Multi-modal area rainfall intelligent prediction method based on ensemble forecast
By constructing a neural network based on convlstm and CNN, the basin precipitation, upstream water vapor and topographic characteristics were extracted, and the uncertainty problem in the medium-term precipitation forecast of a single deterministic numerical weather forecast model was solved, and more accurate flood warning and reservoir scheduling were achieved during the basin flood season.
Patent Information
- Application Number
- CN202510535307.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-27
- Publication Date
- 2025-08-08
AI Technical Summary
In the prior art, a single deterministic numerical weather forecast model has uncertainty in the medium-term precipitation forecast, which is difficult to meet the needs of hydropower enterprises and water conservancy management departments, especially in flood warning and reservoir scheduling during flood seasons of large river basins.
A neural network based on convlstm is constructed to extract precipitation features in the basin area, a convlstm neural network based on convlstm extracts upstream water vapor features, a neural network based on CNN extracts topographic features, and a model is trained through convolutional layers and fully connected layers, combining multiple modal information for prediction.
It improves the accuracy and reliability of medium-term precipitation forecasts, and can effectively conduct flood warnings and reservoir dispatches during flood seasons of large river basins, providing scientific support.
Smart Images
Figure CN120448966A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of meteorology and energy, and in particular to a multi-modal surface rainfall intelligent prediction method based on ensemble forecasting. Background Art
[0002] Most of China lies in the northern temperate zone. Influenced by the East Asian and Indian monsoons in summer, China often experiences persistent heavy rainfall, which rapidly rises river levels and causes floods, severely impacting people's livelihoods and production. Currently, short-term precipitation forecasts three days in advance are highly accurate, providing valuable information for government decision-making and helping to prevent loss of life and property. Hydropower companies and related water management departments, however, require longer-term, medium-term precipitation forecasts to manage reservoirs, plan power generation at power plants, and rationally allocate water resources.
[0003] Currently, the main method for predicting medium-term quantitative precipitation is numerical weather prediction. However, the chaotic nature of the atmosphere and the nonlinear characteristics of atmospheric motion cause even small errors in the initial field to be amplified during the model integration process. Therefore, the forecast results of a single deterministic numerical weather prediction model are subject to significant uncertainty. The emergence of ensemble forecasting technology has provided new insights into the development of deterministic numerical weather prediction. It not only overcomes the shortcomings of a single deterministic model but also leverages the advantages of multi-member forecasting, providing more information relevant to the forecast and improving the usability of numerical forecasts. With the development of ensemble forecasting, the application of medium-term precipitation ensemble forecasts has become increasingly widespread. Summary of the Invention
[0004] In view of the above problems, the present invention is proposed to provide a multi-modal area rainfall intelligent prediction method based on ensemble forecasting, which overcomes the above problems or at least partially solves the above problems.
[0005] According to one aspect of the present invention, a multimodal surface rainfall intelligent prediction method based on ensemble forecast is provided, the intelligent prediction method comprising:
[0006] Construct a neural network based on ConvLSTM to extract precipitation characteristics of the watershed area;
[0007] Construct a convolutional neural network based on ConvLSTM to extract upstream water vapor characteristics;
[0008] Build a CNN-based neural network to extract terrain features;
[0009] The precipitation characteristics of the watershed area, the water vapor characteristics of the upstream, and the terrain characteristics are trained with the observation data through a convolutional layer + a fully connected layer to obtain a training model;
[0010] The total precipitation is predicted based on the trained model.
[0011] Optionally, the step of constructing a ConvLSTM-based neural network to extract precipitation characteristics of a watershed area specifically includes:
[0012] The 10-day sum of basin precipitation is the target, and the results given by numerical weather forecasts are probabilistic forecasts;
[0013] The probability forecast consists of 51 members, and the median, mean, 20th, 30th, 40th, 60th, 70th, 80th, and 90th percentile numerical weather forecasts are extracted as the input of the ConvLSTM neural network model. The area covered by the numerical weather forecast is 5×5, and the input data dimension is 10×5×5×9;
[0014] The convlstm-based neural network includes a batch-normalization layer and a double-layer convlstm+Dropout layer;
[0015] The first convolutional layer has 128 3×3 convolution kernels with a stride of 1. The size of the output feature map after the convolution operation is consistent with the size of the input data, and the output of the entire time series is returned. To prevent overfitting, the neuron dropout rate is 30%.
