CN117290810A - Short-time strong precipitation probability prediction fusion method based on cyclic convolutional neural network - Google Patents
Short-time strong precipitation probability prediction fusion method based on cyclic convolutional neural network Download PDFInfo
- Publication number
- CN117290810A CN117290810A CN202311584724.0A CN202311584724A CN117290810A CN 117290810 A CN117290810 A CN 117290810A CN 202311584724 A CN202311584724 A CN 202311584724A CN 117290810 A CN117290810 A CN 117290810A
- Authority
- CN
- China
- Prior art keywords
- precipitation
- short
- time
- prediction
- hour
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000001556 precipitation Methods 0.000 title claims abstract description 305
- 125000004122 cyclic group Chemical group 0.000 title claims abstract description 45
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 34
- 238000007500 overflow downdraw method Methods 0.000 title claims abstract description 17
- 238000012549 training Methods 0.000 claims abstract description 87
- 238000012360 testing method Methods 0.000 claims abstract description 68
- 238000000034 method Methods 0.000 claims abstract description 41
- 238000012795 verification Methods 0.000 claims abstract description 36
- 238000013145 classification model Methods 0.000 claims abstract description 26
- 230000004927 fusion Effects 0.000 claims abstract description 24
- 238000013528 artificial neural network Methods 0.000 claims description 13
- 230000008569 process Effects 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 10
- 238000003908 quality control method Methods 0.000 claims description 10
- 238000010801 machine learning Methods 0.000 claims description 9
- 238000011156 evaluation Methods 0.000 claims description 7
- 238000010606 normalization Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 6
- 230000015654 memory Effects 0.000 claims description 4
- 238000005070 sampling Methods 0.000 claims description 4
- 238000012216 screening Methods 0.000 claims description 4
- 230000003044 adaptive effect Effects 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 238000012163 sequencing technique Methods 0.000 claims description 3
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims 1
- 238000011160 research Methods 0.000 abstract description 2
- 230000000694 effects Effects 0.000 description 11
- 230000008901 benefit Effects 0.000 description 9
- 238000013136 deep learning model Methods 0.000 description 6
- 238000013213 extrapolation Methods 0.000 description 5
- 230000006872 improvement Effects 0.000 description 3
- 230000007774 longterm Effects 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000013103 analytical ultracentrifugation Methods 0.000 description 1
- 239000003086 colorant Substances 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000013277 forecasting method Methods 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 238000004445 quantitative analysis Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 239000011800 void material Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01W—METEOROLOGY
- G01W1/00—Meteorology
- G01W1/10—Devices for predicting weather conditions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Environmental & Geological Engineering (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Atmospheric Sciences (AREA)
- Biodiversity & Conservation Biology (AREA)
- Ecology (AREA)
- Environmental Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a short-time strong precipitation probability prediction fusion method based on a cyclic convolutional neural network, which relates to the field of atmospheric science research and comprises the following steps: step one, establishing a fusion data set of short-time prediction of precipitation, and dividing the fusion data set into a training set, a verification set and a test set of short-time prediction of precipitation; training a precipitation time sequence prediction model based on a cyclic convolutional neural network by using the precipitation short-time prediction training set and the verification set in the first step; a trained precipitation time sequence prediction model based on a cyclic convolutional neural network is adopted to obtain the time-space prediction result of the hour precipitation amount of 6-18 hours in the future; training the LightGBM classification model by adopting the precipitation space-time prediction result obtained in the step two to obtain a short-time strong precipitation classification model, and obtaining a probability prediction result of a short-time strong precipitation event every hour for 6-18 hours in the future. The method can greatly improve the probability prediction capability of short-time strong rainfall.
Description
Technical Field
The invention relates to the field of atmospheric science research, in particular to a short-time strong precipitation probability prediction fusion method based on a cyclic convolution neural network.
Background
Short-time strong precipitation refers to a very strong precipitation event occurring in a short period of time, typically defined as a precipitation event with a 1 hour rainfall above 20 mm. Short-time strong precipitation is usually formed by convection precipitation, has the characteristics of rapid growth and elimination, strong locality and high disaster-causing degree, and is extremely easy to cause serious secondary disasters such as landslide, urban waterlogging, flood and the like. The urban flood caused by the short-time strong precipitation is one of important factors causing the meteorological disasters of large cities, so that the short-time prediction capability of the short-time strong precipitation is very necessary to be enhanced, the short-time strong precipitation prediction early warning result with consistent and accurate time and space is provided, the early warning decision-making capability of the urban meteorological disasters is favorably enhanced, and the extreme weather and serious risk judging mechanism is well developed.
The traditional short time forecast mainly comprises two methods, namely a numerical weather forecast model and an extrapolation method based on mass data driving.
Numerical weather forecast models utilize the physical characteristics of the relevant atmospheric process to model a wide range of atmospheric physical quantities, the most critical one of which is precipitation. Despite the growing day and month of numerical weather forecast patterns, there are many shortcomings in the aspect of the short forecast. For example, the numerical weather forecast mode is mostly based on various assumption conditions on the atmospheric state, and the uncertainty problem of the mode is always a fundamental problem for restricting the development of the numerical weather forecast. Especially, the current observation capability still has difficulty in accurately expressing various states of the atmosphere, and the schemes of convection cloud parameterization, cloud micro-physical parameterization, boundary layer parameterization and the like in a numerical mode often bring huge strong rainfall prediction errors. Therefore, in order to improve the uncertainty of the mode starting stage, the numerical forecasting mode with high space-time resolution needs a more complex assimilation method and sufficient observation data, but the mode generally needs to consume massive computing resources and computing time, so that the difficulty of the high-resolution numerical weather forecasting business is greatly increased.
Unlike the numerical weather forecast model, extrapolation is driven entirely by massive data characterizing atmospheric conditions. Many studies have shown that extrapolation-based precipitation prediction can achieve significantly better prediction results than traditional numerical weather prediction models with a small amount of computational resources over a short prediction time period (0-2 hours). But after more than 2 hours, the precipitation extrapolation prediction accuracy rate is rapidly reduced, and the precipitation prediction accuracy rate of the numerical weather prediction model is higher.
Therefore, in the aspect of short-time rainfall prediction, a very accurate and high-space-time resolution strong rainfall prediction result is difficult to obtain only by means of a numerical weather prediction or data driving method.
The deep learning model can effectively extract various remarkable characteristics from massive meteorological data, and has strong nonlinear fitting capacity and generalization capacity. Recently, there have been many studies to propose an atmospheric state prediction method based on a deep learning model. For example, a 6-hour lightning prediction model based on a convolved long short-term memory network, lightNet, a 12-hour lightning prediction model based on LightNet improvement. But the predictive capability of deep learning models based solely on observations is still difficult to meet business needs. Recently, zhou Kanghui et al established an hour-by-hour lightning prediction model based on SegNet that fused multisource observation data and numerical prediction results, significantly improving the 6 hour lightning prediction accuracy. However, since the deep learning model based on the convolutional neural network learns complex atmospheric state changes by stacking convolutional layers, it is difficult to capture long-term dependency between time frames that are far away. Therefore, even if the pattern prediction data and the observation data are combined, it is difficult to improve the atmospheric prediction ability for 6 to 12 hours in the future.
