A method and system for forecasting short-term rainfall and floods based on deep learning

Abstract

The present invention relates to the field of hydrological forecasting technology, and in particular to a forecasting method and system for short-term rainfall and floods based on deep learning, comprising the following steps: constructing a radar echo extension prediction model and a rainfall and flood forecasting model based on deep learning, and respectively using a first data set and a second data set to train and test the two constructed models; inputting rainfall radar echo data of a certain period of time in a target basin into the radar echo extension prediction model, and forecasting the changing trend of rainfall radar echoes in a future period of time; inputting the changing trend of rainfall radar echoes in a future period of time into the rainfall and flood forecasting model, and forecasting the rainfall and runoff of the target basin in a future period of time. The present invention can make more accurate short-term forecasts of the rainfall and runoff of the target basin, and effectively improve the forecasting ability and decision-making level of regional water resources management.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrological forecasting, and more specifically to a forecasting method and system for short-term rainfall and flood based on deep learning. Background Art

[0002] With the intensification of global climate change and human activities, extreme weather events (such as heavy rains, typhoons, etc.) have occurred frequently, leading to the frequent occurrence of floods and waterlogging disasters. Floods and waterlogging disasters not only bring huge losses to social and economic development, but also pose a serious threat to human life and safety. Therefore, how to improve the accuracy of flood forecasts and issue flood warnings in advance has become an important topic in hydrological research. Traditional hydrological forecasts rely on historical data and statistical models, but this method is often insufficient in predicting sudden, short-term heavy rainfall events and is difficult to provide timely and accurate flood forecasts.

[0003] Radar technology plays a vital role in weather forecasting. Radar echoes can be used to monitor precipitation, but how to convert radar echoes into quantitative precipitation forecast data has always been an important topic in hydrological and meteorological research. Traditional radar echo-precipitation relationship models are usually established based on statistical methods, but the defect is that they cannot effectively capture the characteristics of different rainfall types (such as convective precipitation and stratiform precipitation), resulting in large forecast errors.

[0004] Short-term quantitative precipitation forecast is a key component of disaster management and hydrological forecasting. In order to improve the accuracy of short-term precipitation forecast, researchers began to use rainfall radar echo data to extrapolate future short-term precipitation conditions. This method attempts to forecast precipitation in a short period of time based on the movement and evolution of radar echoes. However, this extrapolation method is still limited by the radar echo-precipitation relationship model.

[0005] Despite the continuous technological progress in short-term rainfall forecasting and hydrological models, forecast accuracy and timeliness still face some major challenges. In particular, the unpredictability of local heavy rainfall events has become a major obstacle to improving the accuracy of rainfall forecasts. Therefore, it is urgent to improve the spatial resolution and temporal accuracy of rainfall forecasts in the future in order to better capture these extreme weather events that occur in a short period of time. In addition, the high dependence of the coupled forecasting system on high-quality real-time observation data also brings problems with data quality and accessibility. In some areas, the lack of observation sites may lead to uncertainty in the initial conditions of the model, thereby affecting the reliability of the forecast results. Finally, the computational complexity of distributed hydrological models and high-resolution meteorological models is high, which means that large-scale computing resources are required to support the operation of real-time forecasting systems. Therefore, solving these challenges will be the key to improving the technology of coupled forecasting of short-term rainfall and hydrology. Summary of the invention

[0006] In view of this, the present invention provides a method and system for forecasting short-term rainfall and floods based on deep learning, which can make more accurate short-term forecasts of precipitation and runoff in the target basin.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a method for forecasting short-term rainfall and floods based on deep learning, comprising the following steps:

[0009] Construct a radar echo extension prediction model and a rainfall and flood forecasting model based on deep learning, and use the first data set and the second data set to train and test the two constructed models respectively;

[0010] Inputting rainfall radar echo data of a certain period of time in the target watershed into the radar echo extension prediction model to predict the changing trend of rainfall radar echo in the future period of time;

[0011] The changing trend of rainfall radar echoes in the future period is input into the rainfall and flood forecasting model to forecast the rainfall and runoff of the target basin in the future period.

[0012] Furthermore, the process of constructing the first data set includes:

[0013] Acquire historical rainfall radar echo data, and pre-process the data to obtain the first data set;

[0014] The preprocessing methods for historical rainfall radar echo data include: data quality control, data calibration, data interpolation and resampling, noise reduction and smoothing, and normalization.