[0016] The second convolutional layer is similar to the first convolutional layer, returning the output of the last time step of the sequence, and the neuron dropout ratio is also 30%.
[0017] Optionally, constructing a convolution-based ConvLSTM neural network to extract upstream water vapor features specifically includes:
[0018] The upstream region calculates the direction of water vapor movement in the target basin through statistical analysis, and constructs a window that is much larger than the target area and biased towards the direction of water vapor. The mean, median, and 75th percentile of the medium-term ensemble forecast of the numerical weather forecast are extracted as the input of the convolutional ConvLSTM neural network. The time resolution is daily, and the input data dimension is 10×25×25×3.
[0019] The convolutional convlstm neural network consists of a batch-normalization layer and a double-layer convlstm+Dropout layer;
[0020] The first convolutional layer has 128 3×3 convolution kernels with a stride of 1. The size of the output feature map after the convolution operation is consistent with the size of the input data, and the output of the entire time series is returned.
[0021] The neuron discard ratio is 30%;
[0022] The second convolutional layer is similar to the first convolutional layer and returns the output of the last time step of the sequence with a neuron dropout ratio of 30%.
[0023] Optionally, the field of view dimension of the upstream water vapor feature is 25.
[0024] Optionally, the constructing of a CNN-based neural network to extract terrain features specifically includes:
[0025] The area corresponding to the terrain is the size of the watershed area, and the resolution of the terrain information is higher than the resolution of the ensemble forecast of the numerical weather forecast;
[0026] The resampling is 1 / 50 of the ensemble forecast resolution, the terrain extent is 250×250, and the input data dimension is 250×250×1;
[0027] The CNN-based neural network includes a batch-normalization layer and a 7-layer conv2D+MaxPooling2D+Dropout combination model. The convolution layer has 64 3×3 convolution kernels with a stride of 1. The convolution output is max-pooled with a pooling window size of 2×2 and a neuron dropout ratio of 20% to extract terrain features.
[0028] Optionally, the training of the precipitation characteristics of the watershed area, the upstream water vapor characteristics, and the terrain characteristics with observation data through a convolutional layer + a fully connected layer to obtain a training model specifically includes:
[0029] Fusion after extracting multiple modal information;
[0030] Adjust the grid size of the upstream water vapor characteristics and terrain characteristics output to meet the watershed grid size of 5×5, splice them in the channel dimension, and input a 5×5 array of 320 channels;
[0031] Flatten the precipitation characteristics of the watershed area, the upstream water vapor characteristics, and the terrain characteristics, connect them to a fully connected layer with a relu activation function, and add a dropout layer with a discard ratio of 30%;
[0032] The deep learning model is trained using the Adam optimizer, with a cosine annealing learning rate, and a loss function of mean_squared_error by simulating the shape of the cosine curve.
[0033] Optionally, the multiple modal information specifically includes: water vapor direction, basin precipitation and topography.
[0034] The present invention provides a multimodal area rainfall intelligent prediction method based on ensemble forecasting. The method comprises: constructing a ConvLSTM-based neural network to extract basin-wide precipitation characteristics; constructing a ConvLSTM-based neural network to extract upstream water vapor characteristics; constructing a CNN-based neural network to extract terrain characteristics; training the basin-wide precipitation characteristics, upstream water vapor characteristics, and terrain characteristics with observation data through a convolutional layer + fully connected layer method to obtain a training model; and predicting total precipitation based on the training model. Machine learning methods are used to study the forecasting technology for total precipitation in a basin during the flood season, and the sensitivity of the number of ensemble members modeled and the range of precipitation characteristic variables to the forecast effect are studied and analyzed. This method is used to address flood warning and reservoir scheduling issues in large basins during the flood season.