The time sequence prediction of precipitation is a more complex prediction task than lightning prediction, and the time-space distribution of different precipitation intensities needs to be given simultaneously, so that more atmospheric physical states and more complex prediction mechanisms are involved. In addition, the short-time strong precipitation event is a small probability event, has a serious sample imbalance problem, and is difficult to solve the short-time prediction problem of the short-time strong precipitation simply by means of the conventional time sequence prediction model. In the prior art, the prediction of the short-time strong precipitation is mostly to simply select a precipitation threshold of 20mm/h, and the precipitation prediction is directly converted into the deterministic prediction of the short-time strong precipitation. However, this method relies heavily on the prediction of strong precipitation, but often the numerical mode has poor prediction of strong precipitation. Unlike deterministic predictions, probabilistic predictions can give spatial distribution of rare events while also predicting probability of occurrence of rare events, especially probabilistic predictions can be used to estimate event occurrence probability associated with maximum acceptable loss and are therefore more valuable than deterministic predictions in weather risk management, but there is currently no short-term probability prediction model specifically for short-term strong precipitation.
Disclosure of Invention
The invention aims to: aiming at the defects of the short-time probability prediction method of the short-time strong precipitation, the invention provides a short-time strong precipitation probability prediction fusion method based on a cyclic convolutional neural network. Taking the 6 th hour as a forecasting starting point, utilizing regional weather pattern forecasting data of 0-18 hours, fusing ground site hour rainfall observation data of the past 0-5 hours, and forecasting probability distribution of short-time strong rainfall of 12 hours (namely 6-18 hours) in the future by adopting a method of fusing a circular convolution neural network and a LightGBM. The fusion is mainly embodied in two aspects, namely, the ground precipitation observation of 0-5 hours is fused into the atmospheric physical elements of 0-18 hours of regional numerical weather forecast, and the probability forecast fusion method of the short-time strong precipitation is finally formed by fusing the precipitation space-time prediction task and the short-time strong precipitation classification task based on the cyclic convolution neural network. The primary aim is to solve the problems of low prediction accuracy and poor prediction effect when the current short-time strong precipitation prediction method is applied to a precipitation business prediction system.
The technical scheme is as follows: a short-time strong precipitation probability prediction fusion method based on a cyclic convolutional neural network comprises the following steps:
step one, establishing a fusion data set of short-time prediction of precipitation, and dividing the fusion data set into a training set, a verification set and a test set of short-time prediction of precipitation;
training a precipitation time sequence prediction model based on a cyclic convolutional neural network by using the precipitation short-time prediction training set and the verification set in the first step; a trained precipitation time sequence prediction model based on a cyclic convolutional neural network is adopted to obtain an hour precipitation prediction result of 6-18 hours in the future;
training the LightGBM classification model by adopting the precipitation space-time prediction result obtained in the step two to obtain a short-time strong precipitation classification model, and obtaining a probability prediction result of a short-time strong precipitation event every hour for 6-18 hours in the future.
Further, the first step specifically includes the following steps:
step 1.1, collecting hour-by-hour ground site hour rainfall observation data and 0-18 hours multivariate data of regional weather numerical model forecast;
step 1.2, carrying out time matching on the hour-by-hour ground site hour precipitation observation data and the regional weather pattern forecast data collected in the step 1.1 to form a data sequence;
step 1.3, eliminating all samples without precipitation aiming at the data sequence formed in the step 1.2;
step 1.4, carrying out quality control, comparison evaluation and normalization processing on the data sequence formed in the step 1.3;
step 1.5, performing space fusion on the data sequence after quality control formed in the step 1.4;
and step 1.6, obtaining a fusion data set of short-time rainfall prediction fused with the ground hour rainfall observation data and the regional weather numerical forecast data through the steps 1.1-1.5, and dividing the fusion data set into a short-time rainfall prediction training set, a verification set and a test set.
Further, in the step 1.1, the forecast variables of the collected regional weather numerical patterns are 24, including temperatures, humidity, warp direction wind, weft direction wind at 925, 850, 700, 500 and 200 hPa, and barometric pressure, convection instability energy, 2 m temperature and hour precipitation of the ground layer.
Further, the specific operation of step 1.2 is as follows: the regional weather numerical mode starts from 0 hour a day and outputs an hour-by-hour atmospheric element forecast of 0-18 hours in the future every 3 hours; matching the ground site hour precipitation observation data into regional weather pattern prediction data, namely forming a data sequence containing 24 pattern prediction variables and ground site hour precipitation observation data every 3 hours, wherein the time interval of the sequence is 1 hour, and the sequence length is 0-18 hours; wherein 0-5 hours is used as a predictor input sequence and 6-18 hours is used as a predictor test sequence.
Further, the specific operation method in step 1.4 is as follows:
firstly, performing quality control, removing data sequences with errors or suspicious observations, and removing data sequences with missing ground hour precipitation observations;
then, carrying out comparison evaluation, accumulating the hour precipitation observation data, comparing the hour precipitation observation data with the day hour precipitation observation data, and if the two are inconsistent, eliminating all data sequences in the same day;
and finally, respectively carrying out normalization processing on each element in the data sequence.
Further, the spatial fusion method in step 1.5 specifically comprises the following steps: and (3) interpolating the ground site hour precipitation observation data of 0-18 hours and the multivariable forecast data of the weather mode in the area of 0-18 hours into uniform grid coordinates by adopting an inverse distance weighted interpolation method, wherein the horizontal resolution is 0.05 degrees, and the grid points are 120 multiplied by 120.
Further, the specific steps of the second step are as follows:
step 2.1, the adopted rainfall time sequence prediction model is a PredRNNv2 model with four layers of space-time long-short-term memory modules;
training the rainfall time sequence prediction model by using the rainfall short-time prediction training set and the verification set obtained in the step one; the input factors of 0-5 hours are 25, and the input factors comprise 24 variables of ground hour rainfall observation and regional weather numerical model prediction; the input factors of 6-18 hours do not contain ground hour precipitation observation, and only 24 variables for regional weather numerical mode prediction are included; the true value is the observed data of the hour precipitation;
an adaptive moment estimation optimizer is used in the training process, the batch size of each training is 8, the learning rate is 0.001, and the maximum training iteration number is 300; the weighted average absolute error is used as a loss function of precipitation prediction, and the formula is as follows:
（1），
wherein the method comprises the steps ofIs the total sample size, +.>Is the real value of the hour precipitation intensity of the ith sample,/>Precipitation intensity prediction for the ith sampleValue of->Is the input factor set for the i-th sample; />Is the precipitation weight coefficient of the ith sample, and the weight is set as follows:
，
the accuracy rate of each training and verification is output in the training process, and the model with the highest verification accuracy is automatically stored;
step 2.2, selecting a model with highest verification precision, namely a trained precipitation time sequence prediction model based on a cyclic convolutional neural network;
step 2.3, saving the model to predict the result of the short-time precipitation prediction for 6-18 hours on the short-time precipitation prediction verification set; and testing the model on the short-time rainfall prediction test set to obtain the hour-by-hour rainfall space-time prediction result of 6-18 hours on the short-time rainfall prediction test set.