[0015] Furthermore, the data quality control method includes: removing ground clutter, processing the influence of atmospheric refraction, and identifying and removing non-meteorological echoes.

[0016] Furthermore, the data interpolation and resampling method includes:

[0017] The Kriging interpolation method is used to spatially interpolate rainfall radar echo data;

[0018] The linear interpolation method is used to perform time interpolation on rainfall radar echo data.

[0019] Furthermore, the data calibration includes: at any moment, comparing the rainfall radar echo data with the actual precipitation measured by the ground rain gauge, and adjusting the rainfall radar echo data to match the reading of the ground rain gauge.

[0020] Furthermore, the radar echo extension prediction model includes an input layer, a CNN layer, an LSTM layer, an attention mechanism layer and an output layer;

[0021] The input layer is used to receive rainfall radar echo data;

[0022] The CNN layer is used to extract the spatial features of the rainfall radar echo data as the input of the LSTM layer;

[0023] The LSTM layer is used to recursively process the spatial features of the rainfall radar echo data step by time, and output a feature vector containing time series information;

[0024] The attention mechanism layer is used to perform weighted aggregation on the feature vector output by the LSTM layer to obtain a context feature vector;

[0025] The output layer is used to convert the context feature vector output by the attention mechanism layer into a final prediction result to obtain the change trend of the rainfall radar echo in the future period of time.

[0026] Furthermore, the construction process of the second data set includes:

[0027] Convert the latitude and longitude coordinates of the target watershed to the same coordinate system as the rainfall radar echo data;

[0028] Extract the past rainfall radar echo data and rainfall and runoff measured data of the corresponding station location;

[0029] The rainfall radar echo data, rainfall and runoff measured data of each station in the target basin are used as the second data set and split into a training set and a test set.

[0030] Furthermore, the rainfall and flood forecasting model is constructed based on the LSTM network and attention mechanism.

[0031] In a second aspect, the present invention provides a forecasting system for short-term rainfall and floods based on deep learning, comprising:

[0032] An acquisition module is used to obtain rainfall radar echo data of the target basin in the current period;

[0033] The first prediction module is used to input the acquired rainfall radar echo data into the radar echo extension prediction model as described above to obtain the change trend of rainfall radar echo in the future period;

[0034] The second prediction module is used to input the changing trend of rainfall radar echoes in the future period into the rainfall and flood forecasting model as described above, and obtain the rainfall and runoff of the target basin in the future period.

[0035] It can be seen from the above technical solutions that, compared with the prior art, the present invention has the following beneficial effects:

[0036] The present invention first predicts the rainfall radar data for the future period through a radar echo extrapolation prediction model based on deep learning, and then analyzes the rainfall radar data for the future period based on the rainfall and flood forecasting model to obtain the rainfall and runoff of the target river basin in the future period. During the entire prediction process, the forecast results of radar echo extrapolation are combined with ground telemetry data to achieve coordinated coupling of the flood forecasting technology system, thereby being able to more accurately predict precipitation and runoff. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0038] Figure 1 A flowchart of a method for forecasting short-term rainfall and floods based on deep learning provided by the present invention;

[0039] Figure 2 It is the extension of rainfall radar echo data;

[0040] Figure 3 Schematic diagram of the training error change process of the RNN model;

[0041] Figure 4 Schematic diagram of the training error change process of the CNN model;

[0042] Figure 5 This is a schematic diagram of the training error change process of the LSTM model;

[0043] Figure 6 This is a schematic diagram of the training error change process of the CNN-LSTM model;

[0044] Figure 7 This is a schematic diagram of the training error change process of the Transformer model;

[0045] Figure 8 This is a schematic diagram of the training error change process of the CNN-LSTM-Attention model;

[0046] Fig. 9 This is the mean square error distribution diagram of the measured and simulated water level at the reservoir hydrological station. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] like Figure 1 As shown, the embodiment of the present invention discloses a method for forecasting short-term rainfall and flood based on deep learning, comprising the following steps:

[0049] Construct a radar echo extension prediction model and a rainfall and flood forecasting model based on deep learning, and use the first data set and the second data set to train and test the two constructed models respectively;

[0050] Input the rainfall radar echo data of a certain period of time in the target basin into the radar echo extension prediction model to predict the changing trend of rainfall radar echo in the future period of time;

[0051] The changing trend of rainfall radar echoes in the future period is input into the rainfall and flood forecasting model to predict the rainfall and runoff in the target basin in the future period.