[0035] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0037] Figure 1 A flowchart of a multi-modal surface rainfall intelligent prediction method based on ensemble forecasting provided by an embodiment of the present invention;
[0038] Figure 2 A detailed flow chart of a multi-modal area rainfall intelligent prediction method based on ensemble forecast provided by an embodiment of the present invention;
[0039] Figure 3 The range of a river and the distribution of meteorological stations provided in an embodiment of the present invention (the green box is the ensemble forecast spatial range, the blue area is the river basin range, the red box is the downstream range of a river, and the black dots are the distribution of meteorological stations in the downstream of a river);
[0040] Figure 4 A schematic diagram of the comparison of actual prediction samples, intelligent prediction and ensemble average provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0041] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.
[0042] The terms "comprises" and "comprising" and any variations thereof in the description, embodiments, claims and drawings of the present invention are intended to cover non-exclusive inclusions, for example, including a series of steps or units.
[0043] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and embodiments.
[0044] like Figure 1 and Figure 2 As shown, a multi-modal surface rainfall intelligent prediction method based on ensemble forecast includes:
[0045] Construct a neural network based on ConvLSTM to extract precipitation characteristics of the watershed area;
[0046] Construct a convolutional neural network based on ConvLSTM to extract upstream water vapor characteristics;
[0047] Build a CNN-based neural network to extract terrain features;
[0048] The precipitation characteristics of the watershed area, the water vapor characteristics of the upstream, and the terrain characteristics are trained with the observation data through a convolutional layer + a fully connected layer to obtain a training model;
[0049] The total precipitation is predicted based on the trained model.
[0050] Construct a neural network based on ConvLSTM to extract the precipitation characteristics of the basin area, including:
[0051] The 10-day sum of basin precipitation is the target, and the results given by numerical weather forecasts are probabilistic forecasts;
[0052] The probability forecast consists of 51 members, and the median, mean, 20th, 30th, 40th, 60th, 70th, 80th, and 90th percentile numerical weather forecasts are extracted as the input of the ConvLSTM neural network model. The area covered by the numerical weather forecast is 5×5, and the input data dimension is 10×5×5×9;
[0053] The convlstm-based neural network includes a batch-normalization layer and a double-layer convlstm+Dropout layer;
[0054] The first convolutional layer has 128 3×3 convolution kernels with a stride of 1. The size of the output feature map after the convolution operation is consistent with the size of the input data, and the output of the entire time series is returned. To prevent overfitting, the neuron dropout rate is 30%.
[0055] The second convolutional layer is similar to the first convolutional layer, but only returns the output of the last time step of the sequence, and the neuron dropout ratio is also 30%.
[0056] Construct a convolution-based ConvLSTM neural network to extract upstream water vapor features, including:
[0057] The upstream region calculates the direction of water vapor movement in the target basin through statistical analysis, and constructs a window that is much larger than the target area and biased towards the direction of water vapor. The mean, median, and 75th percentile of the medium-term ensemble forecast of the numerical weather forecast are extracted as the input of the convolutional ConvLSTM neural network. The time resolution is daily, and the input data dimension is 10×25×25×3.
[0058] The convolutional convlstm neural network consists of a batch-normalization layer and a double-layer convlstm+Dropout layer;
[0059] The first convolutional layer has 128 3×3 convolution kernels with a stride of 1. The size of the output feature map after the convolution operation is consistent with the size of the input data, and the output of the entire time series is returned. To prevent overfitting, the neuron dropout rate is 30%.
[0060] The second convolutional layer is similar to the first convolutional layer, but only returns the output of the last time step of the sequence, and the neuron dropout ratio is also 30%.
[0061] The horizon dimension of the upstream water vapor feature is 25.
[0062] Construct a CNN-based neural network to extract terrain features, including:
[0063] The area corresponding to the terrain is the size of the watershed area, and the resolution of the terrain information is higher than the resolution of the ensemble forecast of the numerical weather forecast;
[0064] The resampling is 1 / 50 of the ensemble forecast resolution, the terrain extent is 250×250, and the input data dimension is 250×250×1;
[0065] The CNN-based neural network includes a batch-normalization layer and a 7-layer conv2D+MaxPooling2D+Dropout combination model. The convolution layer has 64 3×3 convolution kernels with a stride of 1. The convolution output is max-pooled with a pooling window size of 2×2. To prevent overfitting, the neuron dropout rate is 20% to extract terrain features.
