A method for constructing a short-term precipitation forecast model based on PSD-AUnet

Abstract

The present invention discloses a method for constructing a short-term precipitation forecast model based on PSD-AUnet, which relates to the technical field of meteorological forecasting. Data preparation: collecting CLDAS precipitation data of the past 3 hours; data preprocessing: performing time alignment, normalization and data partitioning on the original data; model construction: building a SmaAt-Unet model including an encoder, a decoder and an attention mechanism; model optimization: integrating a PSD module on the basis of the SmaAt-Unet model to form a PSD-AUnet model; model training: training the PSD-AUnet model using training set data, and updating the model parameters by minimizing the loss function; model prediction: evaluating the model using test set data, and outputting the precipitation prediction result for the next 2 hours. The present invention significantly improves the accuracy of precipitation prediction by introducing an attention mechanism and a PSD module. The method of the present invention has high prediction accuracy and computational efficiency, and can be widely used in the fields of meteorological forecasting, disaster prevention and mitigation, and agricultural production.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and in particular to a method for constructing a short-term precipitation forecasting model based on PSD-AUnet. Background Art

[0002] At present, precipitation prediction is one of the important tasks in meteorological forecasting. Especially when dealing with extreme weather such as heavy precipitation and rainstorms, accurate precipitation prediction can provide important decision-making basis for disaster prevention and mitigation. With the development of deep learning technology, precipitation prediction methods based on convolutional neural networks (CNN) have received widespread attention. In particular, the SmaAt-Unet model has achieved certain results in the field of precipitation prediction due to its lightweight structure and good feature extraction capabilities.

[0003] The SmaAt-Unet model is an improved Unet model. By adopting deep separable convolution and attention mechanism, it can effectively reduce the complexity of the model and enhance the focus on important features, thus achieving good results in precipitation prediction. However, the SmaAt-Unet model also has some shortcomings in precipitation prediction. Specifically, the SmaAt-Unet model still has room for improvement in the accuracy of precipitation prediction under complex weather conditions. In particular, when dealing with high-frequency details and complex spatial relationships, its feature extraction ability is limited, resulting in deviations in the prediction results in fine structures and details.

[0004] In order to further improve the accuracy of precipitation prediction, the present invention proposes a method for constructing a short-term precipitation forecast model based on PSD-AUnet. PSD-AUnet integrates a power spectral density (PSD) module based on SmaAt-Unet, and enhances the feature extraction capability of the model by introducing frequency domain information, effectively making up for the shortcomings of the SmaAt-Unet model in processing high-frequency information, thereby significantly improving the accuracy of precipitation prediction. The PSD module calculates the power spectral density difference between the model prediction and the real data, so that the model can better learn the frequency domain characteristics of precipitation data, thereby achieving better results in capturing fine structure and high-frequency information. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method for constructing a short-term precipitation forecast model based on PSD-AUnet, which aims to improve the accuracy of precipitation prediction by introducing an attention mechanism and a power spectral density (PSD) module.

[0006] To solve the above technical problems, the present invention provides a technical solution: a method for constructing a short-term precipitation forecast model based on PSD-AUnet, comprising the following steps:

[0007] S1. Data preparation: Collect CLDAS precipitation data for the past 3 hours;

[0008] S2, data preprocessing: time alignment, normalization and data partitioning of raw data;

[0009] S3, model construction: build the SmaAt-Unet model;

[0010] S4, model optimization: Integrate the PSD module based on the SmaAt-Unet model to form the PSD-AUnet model;

[0011] S5, model training: Use the training set data to train the PSD-AUnet model and update the model parameters by minimizing the loss function;

[0012] S6. Model prediction: Use the test set data to evaluate the model and output the precipitation prediction results for the next 2 hours.

[0013] Furthermore, the CLDAS precipitation data records precipitation conditions with high temporal and spatial resolution. The CLDAS precipitation data forms a time series at hourly intervals. The data for each time step consists of a multi-channel precipitation image, where each channel represents the precipitation intensity of a time step.

[0014] Furthermore, the detailed steps of the data preprocessing are as follows:

[0015] (1) Time alignment: align data at different time steps to ensure that the input data have the same time interval;

[0016] (2) Normalization: Normalize the precipitation intensity data so that the data range is between [0, 1] to facilitate model training;

[0017] (3) Data partitioning: Divide the data into training set, validation set, and test set.

[0018] Furthermore, the SmaAt-Unet model includes an encoder, a decoder and an attention mechanism.

[0019] Furthermore, the encoder in the SmaAt-Unet model is composed of multiple convolutional layers and pooling layers for extracting features of input data; the decoder is composed of multiple deconvolutional layers for mapping the features back to the spatial resolution of the input data.