[0052] The present invention has three main objectives: first, to establish a radar echo extension prediction model based on deep learning for the extension of rainfall radar data; second, to couple the rainfall radar data with the precipitation and runoff data measured at ground stations to establish a rainfall and flood forecasting model based on rainfall radar data; third, to introduce the extended rainfall radar data into the rainfall and flood forecasting model to predict the short-term rainfall and flood volume in the basin in real time.

[0053] The above steps are further described below.

[0054] S1. Construction, training and testing of radar echo extension prediction model based on deep learning, including:

[0055] S11, obtaining historical rainfall radar echo data, and preprocessing the data as a first data set;

[0056] In order to ensure the quality and accuracy of the data and provide a reliable basis for subsequent meteorological analysis and short-term precipitation forecast, the historical rainfall radar echo data needs to be preprocessed. The specific preprocessing methods mainly include: data quality control, data calibration, data interpolation and resampling, noise reduction and smoothing, and normalization.

[0057] (1) Data quality control: Data quality control is the first step in preprocessing, which aims to identify and correct errors or anomalies in the data, including:

[0058] Removing ground clutter: Signals reflected by ground objects (such as buildings and mountains) may be captured by radar and need to be identified and removed. This patent uses the characteristics of radar echoes in spatial distribution to design spatial filters. For example, ground clutter usually appears as a fixed, spatially uniform pattern in radar images, while weather echoes are different. Based on these characteristics, algorithms can be designed to automatically remove these ground clutter.

[0059] Dealing with atmospheric refraction effects: Atmospheric conditions (such as temperature and humidity) affect the propagation path of microwaves and may cause errors. The refractive index distribution of the atmosphere is estimated and the radar data is corrected accordingly. This usually requires additional meteorological observation data, such as radiosonde or ground observation station data.

[0060] Identify and remove non-weather echoes: Non-weather echoes, such as those caused by flocks of birds, aircraft, or insect swarms, also need to be removed from weather radar data. Weather and non-weather echoes are distinguished by analyzing the characteristics of radar echoes, such as speed, direction, shape, and intensity. For example, the echoes produced by flocks of birds or aircraft may be significantly different from weather echoes in speed or direction.

[0061] (2) Data calibration: Data calibration is the process of ensuring that radar measurements match actual precipitation, including:

[0062] Calibration of radar reflectivity factor: Radar systems are usually equipped with internal calibration devices that periodically check and adjust the radar's transmit power and receive sensitivity. This calibration ensures a consistent relationship between the radar's measurements and the actual precipitation intensity. In addition, external calibration sources such as nearby targets of known reflectivity (such as corner reflectors) are used to calibrate the radar. By comparing the radar's measurements of these known targets with their actual reflectivity, the radar's reflectivity factor can be calibrated. In dual-polarization radars, dual-polarization parameters such as differential reflectivity (ZDR) can be used to compare and calibrate the radar reflectivity factor to improve the accuracy of the measurement.

[0063] Comparison with Ground Rain Gauge Data: Ground rain gauges provide actual precipitation measurements, and comparison with radar measurements can further validate and calibrate radar data. Patented utilization adjusts radar data to match rain gauge readings by comparing radar-measured precipitation areas with actual measurements from ground rain gauges. This includes spatially and temporally matching radar data to ensure consistency with rain gauge data. Statistical regression methods are used to analyze the relationship between radar reflectivity factors and ground rain gauge data. By establishing a statistical relationship between the two, radar measurements can be corrected to more closely match actual precipitation. Advanced data fusion techniques are used to combine radar and rain gauge data to improve the accuracy of precipitation estimates. This fusion takes into account the strengths and limitations of both measurement methods to produce more accurate precipitation estimates.