[0066] The precipitation characteristics of the basin area, upstream water vapor characteristics and terrain characteristics are trained with the observation data through the convolution layer + fully connected layer. The training model specifically includes:
[0067] Fusion after extracting multiple modal information;
[0068] Adjust the grid size of the upstream water vapor characteristics and terrain characteristics output to meet the watershed grid size of 5×5, splice them in the channel dimension, and input a 5×5 array of 320 channels;
[0069] Flatten the precipitation characteristics of the watershed area, the upstream water vapor characteristics, and the terrain characteristics, connect them to a fully connected layer with an activation function of 'relu', and add a dropout layer with a dropout ratio of 30%;
[0070] The deep learning model is trained using the Adam optimizer, with a cosine annealing learning rate, and a loss function of mean_squared_error by simulating the shape of the cosine curve.
[0071] The multiple modal information specifically includes: water vapor direction, basin precipitation and topography.
[0072] The specific design of the model for the multi-modal area rainfall intelligent prediction method based on ensemble forecast is as follows:
[0073] The model is divided into upstream water vapor feature extraction module, basin precipitation feature extraction module, terrain feature extraction module, and the convolution and fully connected layer modules of the model.
[0074] Upstream Water Vapor Feature Extraction Module: Medium- and long-term precipitation is closely related to upstream water vapor in atmospheric motion. The upstream water vapor feature extraction module designed in the patent expands the model's field of view. The upstream region primarily uses statistical analysis to calculate the direction of water vapor movement in the target basin. A window is designed that is significantly larger than the target region and biased toward the incoming water vapor. The module extracts the mean, median, and 75th percentile of the medium-term ensemble forecast of the Numerical Weather Prediction (NWP) as input. The temporal resolution is daily, meaning the input data dimensions are 10×25×25×3. The module consists of a batch-normalization layer and a two-layer ConvLSTM+Dropout layer. The first layer has a filter=128, kernel_size=3, strides=1, and padding='same'. The Dropout layer has a dropout ratio of 30%. The second layer has a filter=128, kernel_size=3, strides=1, and padding='same'. The second layer does not return a sequence, and the Dropout layer also has a dropout ratio of 30%. The field of view dimension is 25, which mainly considers the larger field of view of the basin precipitation water vapor module and the final selected range of the speed of water vapor movement in a day.
[0075] Basin precipitation feature extraction module: The 10-day sum of basin precipitation is the target. The result given by the numerical weather forecast is a probabilistic forecast. How to obtain a deterministic forecast from the probabilistic forecast requires an appropriate means to obtain it from the set members, of which the probabilistic forecast has 51 members. In order to reduce the amount of calculation and model parameters, in addition to the forecast median and average values, the 20th, 30th, 40th, 60th, 70th, 80th, and 90th percentile numerical weather forecasts are also extracted as the input of the model. The area covered by the numerical weather forecast is 5×5, that is, the input data dimension is 10×5×5×9. The module consists of a batch-normalization layer and a double-layer convlstm+Dropout layer, where the first layer has filter=128, kenel_size=3, strides=1, padding='same', and the dropout layer discard ratio is 30%. The second layer has filter=128, kernel_size=3, strides=1, padding='same', the second layer does not return a sequence, and the ratio of neurons discarded in the Dropout layer is also 30%.
[0076] Terrain feature extraction module: Terrain is taken into account in the model as a factor affecting precipitation. The area corresponding to the terrain is the size of the watershed area. Generally, the resolution of terrain information is much higher than the resolution of the ensemble forecast of the numerical weather forecast. It is generally resampled to one-fiftieth of the ensemble forecast resolution, that is, the range of the terrain is 250×250, and the input data dimension is 250×250×1. The module consists of a batch-normalization layer and a 7-layer conv2D+MaxPooling2D+Dropout combination model, where conv2D has filter=128, kernel_size=3, strides=1, padding='same', MaxPooling2D has pooling_size=2, strides=1, and the neuron discarding ratio of Dropout is 20%. Through such a network design, the extraction of terrain features is completed.