[0020] Furthermore, the data processing content of the PSD module is as follows:

[0021] (1) Fourier transform: Perform two-dimensional fast Fourier transform on the model-predicted precipitation results and the actual high-resolution precipitation data to obtain their frequency domain representation;

[0022] (2) Calculate the power spectral density: square the amplitude of the Fourier transform result to obtain the power spectral density of the model prediction and the real data;

[0023] (3) Normalization: To avoid numerical instability and zero division errors, a smoothing term ε is added when calculating the sum of the power spectral density, and then the power spectral density is normalized;

[0024] (4) Calculate the power spectral density difference: Use the Kullback-Leibler divergence to measure the difference between the power spectral density of the model prediction and the real data.

[0025] Furthermore, during the training process of the PSD-AUnet model, mean square error is used as the loss function and a stochastic gradient descent optimizer is used for optimization.

[0026] Furthermore, the PSD-AUnet model improves the model's attention to important features through weighted feature maps, and further enhances the model's feature extraction capability by using frequency domain information, thereby improving the accuracy of precipitation prediction.

[0027] The advantages of the present invention compared with the prior art are:

[0028] 1. The present invention significantly improves the accuracy of precipitation prediction by introducing the attention mechanism and PSD module.

[0029] 2. The present invention utilizes the high temporal and spatial resolution of CLDAS precipitation data to capture more detailed precipitation characteristics and provide richer information for precipitation prediction.

[0030] 3. The model training process of the present invention is simple and efficient, suitable for processing and analyzing large-scale precipitation data, and has broad application prospects.

[0031] 4. The method of the present invention has high prediction accuracy and calculation efficiency, and can be widely used in fields such as weather forecasting, disaster prevention and mitigation, and agricultural production. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a real-time precipitation map predicting the next 0-1 hour.

[0033] Figure 2 It is a precipitation forecast map for the present invention predicting the next 0-1 hour.

[0034] Figure 3 This is the precipitation forecast map predicted by the SmaAt-Unet model for the next 0-1 hour.

[0035] Figure 4 It is a real-time precipitation map predicting the next 1-2 hours.

[0036] Figure 5 It is a precipitation forecast map for predicting the next 1-2 hours according to the present invention.

[0037] Figure 6 This is the precipitation forecast map predicted by the SmaAt-Unet model for the next 1-2 hours. DETAILED DESCRIPTION

[0038] The following is a further detailed description of a method for constructing a short-term precipitation forecast model based on PSD-AUnet according to the present invention in conjunction with the accompanying drawings.

[0039] Combined with Figure 1-6 The specific implementation process of the method for constructing a short-term precipitation forecast model based on PSD-AUnet of the present invention is as follows:

[0040] 1. Data preparation

[0041] The precipitation data used in the present invention is the precipitation intensity data of the past 3 hours provided by CLDAS. CLDAS data records the precipitation conditions in the country with high temporal and spatial resolution. The data is timed every hour to form a time series. The data of each time step consists of a multi-channel precipitation image, where each channel represents the precipitation intensity of a time step.

[0042] 2. Data Preprocessing

[0043] In order to input data into the model, the raw precipitation data needs to be preprocessed first. The preprocessing steps include:

[0044] (1) Time alignment: Align data at different time steps to ensure that the input data have the same time interval.

[0045] (2) Normalization: The precipitation intensity data is normalized so that the data range is between [0, 1] to facilitate model training.

[0046] (3) Data partitioning: Divide the data into training set, validation set, and test set.

[0047] The specific data division ratio can be:

[0048] Training Set: 80% of the dataset, used for model training.

[0049] Validation Set: 10% of the dataset, used to evaluate the performance of the model during training and help select the optimal hyperparameters.

[0050] Test Set: 10% of the dataset, used to independently evaluate the final trained model and test the performance of the model on unseen data.

[0051] When dividing data, the randomness of the data should be ensured. At the same time, the characteristics of time series should be taken into consideration, and the independence of different data sets should be guaranteed as much as possible to avoid information leakage.

[0052] 3. Model Structure

[0053] The structure of the SmaAt-Unet model includes an encoder, a decoder, and an attention mechanism:

[0054] (1) Encoder: It consists of multiple convolutional layers and pooling layers, which are used to extract features of input data.

[0055] (2) Decoder: It consists of multiple deconvolutional layers and is used to map features back to the spatial resolution of the input data.

[0056] (3) Attention mechanism: An attention mechanism is introduced between the encoder and decoder to increase the model's attention to important features by weighting the feature map.