Further, the specific steps of the third step are as follows:
step 3.1, a data set is manufactured again aiming at a short-time strong precipitation event classification task of a machine learning model LightGBM, and the data set is split into a training set, a verification set and a test set;
step 3.2, training the LightGBM model by adopting a training set in the data set of the short-time strong precipitation event classification task in step 3.1; outputting the accuracy rate of each training in the training process, and automatically storing a model with the highest accuracy rate as a trained LightGBM short-time strong precipitation classification model;
and 3.3, testing the trained LightGBM short-time strong precipitation classification model by adopting a test set in the data set of the short-time strong precipitation event classification task in the step 3.1, and obtaining a probability prediction result of the short-time strong precipitation event from hour to hour within 6-18 hours.
Further, step 3.1, the step of reproducing the data set for the short-time strong precipitation event classification task of the machine learning model LightGBM includes:
a) Extracting 6-18 hours forecast test sequence data in a verification set and a test set from the short-time forecast fusion data set of the precipitation formed in the first step, and respectively serving as a training set and a test set of the short-time strong precipitation classification data set; the training set of the short-time strong precipitation classification data set and the ground site hour precipitation observation data in the test set are only used as real fields, and are not used as input factors of short-time strong precipitation classification tasks;
b) Adding the 6-18 hour rainfall space-time prediction results obtained in the short-time rainfall prediction verification set and the test set on the test pre-sequencing column into the training set and the test set of the short-time strong rainfall classification data set respectively;
c) Respectively eliminating the hour precipitation quantity predicted by the regional weather numerical mode from a training set and a testing set of the short-time strong precipitation classification data set;
d) According to a short-time strong precipitation threshold value of 20mm/h, converting training sets of short-time strong precipitation classification data sets and ground site hour precipitation observation data and precipitation space-time prediction results in a test set into short-time strong precipitation marks, namely, when the hour precipitation amount on a grid point is greater than 20mm, 1 is smaller than 20mm and 0 is smaller than 1, so that a data set of short-time strong precipitation event classification tasks is formed; the training set of the short-time strong precipitation classification data set corresponds to the training set in the data set forming the short-time strong precipitation event classification task after conversion, and the testing set of the short-time strong precipitation classification data set corresponds to the testing set in the data set forming the short-time strong precipitation event classification task.
Further, the input factors of the LightGBM classification model in step 3.2 are as follows:
a) 23 forecast variables for regional weather numerical patterns including temperature, humidity, warp, weft, and barometric pressure at ground level, convective instability energy, 2 meters temperature at 925, 850, 700, 500, and 200 hPa;
b) 0-1 distribution of short-time strong precipitation events formed by precipitation space-time prediction results;
c) The real field of the LightGBM classification model is the 0-1 distribution of short-time strong precipitation events of the ground site hour precipitation observation data;
during training, automatically adjusting parameters by using a Bayesian optimization method, and then continuing parameter fine adjustment to enable an objective function to reach the global minimum; the specific parameter scheme is as follows: adopting a single-side sampling algorithm based on gradient, wherein the learning rate is 0.01, the depth of each training tree is 2, the maximum leaf number is 25, the maximum iteration number is 5000, the positive and negative sample proportion is automatically adjusted, the model back screening proportion is 50%, and the maximum barrel number is 100; the binary cross entropy is used as a Loss function (Loss) of a short-time strong precipitation event classification task, and the formula is as follows:
（2），
wherein the method comprises the steps ofIs the total sample size, +.>Is the real value of the short-time strong precipitation classification of the ith sample, < >>Is the short-time strong precipitation classification predictive value of the ith sample,/->Is the input factor set for the i-th sample.
Advantageous effects
1) According to the invention, the short-time prediction problem of short-time strong precipitation is disassembled into two tasks, namely a precipitation prediction task based on a cyclic convolutional neural network for the next 12 hours (6-18 hours in the invention) and a short-time strong precipitation classification task based on the LightGBM. The two-step strategy can exert the advantages of the cyclic convolutional neural network in the aspect of large sample time sequence prediction, and can also fully exert the advantages of the LightGBM in the classification task of short-time strong precipitation such as small probability events.
2) The precipitation time sequence prediction model based on the cyclic convolution neural network can simultaneously utilize time and space information, so that rules and trends in data can be extracted from high-dimensional data, and prediction can be realized more accurately.
3) The short-time strong precipitation is a small probability event, so that the advantages of the LightGBM in a small sample classification task can be fully exerted, and a short-time probability forecasting result of the short-time strong precipitation is directly formed.
4) The multiscale atmospheric circulation thermodynamic element provided by the regional numerical mode is used as a prediction factor, so that more movement and evolution information of a precipitation system can be provided for the cyclic convolution neural network. By adding the ground site hour precipitation observation data of 0-5 hours in the past, the cyclic convolution neural network can learn the time and space information of the evolution of a precipitation system, and the probability prediction capability of short-time strong precipitation is greatly improved.
Drawings
FIG. 1 is a flow chart of a method for data assimilation background error covariance feature statistics for strong convection weather typing;
FIG. 2 is a graph of the strongly convective weather pattern typing results in the east China;
FIG. 3 is a graph of aggregate background error covariance multi-scale feature distribution at different spatial scales;
fig. 4 is an example prediction effect diagram of three schemes.
Detailed Description
The invention will be described in detail below with reference to the drawings and the detailed description.
Based on understanding of deep learning and machine learning, the invention provides a short-time strong precipitation probability prediction fusion method based on a cyclic convolutional neural network, which disassembles short-time prediction problems of short-time strong precipitation into two tasks, namely a precipitation time sequence prediction task based on a deep learning model and a short-time strong precipitation classification task based on a machine learning model. The task disassembly has the advantages that on one hand, the time sequence prediction advantage of the deep learning model is fully exerted, and the imbalance problem of short-time strong rainfall samples is avoided; on the other hand, the advantages of the machine learning model in the small sample classification task are fully developed, and the output result of the classification task is the probability of occurrence of the event, so that the short-time probability prediction result of the short-time strong rainfall can be directly formed.
The invention provides a short-time strong precipitation probability prediction fusion method based on a cyclic convolutional neural network, which comprises the following steps:
step one, establishing a fusion data set of short-time prediction of precipitation, and dividing the fusion data set into a training set, a verification set and a test set of short-time prediction of precipitation.
The specific operation steps of the first step are as follows:
and 1.1, collecting the ground site hour rainfall observation data and the regional weather numerical model forecast 0-18 hours of multivariate data hour by hour.
The forecast variables for the collected regional weather numerical patterns were 24, including temperatures, humidity, warp, weft, and barometric pressure at the ground floor, convective instability energy, 2 meters temperature, and hour precipitation at 925, 850, 700, 500, and 200 hPa.
And step 1.2, performing time matching on the hour-by-hour ground site hour precipitation observation data and the regional weather pattern forecast data collected in the step 1.1 to form a data sequence (25 variables).