[0077] Fully connected layer module: This module is mainly responsible for fusing the extracted information of different modalities (water vapor direction, basin precipitation, and topography). This module first adjusts the grid size output by the upstream water vapor feature module and the terrain feature module so that they both meet the basin grid size (5×5). It then concatenates the data in the channel dimension, inputs a 5×5 array of 320 channels, flattens the three feature modules, and then connects to a fully connected layer with the activation function 'relu'. A dropout layer with a discard ratio of 30% is added to prevent overfitting.
[0078] The deep learning model is trained using the Adam optimizer and a cosine annealing learning rate. By simulating the shape of the cosine curve, the learning rate is gradually reduced and the loss function is mean_squared_error.
[0079] Example 2
[0080] The ensemble forecast product used in this example is the China Meteorological Administration's Global Ensemble Forecast Product (CMA-GEPS), which has a horizontal resolution of 0.5° × 0.5° and a vertical resolution of 87 layers. It includes one control forecast and 30 perturbation members, updated in two batches daily (00 UTC and 12 UTC), with a forecast horizon of 15 days. Output is provided at 3-hour intervals from 0 to 120 hours and at 6-hour intervals from 120 to 360 hours. The precipitation data is hourly data from national ground stations in China, including hourly observations of temperature, air pressure, relative humidity, water vapor pressure, wind, precipitation, and other elements. This data has undergone quality control, and the integrity of each element exceeds 99.9%, with data accuracy approaching 100%.
[0081] The forecast and observation data selected in this embodiment cover a river basin. The target basin is the lower reaches of a river basin. There are 10 meteorological stations in the basin. The distribution is shown in Figure 2. Figure 3 The historical observation and live time period is from May to September 2020-2024, with a total sample size of 1530 batches. The spatial range of CMA ensemble forecast, the range of a river basin and its downstream, and the distribution of meteorological stations are shown in Figure 3 .
[0082] After data quality control and default elimination, 1,399 samples remained. These were randomly sampled and divided into training, test, and prediction sets, with sample sizes of 979, 210, and 210, respectively, using the training set and the ten-day rainfall prediction model established by this invention. Training, feature learning, and model parameter optimization were performed, and model evaluation and selection were performed using the test set to obtain the optimal model. Finally, based on the prediction set data, a ten-day rainfall forecast for the lower reaches of the Yalong River was made.
[0083] The present invention uses common precipitation verification methods in the hydropower field, such as absolute error, root mean square error, pass rate, coefficient of certainty, and accuracy level, to test the surface rainfall and precipitation forecast method. The specific calculations include:
[0084] Absolute error
[0085] The absolute error of a meteorological element is the forecast value minus the measured value. The average of multiple absolute errors represents the average error level of multiple forecasts.
[0086]
[0087] Where, MAE represents the mean absolute error; y o (i) indicates the actual precipitation; y e (i) indicates precipitation forecast.
[0088] Root mean square error
[0089] The square root of the ratio of the sum of the squares of the forecast value and the measured value of a meteorological element to the number of observations n.
[0090]
[0091] Where RMSE represents the root mean square error; o (i) indicates the actual precipitation; y e (i) represents precipitation forecast; n represents the length of the data.
[0092] Pass rate
[0093] A qualified forecast is defined as one in which the forecast error is less than the permissible error. The permissible error is the permissible error range determined based on the requirements for the use of the forecast results and the actual level of forecasting technology. In this invention, 20% of the measured precipitation is used as the permissible error for precipitation forecasts. The qualified rate is the percentage of qualified forecasts divided by the total number of forecasts, representing the overall accuracy level of multiple forecasts. A larger value indicates that the forecast is closer to the actual situation. The calculation formula is as follows:
[0094]
[0095] Where QR represents the qualified rate (rounded to one decimal place), %; n represents the number of qualified forecasts; and m represents the total number of forecasts.