[0064] (3) Data interpolation and resampling: Due to the spatial inhomogeneity of radar data, interpolation and resampling may be required to obtain a more consistent data distribution, including:

[0065] Spatial interpolation: Spatial interpolation is used to fill in missing points in radar data in space, or to resample the data to obtain a more uniform spatial distribution. This patent uses Kriging interpolation, which is a geostatistical interpolation method that considers not only the distance between data points, but also the spatial relationship between them. This method can be effectively used for spatial interpolation of radar data, especially when dealing with meteorological data with spatial correlation.

[0066] Time interpolation: For continuous radar scans, time interpolation is performed to generate stable time series data. The patent uses linear interpolation, which is a basic interpolation method that estimates the value of the intermediate time point by linearly connecting the data of two time points. This method is simple and computationally efficient, and is suitable for radar data series with short time intervals and non-dramatic changes.

[0067] (4) Noise reduction and smoothing. Radar echo data may contain noise, so noise reduction and smoothing are required, as shown below:

[0068] Filtering algorithm: This patent uses a standard Gaussian filter, applying a Gaussian kernel (a bell-shaped weighted average window) to each data point. This weighted average takes into account the values ​​of the neighborhood around the data point, where the center point has the highest weight and the weight gradually decreases with increasing distance. This process can effectively smooth the data and reduce noise.

[0069]

[0070] Where x and y are the pixel coordinates. is the standard deviation of the Gaussian function, are Gaussian weights.

[0071] Smoothing: Smoothing the data to remove small-scale fluctuations in the data. This patent uses a moving average filter, which smoothes the data by averaging a certain number of points in the data point and its neighborhood. Moving average filtering can effectively remove random noise, especially in time series data.

[0072] (5) Data normalization. Normalization technology is an important part of data preprocessing, and is widely used in fields such as machine learning, signal processing, and image processing. The purpose of normalization is to adjust the scale of the data so that it falls within a specific range (usually 0 to 1 or -1 to 1), thereby making the processing process more efficient and stable. This patent uses Min-Max Normalization, which is the most common form of normalization, to scale all eigenvalues ​​to between 0 and 1. The calculation formula is as follows:

[0073]

[0074] In the formula, is the normalized value, and Separate arrays The maximum and minimum values ​​in .

[0075] S12. Next, based on the data from the automatic rain gauges in the small reservoir basin and the preprocessed radar echo data, a CNN-LSTM (convolutional neural network-long short-term memory network) model based on the attention mechanism and deep learning optimization was established as a radar echo extension prediction model to predict the changes in the radar cloud map in the next period for subsequent flood forecasting.

[0076] (1) The functions of each module of the radar echo extension prediction model are introduced as follows:

[0077] Attention mechanism: Attention mechanism is a technique inspired by human visual attention and is widely used in deep learning. It enables the model to focus on key parts of the input data rather than all the information when processing information. The core concepts include weight assignment and the creation of context vectors. Under this mechanism, the model determines the focus by assigning different weights to different parts of the input data, and then combines these weighted input parts into a comprehensive representation, namely the context vector, to guide the model's decision-making process.

[0078] CNN-LSTM model: The CNN-LSTM model combines a convolutional neural network (CNN) and a long short-term memory network (LSTM). CNN is good at extracting spatial features from images and identifying local patterns through convolutional layers. It is often used in image processing tasks such as image classification and object detection. On the other hand, LSTM is a special type of recurrent neural network that specializes in processing sequence data and effectively processes and predicts temporal dependencies in sequences, such as used in language models and time series analysis.

[0079] Deep learning optimization parameters: The present invention uses the Adam method for parameter optimization. The core idea of ​​the Adam method is to adjust the learning rate of each parameter by calculating the first-order and second-order momentum of the gradient. By weighted averaging the gradient, the variance of the gradient update is reduced and the convergence is accelerated. Each parameter has an adaptive learning rate that can be automatically adjusted according to the change of the gradient, which is more effective in handling sparse gradients and dynamically changing objective functions.

[0080] Combined application: When the present invention combines the attention mechanism with the CNN-L / 63TM model, CNN is used to extract key features from complex input data (such as images). These features are then input as a sequence into the LSTM, which processes the time series dependencies. During this process, the attention mechanism plays a role when the LSTM processes the data at each time step, helping the model determine which part of the sequence should be focused on. For example, in the image description task, the model first uses CNN to analyze the image and extract features, and then LSTM generates descriptions based on these features. During this process, the attention mechanism helps the model focus on the relevant parts of the image when generating each word.