The regional weather numerical pattern starts at 0 hour a day and outputs an hour-by-hour atmospheric element forecast for 0-18 hours a time every 3 hours. Matching the ground site hour precipitation observation data into regional weather pattern prediction data, namely forming a data sequence containing 24 pattern prediction variables and ground site hour precipitation observation data every 3 hours, wherein the time interval of the sequence is 1 hour, and the sequence length is 0-18 hours; wherein 0-5 hours is used as a predictor input sequence and 6-18 hours is used as a predictor test sequence.
The purpose of the data set so produced is to forecast the probability distribution of short-time strong precipitation for the next 12 hours (i.e. 6-18 hours) using the regional weather pattern forecast data for 0-18 hours, in combination with the ground site hour precipitation observations of the past 0-5 hours, with the 6 th hour as the forecast origin.
And step 1.3, eliminating all samples without precipitation according to the data sequence formed in the step 1.2.
The judging method of the rainfall-free sample comprises the following steps: and (3) carrying out precipitation sequence screening on precipitation data in the precipitation observation data of the ground station for hours, if more than 5% of the precipitation data of 0-18 hours are recorded by the ground precipitation observation station to obtain precipitation amount more than 0.1 mm/h, considering the sequence as precipitation samples, otherwise, judging the sequence as precipitation-free samples, and eliminating all the precipitation-free samples.
And step 1.4, performing quality control, comparison evaluation and normalization processing on the data sequence formed in the step 1.3.
Firstly, quality control is carried out, data sequences with errors or suspicious observations are removed, and data sequences with shortage of ground hour precipitation observations are removed.
And then, carrying out comparison evaluation, accumulating the hour precipitation observation data, comparing the hour precipitation observation data with the day hour precipitation observation data, and if the hour precipitation observation data and the day hour precipitation observation data are inconsistent, eliminating all data sequences in the same day.
Finally, normalization processing is performed on each element (25 elements in total) in the data sequence.
And step 1.5, performing space fusion on the data sequence after quality control formed in the step 1.4.
And (3) interpolating the ground site hour precipitation observation data of 0-18 hours and the multivariable forecast data of the weather mode in the area of 0-18 hours into uniform grid coordinates by adopting an inverse distance weighted interpolation method, wherein the horizontal resolution is 0.05 degrees, and the grid number is 120 multiplied by 120.
And step 1.6, obtaining a short-time rainfall prediction fusion data set fused with the ground hour rainfall observation data and the regional weather numerical forecast data through the steps of time matching, quality control, space fusion and the like. Further aiming at precipitation time sequence prediction tasks, according to 8:1: the proportion of 1 divides the fusion data set into a rainfall short-time prediction training set, a verification set and a test set.
Training a precipitation time sequence prediction model based on a cyclic convolutional neural network by using the precipitation short-time prediction training set and the verification set in the first step; acquiring an hour precipitation forecast result of 6-18 hours by adopting a trained precipitation time sequence prediction model based on a cyclic convolutional neural network;
training a precipitation time sequence prediction model based on a cyclic convolutional neural network PredRNNv2 by using the precipitation short-time prediction training set and the verification set obtained in the step one; and (3) testing on a test set of a short-time rainfall prediction data set by adopting a trained rainfall time sequence prediction model based on the cyclic convolutional neural network, and obtaining an hour rainfall space-time prediction result for 6-18 hours.
The specific steps of the second step are as follows:
step 2.1, the precipitation time sequence prediction model is a PredRNNv2 model with four layers of space-time long-short-term memory modules (ST-LSTM) (figure 2). The model adopts reverse plan sampling, and the model is forced to learn more information about long-term dynamics by reducing the probability of randomly hiding real observation results, 128 convolution kernels are arranged in each layer of space-time long-short-term memory module, the size of each convolution kernel is 5 multiplied by 5, and the convolution step length is 1. The number of predictors input is 25 and the output channel is 1.
Training the rainfall time sequence prediction model by using the rainfall short-time prediction training set and the verification set obtained in the step one; wherein the input factors of 0-5 hours are 25, and the input factors comprise 24 variables of ground hour rainfall observation and regional weather numerical mode prediction. The 6-18 hour input factor does not contain ground hour precipitation observations, only 24 variables for regional weather numerical pattern forecast. The true value is the observed data of the hour precipitation.
In the training process, an adaptive moment estimation optimizer is used, the batch size of each training is 8, the learning rate is 0.001, and the maximum training iteration number is 300. The weighted average absolute error is used as a Loss function (Loss) of precipitation prediction, and the formula is as follows:
（1），
wherein the method comprises the steps ofIs the total sample size, +.>Is the real value of the hour precipitation intensity of the ith sample,/>Is the precipitation intensity predictive value of the ith sample,/->Is the input factor set for the i-th sample. />Is the precipitation weight coefficient of the ith sample, and the weight is set as follows:
，
the machine learning package Pytorch was used in the experiment and was run on a GPU server equipped with 4 RTX 3090. And outputting the accuracy of each training and verification in the training process, and automatically storing the model with the highest verification accuracy.
And 2.2, selecting a model with highest verification precision, namely a trained precipitation time sequence prediction model based on the cyclic convolutional neural network.
And 2.3, saving the model in the short-time rainfall prediction verification set for 6-18 hours. And testing the model on the short-time rainfall prediction test set to obtain the hour-by-hour rainfall space-time prediction result of 6-18 hours on the short-time rainfall prediction test set.
And thirdly, training the LightGBM classification model by adopting the precipitation space-time prediction result obtained in the second step to obtain a short-time strong precipitation classification model.
Based on the short-time rainfall prediction data set obtained in the second step and the hour-by-hour rainfall space-time prediction result of 6-18 hours on the test set, reconstructing the data set for the short-time strong rainfall event classification task based on the LightGBM, wherein the data set comprises a training set and a test set; training a short-time strong precipitation classification model based on the LightGBM based on the newly manufactured training set; and testing on a short-time strong precipitation classification test set by using a trained short-time strong precipitation classification model based on the LightGBM, and obtaining a probability prediction result of the short-time strong precipitation event from hour to hour for 6-18 hours.
The specific operation of the third step is as follows:
step 3.1, a data set is manufactured again aiming at a short-time strong precipitation event classification task of a machine learning model LightGBM, and the steps are as follows:
a) And (3) extracting 6-18 hour prediction test sequence data in the verification set and the test set from the short-time prediction fusion data set of the precipitation formed in the step one, and respectively serving as a training set and a test set of the short-time strong precipitation classification data set. The training set of the short-time strong precipitation classification data set and the ground site hour precipitation observation data in the test set are only used as real fields, and are not used as input factors of short-time strong precipitation classification tasks.
B) And adding the hour-by-hour rainfall space-time prediction results of 6-18 hours, which are obtained on the short-time rainfall prediction verification set and the test pre-sequencing column of the test set, into the training set and the test set of the short-time strong rainfall classification data set respectively.
C) And respectively eliminating the hour precipitation quantity predicted by the regional weather numerical mode from the training set and the test set of the short-time strong precipitation classification data set.