[0096] Determination coefficient
[0097] The certainty coefficient is the degree of agreement between the precipitation forecast and the actual precipitation observation. The larger the value, the closer the forecast is to the actual precipitation. The calculation formula is as follows:
[0098]
[0099] Where DC represents the coefficient of certainty (take 2 decimal places); y o (i) indicates the actual precipitation; y e (i) indicates precipitation forecast; represents the mean of actual precipitation; n represents the length of the data.
[0100] Accuracy grade
[0101] The accuracy of precipitation forecast is divided into three levels according to the pass rate or the size of the certainty coefficient. The precipitation forecast accuracy levels are detailed in Table 1.
[0102] Table 1 Precipitation forecast accuracy level
[0103] Accuracy grade First Second C Pass rate / % QR≥85.0 85.0>QR≥70.0 70.0>QR≥60.0 Determination coefficient DC>0.90 0.90>DC≥0.70 0.70>QR≥0.50
[0104] The prediction results of 210 prediction samples of a river are compared with the actual precipitation and compared with the original ensemble average ( Figure 4 ), it can be seen that after the multimodal intelligent prediction method of the present invention is applied to the ensemble forecast product of the lower reaches of a river, the ten-day surface rainfall forecast (blue line) is closer to the actual surface rainfall (red line), especially the forecast error of large-scale precipitation is significantly reduced compared with the ensemble average (black line).
[0105] Table 2 compares the absolute error, root mean square error, pass rate, coefficient of certainty, and accuracy level of intelligent prediction and ensemble average. It can be seen that all test indicators show that intelligent prediction is better than ensemble average, and it has made significant improvements over ensemble average. The pass rate and coefficient of certainty of intelligent prediction are both Class B, while the ensemble average is much lower than Class C.
[0106] Table 2 Comparison of intelligent prediction and ensemble average test indicators
[0107]
[0108]
[0109] This paper uses machine learning methods to study the forecasting technology of total precipitation in a river basin during the flood season based on actual precipitation observation data, combined with medium-term ensemble forecast members and ensemble statistics. It also studies and analyzes the sensitivity of the number of ensemble member modeling and the range of precipitation characteristic variables to the forecast effect, which is used to solve flood warning and reservoir scheduling in large river basins during the flood season.
[0110] Beneficial Effects: This invention establishes a multimodal surface rainfall intelligent prediction method based on ensemble forecasting. By introducing the original medium- and long-term probabilistic forecast results, terrain characteristics, and upstream water vapor characteristics, it shows good results in forecasting the total precipitation in the next 10 days. It can also perform medium-term surface rainfall forecasts in some basins that are mainly affected by upstream water vapor and complex terrain. It provides a relatively simple and fast method for medium-term deterministic forecasting based on artificial intelligence technology. The invention can be applied in the intersection of hydrology and meteorology, providing certain scientific support for water conservancy departments and related hydropower enterprises in flood control and peak shifting, reservoir scheduling, and flood control and drought relief.
[0111] This paper uses medium-term ensemble forecasts from Numerical Weather Prediction (NWP) and precipitation monitoring information from meteorological stations (automatic and national). It uses ConvLSTM and CNN to build a multimodal neural network. This innovative technology, which takes into account water vapor direction and topographic characteristics, proposes a forecast for total precipitation in a river basin during the flood season. It provides a 10-day forecast of total precipitation in the river basin.
[0112] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A multi-modal surface rainfall intelligent prediction method based on ensemble forecast, characterized in that: The intelligent prediction method comprises: Construct a neural network based on ConvLSTM to extract precipitation characteristics of the watershed area; Construct a convolutional neural network based on ConvLSTM to extract upstream water vapor characteristics; Build a CNN-based neural network to extract terrain features; The precipitation characteristics of the watershed area, the water vapor characteristics of the upstream, and the terrain characteristics are trained with the observation data through a convolutional layer + a fully connected layer to obtain a training model; The total precipitation is predicted based on the trained model.