[0081] This combination enables the model to not only understand the content of the image or sequence data, but also focus on the most important parts of information more effectively when solving specific tasks. This hybrid approach is applicable to a variety of complex tasks such as image and video understanding, natural language processing, etc.

[0082] (2) Model architecture, the radar echo extension prediction model includes input layer, CNN layer, LSTM layer, attention mechanism layer and output layer.

[0083] The input layer is used to receive rainfall radar echo data and convert the data into a format suitable for model processing, such as a time series matrix.

[0084] The CNN layer is used to extract the spatial features of the rainfall radar echo data as the input of the LSTM layer; the CNN layer contains multiple convolutional layers and pooling layers. The convolution operation can learn the spatial or local features in the data, while the pooling layer is used to reduce the dimension and size of the features while retaining important features.

[0085] The convolution layer performs convolution operations on the input data by designing convolution kernels of appropriate sizes to abstractly express the original data. The feature map of the convolution layer can be expressed as: ,in is the convolution operation, W is the weight vector of the convolution kernel, b is the offset, and f is the activation function.

[0086] The pooling layer performs downsampling on the convolution output, retaining strong features and removing weak features, while reducing the number of parameters and preventing overfitting.

[0087] The LSTM layer is used to learn the long-term dependencies and contextual information of the input data. The LSTM network controls the flow of information through a gating mechanism (input gate, forget gate, output gate), thereby avoiding the problem of gradient vanishing or gradient exploding when processing long sequences. In the present invention, the LSTM layer receives the spatial features of the rainfall radar echo data extracted by the CNN layer, performs recursive processing time-step by time-step, and outputs a feature vector containing time series information.

[0088] The attention mechanism layer is used to perform weighted aggregation on the feature vectors output by the LSTM layer to obtain the context feature vector, thereby increasing the model's attention to important features. The attention mechanism layer can be designed and implemented differently according to different tasks and data sets.

[0089] In the attention-based CNN-LSTM model, the attention mechanism is usually combined with the CNN layer or the LSTM layer to extract significant fine-grained features and facilitate LSTM to capture temporal regularities more effectively.

[0090] Attention mechanisms can be classified into hard attention and soft attention. The hard attention mechanism selects the focused area as input, but it is difficult to train and has poor versatility. The soft attention mechanism weights all input features one by one, focusing on specific spaces and channels to extract significant fine-grained features of the time series. Mathematically, the soft attention mechanism can construct a context vector by calculating the weighted average of the input and the key.

[0091] The output layer is used to convert the context feature vector output by the attention mechanism layer into the final prediction result, and obtain the changing trend of rainfall radar echoes in the future. The output layer usually contains a fully connected layer (Dense layer), which is used to connect and process the feature vectors and generate the final prediction value through the activation function.

[0092] (3) Model parameter setting, where the CNN layer parameter settings are as follows:

[0093] Convolution kernel size (kernel_size): Choose a convolution kernel size of 3x3 or 5x5, depending on the characteristics of the input data and the scale of features you want to extract.

[0094] Number of convolution kernels (filters / channels): Based on the complexity of the task, 32 convolution kernels are selected. More convolution kernels can extract richer features, but will also increase the amount of calculation and model complexity.

[0095] Activation function: Use ReLU model to learn complex features.

[0096] Pooling layer: Average pooling is used to reduce the size of the feature map, reduce the amount of calculation, and retain important features.

[0097] The LSTM layer parameters are set as follows:

[0098] Hidden state size (hidden_size): Based on the complexity of the task and the size of the dataset, 64 hidden units are selected. More hidden units can capture more complex time series information, but will also increase the amount of computation and memory consumption.

[0099] Number of LSTM layers (num_layers): Select 2 layers of LSTM. The more layers there are, the more complex the time series information that the model can capture, but it will also increase the training time and the risk of overfitting.

[0100] Attention mechanism layer parameter settings:

[0101] Attention heads (multi-head attention): When using the multi-head attention mechanism, you need to set the number of heads. The more heads there are, the richer features the model can capture, but it will also increase the amount of calculation and model complexity.

[0102] Dimension scaling: When calculating the attention weights, the input dimensions need to be scaled to prevent the dot product result from being too large and causing the softmax function gradient to disappear.