D) According to the short-time strong precipitation threshold value of 20mm/h, the training set of the short-time strong precipitation classification data set and the ground site hour precipitation observation data and the precipitation space-time prediction result in the test set are converted into short-time strong precipitation marks, namely, when the hour precipitation amount on the grid point is larger than 20mm, 1 is smaller than 20mm, and 0 is smaller, so that the data set of the short-time strong precipitation event classification task is formed. The training set of the short-time strong precipitation classification data set corresponds to the training set in the data set forming the short-time strong precipitation event classification task after conversion, and the testing set of the short-time strong precipitation classification data set corresponds to the testing set in the data set forming the short-time strong precipitation event classification task.
And 3.2, training the LightGBM model by adopting a training set in the data set of the short-time strong precipitation event classification task in the step 3.1. And outputting the accuracy rate of each training in the training process, and automatically storing the model with the highest accuracy rate as a trained LightGBM short-time strong precipitation classification model.
The input factors for the LightGBM classification model are as follows:
a) 23 forecast variables for regional weather numerical patterns including temperature, humidity, warp, weft, and barometric pressure at ground level, convective instability energy, 2 meters temperature at 925, 850, 700, 500, and 200 hPa.
B) And 0-1 distribution of short-time strong precipitation events formed by precipitation space-time prediction results.
C) The real field of the LightGBM classification model is the 0-1 distribution of short-time strong precipitation events of the ground site hour precipitation observation data.
The output result of the LightGBM classification model is: probability of strong precipitation for short time (0-1).
During training, the Bayesian optimization method is used for automatic parameter adjustment, and then parameter fine adjustment is continued, so that the objective function reaches the global minimum. The parameter scheme is as follows: the learning rate is 0.01 by adopting a gradient-based unilateral sampling algorithm, the depth of each training tree is 2, the maximum leaf number is 25, the maximum iteration number is 5000, the positive and negative sample proportion is automatically adjusted, the model back screening proportion is 50%, and the maximum barrel number is 100. The binary cross entropy is used as a Loss function (Loss) of a short-time strong precipitation event classification task, and the formula is as follows:
（2），
wherein the method comprises the steps ofIs the total sample size, +.>Is the real value (0 or 1) of the short-time strong precipitation classification of the ith sample,/->Is the short-time strong precipitation classification predictive value (0-1) of the ith sample, +.>Is the input factor set for the i-th sample.
The machine learning software package Pytorch was used in the test and was run on a server equipped with 4 RTX3090 graphics cards.
And 3.3, testing the trained LightGBM short-time strong precipitation classification model by adopting a test set in the data set of the short-time strong precipitation event classification task in the step 3.1, and obtaining a probability prediction result of the short-time strong precipitation event from hour to hour within 6-18 hours.
In order to illustrate the improvement effect of the algorithm, in the embodiment, a plurality of prediction schemes are adopted to carry out a short-time strong precipitation probability prediction test. The selection scheme is as follows:
scheme one: and obtaining the predicted result of the 6-18 hours of precipitation only by using the regional weather numerical mode. And then converting the predicted precipitation result from hour to hour into the predicted probability result of the strong precipitation by adopting a trained short-time strong precipitation classification model.
Scheme II: and (3) adopting a precipitation time sequence prediction model based on a cyclic convolution neural network, and repeating the training and verification process in the second step by taking 24 forecast variables output by the regional weather numerical mode as input factors. And obtaining an hour rainfall prediction result for 6-18 hours by using the trained rainfall time sequence prediction model. And then, converting the prediction of the hour precipitation into a probability prediction result of the hour strong precipitation event by adopting the short-time strong precipitation classification model trained in the step three.
Scheme III: the precipitation time sequence prediction model and the short-time strong precipitation classification model which are trained in the second step and the third step and based on the cyclic convolutional neural network are adopted to obtain the probability prediction result of the short-time strong precipitation for 6-18 hours. In contrast to scheme two, scheme three adds 0-5 hours of ground site hour precipitation observation data to the input factor of step two.
The short-time strong precipitation probability prediction results of each regimen were evaluated by a receiver operating characteristic curve (Receiver Operating Characteristic curve, ROC curve for short) and a brier skill score (Brier Skill Score, BSS for short).
ROC curves have the void fraction (PFOD) as the abscissa and the hit fraction (POD) as the ordinate. Quantitative analysis was performed using AUC (Area Under ROC Curve) area, the size of the area under the ROC curve. The prediction effect is better when the AUC is between 0.5 and 1 and is closer to 1. And (3) to (4) are shown, wherein the number of short-time strong precipitation is correctly identified, the number of non-short-time strong precipitation is incorrectly identified as the number of short-time strong precipitation, the number of non-short-time strong precipitation is correctly identified, and the number of short-time strong precipitation is incorrectly identified as the number of non-short-time strong precipitation.
(3)
(4)
The Boolean skill Score (Brier Skill Score, BSS) is calculated based on the Boolean Score (Brier Score, BSS) to measure model predictive power of the second and third solutions relative to the first solutionLifting. The closer the BSS value is to 1, the better the forecasting effect. As shown in formulas (5) - (6), wherein +.>For scheme one->，/>Can represent schemes II and III>，/>Predicting the total number of samples->For predicting probability (0-1), +.>To observe the fact (0 or 1).
(5)
(6)
The prediction effect evaluation is shown in fig. 3. The overall ROC curve for scheme one is lowest, followed by scheme two, and the ROC curve for scheme three is highest, as in (a) of fig. 3. The AUC of the third scheme is the highest in different forecasting time periods, and the maximum value of the AUC reaches 0.96; second, scheme two, AUC max 0.85; the AUC value for scheme one is the lowest and the AUC maximum is only 0.84. Compared with the scheme I, the BSS of the scheme II is 0.15-0.18, which shows that the scheme II has certain improvement. The maximum BSS value of the scheme III reaches 0.35 and the minimum BSS value reaches 0.25, which shows that the forecasting effect of the scheme III is greatly improved. The comparison of the prediction results of various schemes shows that the short-time strong precipitation probability prediction model has the best effect.
Example predictive effects of the three schemes are compared as follows:
four rows from top to bottom in fig. 4 are observation live, scenario three, scenario two, scenario one, respectively. The predictions of different durations are listed from left to right 5. Observing that the live black and white filling color is the real precipitation distribution, wherein the black and white filling colors of the first scheme, the second scheme and the third scheme are the occurrence probability of short-time strong precipitation, and the black cross is the real short-time strong precipitation occurrence distribution.
As can be seen from the analysis in fig. 4, the probability distribution of short-time strong precipitation predicted by the first scheme and the second scheme is relatively average, and is between 50% and 75%, and the probability distribution of real short-time strong precipitation is inconsistent. According to the technical scheme (scheme III), the prediction probability (90% -95%) of short-time strong rainfall is greatly improved, the prediction probability is consistent with the real prediction height, and the prediction capacity is high.