2. The multi-modal surface rainfall intelligent prediction method based on ensemble forecast according to claim 1, characterized in that: The construction of the ConvLSTM-based neural network to extract the precipitation characteristics of the watershed area specifically includes: The 10-day sum of basin precipitation is the target, and the results given by numerical weather forecasts are probabilistic forecasts; The probability forecast consists of 51 members, and the median, mean, 20th, 30th, 40th, 60th, 70th, 80th, and 90th percentile numerical weather forecasts are extracted as the input of the ConvLSTM neural network model. The area covered by the numerical weather forecast is 5×5, and the input data dimension is 10×5×5×9; The convlstm-based neural network includes a batch-normalization layer and a double-layer convlstm+Dropout layer; The first convolutional layer has 128 3×3 convolution kernels with a stride of 1. The size of the output feature map after the convolution operation is consistent with the size of the input data, and the output of the entire time series is returned. To prevent overfitting, the neuron dropout rate is 30%. The second convolutional layer is similar to the first convolutional layer, returning the output of the last time step of the sequence, and the neuron dropout ratio is also 30%.
3. The multi-modal surface rainfall intelligent prediction method based on ensemble forecast according to claim 1, characterized in that: The construction of the convolution-based ConvLSTM neural network to extract upstream water vapor features specifically includes: The upstream region calculates the direction of water vapor movement in the target basin through statistical analysis, and constructs a window that is much larger than the target area and biased towards the direction of water vapor. The mean, median, and 75th percentile of the medium-term ensemble forecast of the numerical weather forecast are extracted as the input of the convolutional ConvLSTM neural network. The time resolution is daily, and the input data dimension is 10×25×25×3. The convolutional convlstm neural network consists of a batch-normalization layer and a double-layer convlstm+Dropout layer; The first convolutional layer has 128 3×3 convolution kernels with a stride of 1. The size of the output feature map after the convolution operation is consistent with the size of the input data, and the output of the entire time series is returned. The neuron discard ratio is 30%; The second convolutional layer is similar to the first convolutional layer and returns the output of the last time step of the sequence with a neuron dropout ratio of 30%.
4. The multi-modal surface rainfall intelligent prediction method based on ensemble forecast according to claim 1 is characterized in that: The field of view dimension of the upstream water vapor feature is 25.
5. The multi-modal surface rainfall intelligent prediction method based on ensemble forecast according to claim 1, characterized in that: The construction of the CNN-based neural network to extract terrain features specifically includes: The area corresponding to the terrain is the size of the watershed area, and the resolution of the terrain information is higher than the resolution of the ensemble forecast of the numerical weather forecast; The resampling is 1 / 50 of the ensemble forecast resolution, the terrain extent is 250×250, and the input data dimension is 250×250×1; The CNN-based neural network includes a batch-normalization layer and a 7-layer conv2D+MaxPooling2D+Dropout combination model. The convolution layer has 64 3×3 convolution kernels with a stride of 1. The convolution output is max-pooled with a pooling window size of 2×2 and a neuron dropout ratio of 20% to extract terrain features.
6. The multi-modal surface rainfall intelligent prediction method based on ensemble forecast according to claim 1, characterized in that: The training of the precipitation characteristics of the watershed area, the water vapor characteristics of the upstream, and the terrain characteristics with the observation data through a convolutional layer + a fully connected layer to obtain a training model specifically includes: Fusion after extracting multiple modal information; Adjust the grid size of the upstream water vapor characteristics and terrain characteristics output to meet the watershed grid size of 5×5, splice them in the channel dimension, and input a 5×5 array of 320 channels; Flatten the precipitation characteristics of the watershed area, the upstream water vapor characteristics, and the terrain characteristics, connect them to a fully connected layer with a relu activation function, and add a dropout layer with a discard ratio of 30%; The deep learning model is trained using the Adam optimizer, a cosine annealing learning rate, and a loss function of mean_squared_error by simulating the shape of the cosine curve.
7. The multi-modal surface rainfall intelligent prediction method based on ensemble forecast according to claim 1, characterized in that: The multiple modal information specifically includes: water vapor direction, basin precipitation and topography.