[0103] Other parameter settings:

[0104] Learning rate (learning_rate): Based on the size of the dataset and training stability, select 0.001 as the learning rate. A learning rate that is too high may cause the model to fail to converge, while a learning rate that is too low will result in a slow training speed.

[0105] Batch size (batch_size): Select 8 as the batch size. A larger batch size will make the model training more stable, but it will also increase the GPU memory consumption and training time.

[0106] Iterations (num_epochs): Select 100 as the number of iterations. More iterations can improve model performance, but may also lead to overfitting.

[0107] Optimizer: Choosing to use the Adam optimizer usually has a faster convergence speed and better performance.

[0108] Loss function: Depending on the type of regression task, choose to use the mean square error loss function.

[0109] (4) Model training and testing.

[0110] The rainfall radar echo data in the first dataset is raster data with a temporal resolution of 5 minutes, a spatial resolution of 1 km*1 km, and a detection range of 100 km. 2 , the first data set is divided into a training set and a test set. When training the model, the input data is the processed rainfall radar echo data, and the output data is the rainfall radar echo data of the next period.

[0111] For example, if 16 radar data with 5-minute intervals are input, the 16*5=80 minutes are the data from 12:00 to 13:20, and the output data is the radar echo trend of the next 80 minutes, that is, the radar echo from 13:20 to 14:40. The predicted output becomes the extension of the rainfall radar echo data. Based on the future trend of the rainfall radar echo, the potential rainfall in the future can be calculated.

[0112] Afterwards, the model is validated using an independent test set to assess its accuracy and reliability. The model is then fine-tuned based on the validation results, including adjusting model parameters and selecting more appropriate features.

[0113] S2. Construction, training and testing of rainfall and flood forecasting models, including:

[0114] S21. Site positioning: To locate the latitude and longitude coordinates of each site in the target basin and the rainfall radar data in the polar coordinate system, the latitude and longitude coordinates must first be converted into the same coordinate system as the rainfall radar echo data (usually a plane rectangular coordinate system), and then the plane rectangular coordinate system data is converted into a polar coordinate system, or the relative position of the two is directly determined by mathematical methods in the plane rectangular coordinate system. This usually involves steps such as the conversion of geographic coordinate systems, the application of map projection technology, and mathematical coordinate transformation to ensure that the two can be compared and located in the same spatial reference system. In actual operation, it needs to be completed with the help of professional geographic information system software or tools.

[0115] S22. Coupling rainfall radar echo data with measured data: extract the past rainfall radar echo data and the measured data of rainfall and runoff at the location of the corresponding station; use the rainfall radar echo data, rainfall and runoff measured data of each station in the target basin as the second data set, and split it into a training set and a test set.

[0116] S23, training and testing the rainfall and flood forecasting model based on the second data set. In this step, the rainfall and flood forecasting model is similar to the radar echo extension prediction model, except that the rainfall and flood forecasting model does not include a CNN layer.

[0117] S3. Predict basin rainfall and cross-section flood peaks.

[0118] S31. After the two models are verified, the rainfall radar echo data of a certain period of time in the target basin is first input into the radar echo extension prediction model to predict the changing trend of rainfall radar echo in the future.

[0119] S32. Next, the changing trend of the rainfall radar echo in the future period is input into the rainfall and flood forecasting model to forecast the rainfall and runoff of the target basin in the future period. The flood peak of the basin section can be determined based on the runoff.

[0120] In the prediction process, rainfall radar data has two functions. The first is to make a long series of coupled simulations of rainfall radar data and actual rainfall and runoff data, so that rainfall and runoff can be calculated through radar echoes. The second is the extension of radar echoes, which can be understood as weather forecasts and is used to predict the trend of future radar echoes. In this way, by combining the two models, it is possible to predict short-term rainfall and runoff in the future.

[0121] In other embodiments, the present invention further provides a forecasting system for short-term rainfall and floods based on deep learning, comprising:

[0122] An acquisition module is used to obtain rainfall radar echo data of the target basin in the current period;

[0123] The first prediction module is used to input the acquired rainfall radar echo data into the radar echo extension prediction model to obtain the change trend of rainfall radar echo in the future period of time;

[0124] The second prediction module is used to input the changing trend of rainfall radar echoes in the future period into the rainfall and flood forecasting model to obtain the rainfall and runoff of the target basin in the future period.