[0057] The PSD module is integrated on the basis of the SmaAt-Unet model to form the PSD-AUnet model; the PSD-AUnet model not only introduces the attention mechanism between the encoder and the decoder to improve the model's attention to important features through weighted feature maps, but also integrates the PSD module (power spectral density module) to further enhance the feature extraction ability of the model by using frequency domain information, thereby improving the accuracy of precipitation prediction.

[0058] The data processing content of the PSD module includes:

[0059] (1) Fourier transform: The model-predicted precipitation results (denoted as PRED) and the actual precipitation data (denoted as LABEL) are respectively subjected to two-dimensional fast Fourier transform to obtain their frequency domain representation.

[0060] (2) Calculate the power spectral density (PSD): Take the square of the amplitude of the Fourier transform result to obtain the power spectral density of the model prediction and the real data.

[0061] (3) Normalization: To avoid numerical instability and zero division errors, a small smoothing term ε (e.g., ε = 1e-9) is added when calculating the sum of the power spectral density, and then the PSD is normalized.

[0062] (4) Calculate the power spectral density difference: Use Kullback-Leibler divergence (KL divergence) to measure the difference between the power spectral density of the model prediction and the real data.

[0063] 4. Model Training

[0064] (1) Data preparation: First, we collect and preprocess precipitation data. The input data is the precipitation data for the past 3 hours, organized in the form of time series. The output label is the precipitation data for the next 2 hours.

[0065] (2) Model initialization: Build the PSD-AUnet model, integrate the PSD module based on SmaAt-Unet, and define the model structure and the required loss function. The loss function includes mean square error (MSE) and PSD loss, which are used to measure the prediction error of the model in the time domain and frequency domain.

[0066] (3) Training process:

[0067] The model is trained using the Stochastic Gradient Descent (SGD) optimizer, and the learning rate can be adjusted based on the performance of the model.

[0068] In each iteration, the model passes the input data (precipitation data of the past 3 hours) to PSD-AUnet to obtain the precipitation forecast results for the next 2 hours.

[0069] Calculate the loss function, including MSE loss in the time domain and PSD loss in the frequency domain.

[0070] The parameters of the model are updated by minimizing the comprehensive loss, so that the model can better fit the data in both time domain and frequency domain.

[0071] (4) Verification and adjustment:

[0072] After each training cycle, the validation set data is used to evaluate the predictive performance of the model.

[0073] According to the verification results, the model's hyperparameters, such as learning rate, batch size, etc., are adjusted to further improve the model's generalization ability.

[0074] (5) Training stop condition: The training process continues until the validation set loss no longer decreases significantly, or the preset number of training rounds is reached.

[0075] (6) Model evaluation: Use the test set data to perform a final evaluation of the trained model, compare the model's prediction results with the real data, and evaluate the model's applicability and accuracy in different scenarios.

[0076] 5. Model prediction

[0077] After training, the model is evaluated using the test set data to obtain the precipitation forecast results for the next 2 hours. The prediction accuracy of the model is evaluated by comparing it with the actual precipitation data.

[0078] In order to verify the effectiveness of the precipitation extrapolation model based on SmaAt-Unet and PSD modules proposed in the present invention, the precipitation data of a certain area was used as the experimental object, and the prediction performance of the model was comprehensively evaluated. The experimental results show that the model has achieved high accuracy in the precipitation prediction of the next 2 hours. Compared with the traditional SmaAt-Unet model, the model of the present invention has significantly improved in accuracy (TS score), and can provide more timely and accurate precipitation prediction results. The evaluation results are shown in Table 1.

[0079] Table 1: Comparative evaluation of precipitation prediction between SmaAt-Unet model and PSD-AUnet model

[0080] Model 1 hour TS 2 hours TS SmaAt-Unet 0.1057 0.0239 PSD-AUnet 0.1599 0.0548

[0081] In order to deeply evaluate the performance of the precipitation extrapolation model based on SmaAt-Unet plus PSD module proposed in this paper, a detailed comparative analysis of a precipitation process was carried out. Using three key indicators, TS (Threat Score), POD (Probability of Detection) and FAR (False Alarm Rate), the performance of the traditional SmaAt-Unet model and the PSD-AUnet model of this invention in 1-hour and 2-hour forecasts were compared.