To sum up:
compared with the traditional numerical forecasting method, the technical scheme of the invention utilizes the cyclic convolution neural network to conduct future 12-hour (6-18 hours in the invention) rainfall extrapolation forecasting based on the multiscale atmospheric thermal power elements of regional numerical mode forecasting, the effect of rainfall forecasting is improved by designing a loss function with weight, the BSS score is between 0.15 and 0.19, and the average AUC is improved from 0.82 to 0.83.
Compared with the thermodynamic elements predicted by using the regional numerical mode, the technical scheme provided by the invention can enable the cyclic convolution neural network to learn the time and space information of the evolution of the precipitation system more fully by observing the ground precipitation at the moment before adding, and greatly improve the probability prediction capability of short-time strong precipitation. BSS peak scores increased from 0.19 to 0.35 and average AUCs increased from 0.83 to 0.88.
According to the technical scheme, a two-step strategy is adopted, the short-time prediction problem of the short-time strong precipitation is disassembled into two tasks, namely a precipitation prediction task based on a cyclic convolutional neural network for 12 hours (6-18 hours in the invention) in the future, and a short-time strong precipitation classification task based on the LightGBM. The two-step strategy can exert the advantages of the cyclic convolutional neural network in the aspect of large sample time sequence prediction, and can also fully exert the advantages of the LightGBM in the classification task of short-time strong precipitation such as small probability events.
The cyclic convolution neural network can simultaneously utilize time and space information to extract rules and trends in the data from the high-dimensional data, and can more accurately realize forecasting. On the other hand, the short-time strong precipitation is a small probability event, so that the advantages of the LightGBM in a small sample classification task can be fully exerted, and a short-time probability forecasting result of the short-time strong precipitation is directly formed. The multiscale atmospheric circulation thermodynamic element provided by the regional numerical mode is used as a prediction factor, so that more movement and evolution information of a precipitation system can be provided for the cyclic convolution neural network. By adding the ground site hour precipitation observation data of 0-5 hours in the past, the cyclic convolution neural network can learn the time and space information of the evolution of a precipitation system, and the probability prediction capability of short-time strong precipitation is greatly improved.
Claims (10)
1. A short-time strong precipitation probability prediction fusion method based on a cyclic convolution neural network is characterized by comprising the following steps:
step one, establishing a fusion data set of short-time prediction of precipitation, and dividing the fusion data set into a training set, a verification set and a test set of short-time prediction of precipitation;
training a precipitation time sequence prediction model based on a cyclic convolutional neural network by using the precipitation short-time prediction training set and the verification set in the first step; a trained precipitation time sequence prediction model based on a cyclic convolutional neural network is adopted to obtain an hour precipitation prediction result of 6-18 hours in the future;
training the LightGBM classification model by adopting the precipitation space-time prediction result obtained in the step two to obtain a short-time strong precipitation classification model, and obtaining a probability prediction result of a short-time strong precipitation event every hour for 6-18 hours in the future.
2. The method for fusion of short-time strong precipitation probability forecast based on cyclic convolutional neural network according to claim 1, wherein the first step specifically comprises the following steps:
step 1.1, collecting hour-by-hour ground site hour rainfall observation data and 0-18 hours multivariate data of regional weather numerical model forecast;
step 1.2, carrying out time matching on the hour-by-hour ground site hour precipitation observation data and the regional weather pattern forecast data collected in the step 1.1 to form a data sequence;
step 1.3, eliminating all samples without precipitation aiming at the data sequence formed in the step 1.2;
step 1.4, carrying out quality control, comparison evaluation and normalization processing on the data sequence formed in the step 1.3;
step 1.5, performing space fusion on the data sequence after quality control formed in the step 1.4;
and step 1.6, obtaining a fusion data set of short-time rainfall prediction fused with the ground hour rainfall observation data and the regional weather numerical forecast data through the steps 1.1-1.5, and dividing the fusion data set into a short-time rainfall prediction training set, a verification set and a test set.
3. The method according to claim 2, wherein in the step 1.1, the collected forecast variables of the regional weather numerical model are 24, including temperature, humidity, wind direction and wind direction at 925, 850, 700, 500 and 200 hPa, and air pressure, energy of unsteady convection, 2 m temperature and water precipitation amount at the ground floor.
4. The short-time strong precipitation probability prediction fusion method based on the cyclic convolutional neural network according to claim 3, wherein the specific operation of the step 1.2 is as follows: the regional weather numerical mode starts from 0 hour a day and outputs an hour-by-hour atmospheric element forecast of 0-18 hours in the future every 3 hours; matching the ground site hour precipitation observation data into regional weather pattern prediction data, namely forming a data sequence containing 24 pattern prediction variables and ground site hour precipitation observation data every 3 hours, wherein the time interval of the sequence is 1 hour, and the sequence length is 0-18 hours; wherein 0-5 hours is used as a predictor input sequence and 6-18 hours is used as a predictor test sequence.
5. The short-time strong precipitation probability prediction fusion method based on the cyclic convolutional neural network according to claim 4, wherein the specific operation method of the step 1.4 is as follows:
firstly, performing quality control, removing data sequences with errors or suspicious observations, and removing data sequences with missing ground hour precipitation observations;
then, carrying out comparison evaluation, accumulating the hour precipitation observation data, comparing the hour precipitation observation data with the day hour precipitation observation data, and if the two are inconsistent, eliminating all data sequences in the same day;
and finally, respectively carrying out normalization processing on each element in the data sequence.
6. The short-time strong precipitation probability prediction fusion method based on the cyclic convolutional neural network, which is characterized in that the spatial fusion method in the step 1.5 specifically comprises the following steps: and (3) interpolating the ground site hour precipitation observation data of 0-18 hours and the multivariable forecast data of the weather mode in the area of 0-18 hours into uniform grid coordinates by adopting an inverse distance weighted interpolation method, wherein the horizontal resolution is 0.05 degrees, and the grid points are 120 multiplied by 120.
7. The short-time strong precipitation probability prediction fusion method based on the cyclic convolutional neural network according to claim 1, wherein the specific steps of the second step are as follows:
step 2.1, the adopted rainfall time sequence prediction model is a PredRNNv2 model with four layers of space-time long-short-term memory modules;
training the rainfall time sequence prediction model by using the rainfall short-time prediction training set and the verification set obtained in the step one; the input factors of 0-5 hours are 25, and the input factors comprise 24 variables of ground hour rainfall observation and regional weather numerical model prediction; the input factors of 6-18 hours do not contain ground hour precipitation observation, and only 24 variables for regional weather numerical mode prediction are included; the true value is the observed data of the hour precipitation;
an adaptive moment estimation optimizer is used in the training process, the batch size of each training is 8, the learning rate is 0.001, and the maximum training iteration number is 300; the weighted average absolute error is used as a loss function of precipitation prediction, and the formula is as follows:
（1），
wherein the method comprises the steps ofIs the total sample size, +.>Is the real value of the hour precipitation intensity of the ith sample,/>Is the precipitation intensity predictive value of the ith sample,/->Is the input factor set for the i-th sample; />Is the precipitation weight coefficient of the ith sample, and the weight is set as follows:
，
the accuracy rate of each training and verification is output in the training process, and the model with the highest verification accuracy is automatically stored;
step 2.2, selecting a model with highest verification precision, namely a trained precipitation time sequence prediction model based on a cyclic convolutional neural network;
step 2.3, saving the model to predict the result of the short-time precipitation prediction for 6-18 hours on the short-time precipitation prediction verification set; and testing the model on the short-time rainfall prediction test set to obtain the hour-by-hour rainfall space-time prediction result of 6-18 hours on the short-time rainfall prediction test set.