[0125] Generally speaking, short-term precipitation forecast refers to the prediction of precipitation events that will occur within a few hours. Such precipitation events are characterized by suddenness and localization, and they may bring strong weather changes, such as heavy rain or thunderstorms. Due to the uncertainty and large intensity changes of short-term precipitation, accurately forecasting such precipitation events is a major challenge for meteorologists.

[0126] Among the methods of short-term precipitation forecasting, meteorological radar monitoring is a key technology. Radar can monitor the development of clouds and precipitation areas in real time, thereby predicting the possibility and intensity of short-term precipitation. In addition, satellite remote sensing technology is also widely used in precipitation forecasting. By analyzing the changes in cloud systems in satellite images, meteorologists can assist in judging the development trend of precipitation events. In addition, numerical weather forecast models and statistical methods based on historical data are also important means of predicting short-term precipitation.

[0127] Ground-based radar plays an important role in short-term precipitation forecasting. This type of radar detects precipitation particles by sending microwaves and receiving their reflected echoes. The intensity of the radar echo can show the size and distribution of precipitation particles, which can be used to predict the occurrence and intensity of short-term precipitation. However, this method also has some limitations. For example, the spatial resolution of the radar is limited and it may not be able to accurately capture localized areas of heavy precipitation. In addition, the radar's blind spots, such as terrain obstruction, may cause precipitation in some areas to be undetectable. Moreover, the processing and interpretation of radar data requires expertise and may involve uncertainties.

[0128] The rainfall radar data extension proposed in this invention is based on the attention mechanism and the CNN-LSTM (convolutional neural network-long short-term memory network) model optimized by deep learning, which can effectively extend the rainfall radar data to predict the changes in the radar cloud map in the next period, such as Figure 2 shown.

[0129] Next, the performance of the method of the present invention is verified.

[0130] By using the first 16 radar images (80 minutes), the radar echo characteristics of the last 16 images (80 minutes) are predicted. The prediction errors of each model are shown in Table 1.

[0131] Table 1 Errors of six models in predicting short-term radar echoes

[0132]

[0133] Figures 3 to 8 The error changes of the six models in the number of iterations during the training process. Combined with the training time and training results in Table 1, it can be seen that as a simple sequence data model, RNN has the shortest training time, but the error is also the largest among the six models. Figure 3As shown in the figure, as an improved model that considers the relationship between sequence data and RNN, LSTM can greatly improve the accuracy of the model while keeping the running time basically the same as RNN. Figure 5 As shown in Figure 2, the CNN model is a model that can extract local features of an image. Its simulation error is lower than that of the RNN model, but its running time is significantly increased, as shown in Figure 2. Figure 4 shown.

[0134] The CNN-LSTM model combines the characteristics of the CNN model and the LSTM model. Its running time is slightly longer than that of the CNN model, but its simulation error is significantly lower than that of the previous three models. Figure 6 The Transformer model is composed of multiple encoders and decoders stacked together. Since this patent does not use parallel computing, its running time is the longest. It also shows good performance in the calculation of multi-band large-scale data sets. It has the smallest error in predicting short-term radar echoes, as shown in Figure 7 As shown in Figure 2, the operation time of the CNN-LSTM-Attention model has been significantly improved due to the addition of the attention mechanism. The more efficient attention mechanism has greatly improved the training and prediction efficiency, although its error in training is slightly higher than that of the Transformer model, as shown in Figure 2. Figure 8 However, due to its lower computation time, we believe that the CNN-LSTM-Attention model has both computational efficiency and prediction accuracy among the six models at this stage, and is more suitable for the current short-term radar echo extension work.

[0135] The short-term rainfall data based on the real-time correction of radar echo extrapolation are input into the rainfall and flood forecasting model to predict the inflow of small reservoirs in the basin. Experiments show that the hydrological simulation values ​​based on the real-time correction of radar echo extrapolation are highly consistent with the measured values, and basically remain on a 1:1 straight line.