[0082] 1-hour forecast comprehensive analysis:

[0083] Combined with Figure 1-3 , in the 1-hour precipitation forecast, the PSD-AUnet model performs significantly better than SmaAt-Unet. Specifically, the TS value of PSD-AUnet is 66.7%, which is 7.5 percentage points higher than SmaAt-Unet's 59.2%, indicating that the model has significantly improved in accurately predicting precipitation areas. At the same time, the POD value of PSD-AUnet reaches 72.3%, which is higher than SmaAt-Unet's 63.1%, which means that the new model detects actual precipitation events more effectively. Although the FAR value of PSD-AUnet is slightly higher at 10.3%, which is 0.9 percentage points higher than SmaAt-Unet's 9.4%, the overall false alarm rate is still at a low level and within an acceptable range. Overall, PSD-AUnet significantly improves the accuracy and detection rate in 1-hour forecasts, and the slightly increased false alarm rate does not weaken its overall advantage.

[0084] 2-hour forecast comprehensive analysis:

[0085] Combined with Figure 4-6, in the 2-hour precipitation forecast, the performance of the PSD-AUnet model has improved more significantly. Its TS value reached 49.6%, an increase of 28.1 percentage points compared to SmaAt-Unet's 21.5%, showing that PSD-AUnet can still maintain a high prediction accuracy under a longer prediction time. In terms of POD indicators, PSD-AUnet reached 51.3%, much higher than SmaAt-Unet's 22.6%, indicating that the new model is more sensitive in capturing precipitation events. In addition, the FAR value of PSD-AUnet is 6.3%, which is significantly lower than SmaAt-Unet's 18.4%, a decrease of 12.1 percentage points, which shows that the new model effectively reduces the false alarm rate in long-term prediction and improves the reliability of the prediction results. Overall, PSD-AUnet not only improves the accuracy and detection rate in the 2-hour prediction, but also significantly reduces the false alarm rate, reflecting excellent model performance.

[0086] The precipitation extrapolation method based on PSD-AUnet and CLDAS precipitation data proposed in this paper successfully predicts the precipitation in the next 2 hours by using the CLDAS precipitation data of the past 3 hours. This method has high prediction accuracy and computational efficiency and can be widely used in meteorological forecasting, disaster prevention and mitigation, agricultural production and other fields.

[0087] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A method for constructing a short-term precipitation forecast model based on PSD-AUnet, characterized in that: The following steps are included: S1. Data preparation: Collect CLDAS precipitation data for the past 3 hours; S2, data preprocessing: time alignment, normalization and data partitioning of raw data; S3, model construction: build the SmaAt-Unet model; The SmaAt-Unet model includes an encoder, a decoder, and an attention mechanism; The encoder in the SmaAt-Unet model consists of multiple convolutional layers and pooling layers for extracting features of input data; the decoder consists of multiple deconvolutional layers for mapping features back to the spatial resolution of the input data; S4, model optimization: Integrate the PSD module based on the SmaAt-Unet model to form the PSD-AUnet model; The PSD-AUnet model improves the model's attention to important features through weighted feature maps, and further enhances the model's feature extraction capability using frequency domain information, thereby improving the accuracy of precipitation prediction; The data processing content of the PSD module is as follows: (1) Fourier transform: Perform two-dimensional fast Fourier transform on the model-predicted precipitation results and the actual high-resolution precipitation data to obtain their frequency domain representation; (2) Calculate the power spectral density: square the amplitude of the Fourier transform result to obtain the power spectral density of the model prediction and the real data; (3) Normalization: To avoid numerical instability and zero division errors, a smoothing term ε is added when calculating the sum of the power spectral density, and then the power spectral density is normalized; (4) Calculate the power spectral density difference: Use the Kullback-Leibler divergence to measure the difference between the power spectral density of the model prediction and the real data; S5, model training: Use the training set data to train the PSD-AUnet model and update the model parameters by minimizing the loss function; S6. Model prediction: Use the test set data to evaluate the model and output the precipitation prediction results for the next 2 hours.

2. The method for constructing a short-term precipitation forecast model based on PSD-AUnet according to claim 1, characterized in that: The CLDAS precipitation data records precipitation conditions with temporal and spatial resolution. The CLDAS precipitation data forms a time series at hourly intervals. The data of each time step consists of multi-channel precipitation images, where each channel represents the precipitation intensity of a time step.

3. The method for constructing a short-term precipitation forecast model based on PSD-AUnet according to claim 2, characterized in that: The detailed steps of data preprocessing are as follows: (1) Time alignment: align data at different time steps to ensure that the input data have the same time interval; (2) Normalization: Normalize the precipitation intensity data so that the data range is between [0, 1] to facilitate model training; (3) Data partitioning: Divide the data into training set, validation set, and test set.

4. The method for constructing a short-term precipitation forecast model based on PSD-AUnet according to claim 3, characterized in that: During the training process of the PSD-AUnet model, mean square error is used as the loss function and the stochastic gradient descent optimizer is used for optimization.

Images