8. The short-time strong precipitation probability prediction fusion method based on the cyclic convolutional neural network according to claim 1, wherein the specific steps of the third step are as follows:
step 3.1, a data set is manufactured again aiming at a short-time strong precipitation event classification task of a machine learning model LightGBM, and the data set is split into a training set, a verification set and a test set;
step 3.2, training the LightGBM model by adopting a training set in the data set of the short-time strong precipitation event classification task in step 3.1; outputting the accuracy rate of each training in the training process, and automatically storing a model with the highest accuracy rate as a trained LightGBM short-time strong precipitation classification model;
and 3.3, testing the trained LightGBM short-time strong precipitation classification model by adopting a test set in the data set of the short-time strong precipitation event classification task in the step 3.1, and obtaining a probability prediction result of the short-time strong precipitation event from hour to hour within 6-18 hours.
9. The method for predicting and fusing short-time strong precipitation probability based on a cyclic convolutional neural network according to claim 8, wherein the step 3.1 of reconstructing a data set for a short-time strong precipitation event classification task of a machine learning model LightGBM comprises the steps of:
a) Extracting 6-18 hours forecast test sequence data in a verification set and a test set from the short-time forecast fusion data set of the precipitation formed in the first step, and respectively serving as a training set and a test set of the short-time strong precipitation classification data set; the training set of the short-time strong precipitation classification data set and the ground site hour precipitation observation data in the test set are only used as real fields, and are not used as input factors of short-time strong precipitation classification tasks;
b) Adding the 6-18 hour rainfall space-time prediction results obtained in the short-time rainfall prediction verification set and the test set on the test pre-sequencing column into the training set and the test set of the short-time strong rainfall classification data set respectively;
c) Respectively eliminating the hour precipitation quantity predicted by the regional weather numerical mode from a training set and a testing set of the short-time strong precipitation classification data set;
d) According to a short-time strong precipitation threshold value of 20mm/h, converting training sets of short-time strong precipitation classification data sets and ground site hour precipitation observation data and precipitation space-time prediction results in a test set into short-time strong precipitation marks, namely, when the hour precipitation amount on a grid point is greater than 20mm, 1 is smaller than 20mm and 0 is smaller than 1, so that a data set of short-time strong precipitation event classification tasks is formed; the training set of the short-time strong precipitation classification data set corresponds to the training set in the data set forming the short-time strong precipitation event classification task after conversion, and the testing set of the short-time strong precipitation classification data set corresponds to the testing set in the data set forming the short-time strong precipitation event classification task.
10. The method for fusion of short-time strong precipitation probability prediction based on cyclic convolutional neural network according to claim 8, wherein in step 3.2
The input factors for the LightGBM classification model are as follows:
a) 23 forecast variables for regional weather numerical patterns including temperature, humidity, warp, weft, and barometric pressure at ground level, convective instability energy, 2 meters temperature at 925, 850, 700, 500, and 200 hPa;
b) 0-1 distribution of short-time strong precipitation events formed by precipitation space-time prediction results;
c) The real field of the LightGBM classification model is the 0-1 distribution of short-time strong precipitation events of the ground site hour precipitation observation data;
during training, automatically adjusting parameters by using a Bayesian optimization method, and then continuing parameter fine adjustment to enable an objective function to reach the global minimum; the specific parameter scheme is as follows: adopting a single-side sampling algorithm based on gradient, wherein the learning rate is 0.01, the depth of each training tree is 2, the maximum leaf number is 25, the maximum iteration number is 5000, the positive and negative sample proportion is automatically adjusted, the model back screening proportion is 50%, and the maximum barrel number is 100; the binary cross entropy is used as a Loss function (Loss) of a short-time strong precipitation event classification task, and the formula is as follows:
（2），
wherein the method comprises the steps ofIs the total sample size, +.>Is the ith sampleShort-time strong precipitation classification true value, +.>Is the short-time strong precipitation classification predictive value of the ith sample,/->Is the input factor set for the i-th sample.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311584724.0A CN117290810B (en) | 2023-11-27 | 2023-11-27 | Short-time strong precipitation probability prediction fusion method based on cyclic convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311584724.0A CN117290810B (en) | 2023-11-27 | 2023-11-27 | Short-time strong precipitation probability prediction fusion method based on cyclic convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117290810A true CN117290810A (en) | 2023-12-26 |
| CN117290810B CN117290810B (en) | 2024-02-02 |
Family
ID=89239407
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311584724.0A Active CN117290810B (en) | 2023-11-27 | 2023-11-27 | Short-time strong precipitation probability prediction fusion method based on cyclic convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117290810B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117933316A (en) * | 2024-03-25 | 2024-04-26 | 南京大学 | Groundwater level probability forecasting method based on interpretable Bayesian convolution network |
| CN117933316B (en) * | 2024-03-25 | 2024-05-31 | 南京大学 | Groundwater level probability forecasting method based on interpretable Bayesian convolution network |
Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111815037A (en) * | 2020-06-24 | 2020-10-23 | 河海大学 | Interpretable short-critical extreme rainfall prediction method based on attention mechanism |
| CN112651463A (en) * | 2021-01-05 | 2021-04-13 | 天津大学 | Construction method of double-forecast model of hail weather in plateau area |
| CN114609698A (en) * | 2022-01-20 | 2022-06-10 | 成都信息工程大学 | Short-time strong precipitation probability forecasting method based on proportioning method idea and dichotomy |
| US20220221616A1 (en) * | 2021-01-08 | 2022-07-14 | Innovim, Llc | Landfalling event atmospheric river neural network (learn2) forecast tool |
| CN115545159A (en) * | 2022-09-26 | 2022-12-30 | 中国人民解放军国防科技大学 | Average sea surface temperature forecasting method of deep neural network |