[0136] Based on the relationship between the measured value and the simulated value, the present invention further analyzes the mean square error (MSE) between the two. MSE is a commonly used error evaluation index, especially in regression models, used to measure the difference between the predicted value and the true value. The smaller the MSE value, the closer the model's prediction is to the true value. Compared with the measured value , can be calculated using the following formula:

[0137]

[0138] Depend on Fig. 9It can be seen that, overall, the MSE of the simulation value of the present invention is basically maintained in the range of about -0.42 to 0.60 (m). From the perspective of error probability density (thick solid line fitting curve), the error of the simulation value is mainly distributed in the range of ±0.2m, and the MSE of most (nearly 70%) stations is within a smaller error range (-0.1146 to 0.091).

[0139] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0140] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for forecasting short-term rainfall and floods based on deep learning, characterized in that: The following steps are involved: Constructing a radar echo extension prediction model and a rainfall and flood forecasting model based on deep learning, and using the first data set and the second data set to train and test the two constructed models respectively; the radar echo extension prediction model adopts a CNN-LSTM model; Inputting rainfall radar echo data of a certain period of time in the target watershed into the radar echo extension prediction model to predict the changing trend of rainfall radar echo in the future period of time; Inputting the changing trend of rainfall radar echoes in the future period into the rainfall and flood forecasting model to forecast the rainfall and runoff of the target basin in the future period; The process of constructing the first data set includes: obtaining historical rainfall radar echo data, preprocessing the data, and using the data set as the first data set, wherein the preprocessing method of the historical rainfall radar echo data includes data calibration; The data calibration includes: for any moment, comparing the rainfall radar echo data with the actual precipitation measured by the ground rain gauge, and adjusting the rainfall radar echo data to match the reading of the ground rain gauge; wherein the radar echo data is corrected by establishing a statistical relationship between the radar reflectivity factor and the ground rain gauge data; The construction process of the second data set includes: Convert the latitude and longitude coordinates of the target watershed to the same coordinate system as the rainfall radar echo data; Extract the past rainfall radar echo data and rainfall and runoff measured data of the corresponding station location; The rainfall radar echo data, rainfall and runoff measured data of each station in the target basin are used as the second data set and split into a training set and a test set.

2. The method for forecasting short-term rainfall and floods based on deep learning according to claim 1, characterized in that: The preprocessing methods for historical rainfall radar echo data also include: data quality control, data interpolation and resampling, noise reduction and smoothing, and normalization.

3. The method for forecasting short-term rainfall and floods based on deep learning according to claim 2, characterized in that: The data quality control method includes: removing ground clutter, processing the influence of atmospheric refraction, and identifying and removing non-meteorological echoes.

4. The method for forecasting short-term rainfall and floods based on deep learning according to claim 2, characterized in that: The data interpolation and resampling methods include: The Kriging interpolation method is used to spatially interpolate rainfall radar echo data; The linear interpolation method is used to perform time interpolation on rainfall radar echo data.

5. The method for forecasting short-term rainfall and floods based on deep learning according to claim 1, characterized in that: The radar echo extension prediction model includes an input layer, a CNN layer, an LSTM layer, an attention mechanism layer and an output layer; The input layer is used to receive rainfall radar echo data; The CNN layer is used to extract the spatial features of the rainfall radar echo data as the input of the LSTM layer; The LSTM layer is used to recursively process the spatial features of the rainfall radar echo data step by time, and output a feature vector containing time series information; The attention mechanism layer is used to perform weighted aggregation on the feature vector output by the LSTM layer to obtain a context feature vector; The output layer is used to convert the context feature vector output by the attention mechanism layer into a final prediction result to obtain the change trend of the rainfall radar echo in the future period of time.

6. The method for forecasting short-term rainfall and floods based on deep learning according to claim 1, characterized in that: The rainfall and flood forecasting model is constructed based on LSTM network and attention mechanism.

7. A forecasting system for short-term rainfall and floods based on deep learning, characterized in that: include: An acquisition module is used to obtain rainfall radar echo data of the target basin in the current period; A first prediction module, used for inputting the acquired rainfall radar echo data into the radar echo extension prediction model as described in any one of claims 1 to 6, to obtain the change trend of the rainfall radar echo in a future period of time; The second prediction module is used to input the change trend of the rainfall radar echo in the future period into the rainfall and flood forecasting model as described in any one of claims 1-6 to obtain the rainfall and runoff of the target basin in the future period.

Images