeUnet and phenological knowledge according to claim 1, characterized in that: In step (1), the meteorological element observation data includes the precipitation data, pressure field data, humidity data and wind field data at the current time and the past 2 hours; the bird sound abnormal information sample includes the information at the current time and the past 2 hours; the preprocessing is as follows: Based on the inverse distance weighted interpolation, the station data is interpolated into uniform grid data; the outliers in the grid data are removed; the calculation formula of the inverse distance weighted interpolation is as follows: ; Where n represents the total number of sites around the target grid point, represents the observation data of the ith station, represents the distance from the i-th station to the target grid point, Represents the interpolated grid data.
3. The method for short-term precipitation forecasting based on TimeUnet and phenological knowledge according to claim 1, characterized in that: Step (2) includes the following steps: (21) Based on short-time Fourier transform, the bird sound characteristic spectrum containing time-frequency domain characteristic information is obtained; the formula is as follows: ; in, is the extracted bird sound feature map. is the original sound spectrum at time T, is the analysis window function, is a complex function; (22) De-dimensionalize the meteorological element data to obtain the characteristic map of meteorological elements; the formula is as follows: ; Among them, f is the original meteorological element data, is the average value of meteorological element data, is the standard deviation of meteorological element data.
4. The method for short-term precipitation forecasting based on TimeUnet and phenological knowledge according to claim 1, characterized in that: In step (3), the TimeUnet network includes two modules: MFCM-Unet and TimesNet, and includes the following steps: (31) Inputting the meteorological element feature map obtained in step (2) into the MFCM-Unet network to extract meteorological information of different spatial scales; the MFCM-Unet network includes an encoder, a channel attention mechanism based on multiple features, and a decoder; wherein the encoder is used to extract features of different spatial scales; the channel attention mechanism based on multiple features is used to give higher weights to important forecast features to highlight important information; the decoder is used to reconstruct and generate the final forecast; at the same time, the bird sound feature map TimesNet obtained in step (2) captures abnormal bird sound information of different time scales; including: signal reading, signal decomposition and construction of a two-dimensional signal field, and signal reconstruction; (32) Design a customized loss function, the formula is as follows: ; in, , , , They represent the root mean square error of light rain, moderate rain, heavy rain and rainstorm respectively, , , , Respectively , , , The weight of is as follows: ; in, represents the weight of the i-th precipitation, Represents the number of samples of the i-th type of precipitation; precipitation includes: light rain, moderate rain, heavy rain, and rainstorm.
5. A short-term precipitation forecast system based on TimeUnet and phenological knowledge, characterized in that: include: Collection module: used to collect meteorological element observation data and bird sound abnormal information samples, and pre-process the data; Dataset module: used to extract features from preprocessed meteorological element observation data and bird sound abnormal information samples to construct feature datasets; TimeUnet module: used to build the TimeUnet deep learning model that combines TimesNet and Unet and design a customized loss function; Training module: used to train the TimeUnet model and adjust the hyperparameters in the model to obtain the optimal model; Forecast module: used to generate short-term precipitation forecast products based on real-time meteorological element observation data and bird sound abnormal information samples.
6. The short-term precipitation forecast system based on TimeUnet and phenological knowledge according to claim 5 is characterized in that: In the acquisition module, meteorological element observation data include precipitation data, pressure field data, humidity data and wind field data at the current time and in the past 2 hours; bird sound abnormal information samples include information at the current time and in the past 2 hours; preprocessing is as follows: interpolate the station data into uniform grid data based on inverse distance weighted interpolation; remove outliers in the grid data; the calculation formula of inverse distance weighted interpolation is as follows: ; Where n represents the total number of sites around the target grid point, represents the observation data of the ith station, represents the distance from the i-th station to the target grid point, Represents the interpolated grid data.
7. The short-term precipitation forecast system based on TimeUnet and phenological knowledge according to claim 5 is characterized in that: The dataset module includes the following steps: (21) Based on short-time Fourier transform, the bird sound characteristic spectrum containing time-frequency domain characteristic information is obtained; the formula is as follows: ; in, is the extracted bird sound feature map. is the original sound spectrum at time T, is the analysis window function, is a complex function; (22) De-dimensionalize the meteorological element data to obtain the characteristic map of meteorological elements; the formula is as follows: ; Among them, f is the original meteorological element data, is the average value of meteorological element data, is the standard deviation of meteorological element data.
8. The short-term precipitation forecasting system based on TimeUnet and phenological knowledge according to claim 5 is characterized in that: In the TimeUnet module, the TimeUnet network includes two modules, MFCM-Unet and TimesNet, including the following steps: (31) Inputting the meteorological element feature map obtained in step (2) into the MFCM-Unet network to extract meteorological information of different spatial scales; the MFCM-Unet network includes an encoder, a channel attention mechanism based on multiple features, and a decoder; wherein the encoder is used to extract features of different spatial scales; the channel attention mechanism based on multiple features is used to give higher weights to important forecast features to highlight important information; the decoder is used to reconstruct and generate the final forecast; at the same time, the bird sound feature map TimesNet obtained in step (2) captures abnormal bird sound information of different time scales; including: signal reading, signal decomposition and construction of a two-dimensional signal field, and signal reconstruction; (32) Design a customized loss function, the formula is as follows: ; in, , , , They represent the root mean square error of light rain, moderate rain, heavy rain and rainstorm respectively, , , , Respectively , , , The weight of is as follows: ; in, represents the weight of the i-th precipitation, Represents the number of samples of the i-th type of precipitation; precipitation includes: light rain, moderate rain, heavy rain, and rainstorm.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into a processor, a method for forecasting short-term precipitation based on TimeUnet and phenological knowledge according to any one of claims 1 to 4 is implemented.
10. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, a method for forecasting short-term precipitation based on TimeUnet and phenological knowledge according to any one of claims 1 to 4 is implemented.
Citation Information
Patent Citations
Deep learning-based short temporary rainfall forecasting method and device, electronic equipment and storage medium
CN114492952A
Short-time heavy rainfall forecasting method fusing self-attention module and Unet model
CN117008217A
Short temporary rainfall ensemble forecasting and downscaling method based on multiple meteorological elements
CN117233870A
Short temporary rainfall forecasting method based on improved VIT neural network
CN118033590A
Sub-season rainfall forecasting method based on LA-UNET-LSTM
CN118277767A
Cited By
Multi-source meteorological element rainfall prediction method based on spatio-temporal information conversion equation
CN120276074A