| CN115860268A (en) * | 2023-02-20 | 2023-03-28 | 南京信息工程大学 | Short-time strong precipitation prediction method based on CNN-XGBOOST model |
| CN115933008A (en) * | 2022-11-22 | 2023-04-07 | 广东电网有限责任公司广州供电局 | Strong convection weather forecast early warning method |
| CN116628626A (en) * | 2023-06-19 | 2023-08-22 | 成都信息工程大学 | Short-time strong precipitation typing method based on feature cross fusion |
| CN116644848A (en) * | 2023-05-24 | 2023-08-25 | 上海能源科技发展有限公司 | Multi-source data-based medium-long term area wind power prediction method |
| CN116865251A (en) * | 2023-07-04 | 2023-10-10 | 西安交通大学 | Short-term load probability prediction method and system |
| CN116911178A (en) * | 2023-07-10 | 2023-10-20 | 浙江省水利水电勘测设计院有限责任公司 | Method and system for predicting capacity of small and medium-sized reservoirs based on weather forecast |
| CN117008217A (en) * | 2023-07-11 | 2023-11-07 | 哈尔滨工程大学 | Short-time strong precipitation prediction method integrating self-attention module and Unet model |
| CN117111181A (en) * | 2023-09-05 | 2023-11-24 | 浙江省气象台 | Short-time strong precipitation probability prediction method and system |
-
2023
- 2023-11-27 CN CN202311584724.0A patent/CN117290810B/en active Active
Patent Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111815037A (en) * | 2020-06-24 | 2020-10-23 | 河海大学 | Interpretable short-critical extreme rainfall prediction method based on attention mechanism |
| CN112651463A (en) * | 2021-01-05 | 2021-04-13 | 天津大学 | Construction method of double-forecast model of hail weather in plateau area |
| US20220221616A1 (en) * | 2021-01-08 | 2022-07-14 | Innovim, Llc | Landfalling event atmospheric river neural network (learn2) forecast tool |
| CN114609698A (en) * | 2022-01-20 | 2022-06-10 | 成都信息工程大学 | Short-time strong precipitation probability forecasting method based on proportioning method idea and dichotomy |
| CN115545159A (en) * | 2022-09-26 | 2022-12-30 | 中国人民解放军国防科技大学 | Average sea surface temperature forecasting method of deep neural network |
| CN115933008A (en) * | 2022-11-22 | 2023-04-07 | 广东电网有限责任公司广州供电局 | Strong convection weather forecast early warning method |
| CN115860268A (en) * | 2023-02-20 | 2023-03-28 | 南京信息工程大学 | Short-time strong precipitation prediction method based on CNN-XGBOOST model |
| CN116644848A (en) * | 2023-05-24 | 2023-08-25 | 上海能源科技发展有限公司 | Multi-source data-based medium-long term area wind power prediction method |
| CN116628626A (en) * | 2023-06-19 | 2023-08-22 | 成都信息工程大学 | Short-time strong precipitation typing method based on feature cross fusion |
| CN116865251A (en) * | 2023-07-04 | 2023-10-10 | 西安交通大学 | Short-term load probability prediction method and system |
| CN116911178A (en) * | 2023-07-10 | 2023-10-20 | 浙江省水利水电勘测设计院有限责任公司 | Method and system for predicting capacity of small and medium-sized reservoirs based on weather forecast |
| CN117008217A (en) * | 2023-07-11 | 2023-11-07 | 哈尔滨工程大学 | Short-time strong precipitation prediction method integrating self-attention module and Unet model |
| CN117111181A (en) * | 2023-09-05 | 2023-11-24 | 浙江省气象台 | Short-time strong precipitation probability prediction method and system |
Non-Patent Citations (7)
| Title |
|---|
| ANGELICA N. CASERI 等: "A convolutional recurrent neural network for strong convective rainfall nowcasting using weather radar data in Southeastern Brazil", 《ARTIFICIAL INTELLIGENCE IN GEOSCIENCES》, pages 1 - 6 * |
| WU ZHI-PENG 等: "A COMBINED VERIFICATION METHOD FOR PREDICTABILITY OF PERSISTENT HEAVY RAINFALL EVENTS OVER EAST ASIA BASED ON ENSEMBLE FORECAST", 《JOURNAL OF TROPICAL METEOROLOGY》, pages 1 - 12 * |
| XIAORAN ZHUANG 等: "Spatial Predictability of Heavy Rainfall Events in East China and the Application of Spatial-Based Methods of Probabilistic Forecasting", 《ATMOSPHERE》, pages 1 - 19 * |
| 庄潇然 等: "基于深度学习的融合降水临近预报方法及其在中国东部地区的应用研究", 《气象学报》, pages 286 - 303 * |
| 王清霖: "基于机器学习的短临雷雨大风预测研究", 《中国优秀硕士学位论文全文数据库 (基础科学辑)》, pages 009 - 116 * |
| 郑玉 等: "基于循环神经网络改进雷达定量估测强降水", 《中国科技论文》, vol. 15, no. 5, pages 585 - 592 * |
| 郑玉: "基于深度学习的能见度与降水临近预报技术研究", 《中国优秀硕士学位论文全文数据库 (基础科学辑)》, pages 009 - 179 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117933316A (en) * | 2024-03-25 | 2024-04-26 | 南京大学 | Groundwater level probability forecasting method based on interpretable Bayesian convolution network |
| CN117933316B (en) * | 2024-03-25 | 2024-05-31 | 南京大学 | Groundwater level probability forecasting method based on interpretable Bayesian convolution network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117290810B (en) | 2024-02-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113468803B (en) | WOA-GRU flood flow prediction method and system based on improvement | |
| CN113496104A (en) | Rainfall forecast correction method and system based on deep learning | |
| CN113837499A (en) | Ultra-short-term wind power prediction method and system | |
| CN112561058A (en) | Short-term photovoltaic power prediction method based on Stacking-ensemble learning | |
| CN116128141B (en) | Storm surge prediction method and device, storage medium and electronic equipment | |
| CN114462718A (en) | CNN-GRU wind power prediction method based on time sliding window | |
| CN112069955B (en) | Typhoon intensity remote sensing inversion method based on deep learning | |
| CN113592132B (en) | Rainfall objective forecasting method based on numerical weather forecast and artificial intelligence | |
| CN113554466A (en) | Short-term power consumption prediction model construction method, prediction method and device | |
| CN114611608A (en) | Sea surface height numerical value prediction deviation correction method based on deep learning model | |
| CN114882373A (en) | Multi-feature fusion sandstorm prediction method based on deep neural network | |
| CN116525135B (en) | Method for predicting epidemic situation development situation by space-time model based on meteorological factors | |
| CN117290810B (en) | Short-time strong precipitation probability prediction fusion method based on cyclic convolutional neural network | |
| CN115952924B (en) | Short-term wind power prediction method based on optimized VMD and characteristic thermodynamic diagram | |
| CN116187540B (en) | Wind power station ultra-short-term power prediction method based on space-time deviation correction | |
| CN111488974A (en) | Deep learning neural network-based ocean wind energy downscaling method | |
| CN116663404A (en) | Flood forecasting method and system coupling artificial intelligence and Bayesian theory | |
| CN103268416A (en) | Wind speed prediction method based on STCP-BP | |
| CN116167508A (en) | Short-term photovoltaic output rapid prediction method and system based on meteorological factor decomposition | |
| CN115860165A (en) | Neural network basin rainfall runoff forecasting method and system considering initial loss | |
| CN114372615A (en) | Short-term photovoltaic power prediction method and system considering spatial correlation | |
| CN113537573A (en) | Wind power operation trend prediction method based on dual space-time feature extraction | |
| CN114091741A (en) | Typhoon forecast prediction method and device | |
| CN117909888B (en) | Intelligent artificial intelligence climate prediction method | |
| CN116992781B (en) | Multi-step multi-element storm forecasting method based on deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |