CN110459036B - Mountain torrent early warning method based on deep learning - Google Patents
Mountain torrent early warning method based on deep learning Download PDFInfo
- Publication number
- CN110459036B CN110459036B CN201910846384.1A CN201910846384A CN110459036B CN 110459036 B CN110459036 B CN 110459036B CN 201910846384 A CN201910846384 A CN 201910846384A CN 110459036 B CN110459036 B CN 110459036B
- Authority
- CN
- China
- Prior art keywords
- layer
- early warning
- water level
- input
- output
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G08—SIGNALLING
- G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
- G08B21/00—Alarms responsive to a single specified undesired or abnormal condition and not otherwise provided for
- G08B21/02—Alarms for ensuring the safety of persons
- G08B21/10—Alarms for ensuring the safety of persons responsive to calamitous events, e.g. tornados or earthquakes
-
- 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
- Y02A50/00—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE in human health protection, e.g. against extreme weather
Abstract
The invention discloses a mountain torrent early warning method based on deep learning, which adopts an RNN model with an LTSM function, and adds a long-short term memory (LTSM) method into a model algorithm to effectively solve the problems of gradient dissipation and gradient explosion in the RNN, so that the forecasting of small watersheds or mountain torrent areas is more effectively carried out, the reliable mountain torrent disaster early warning forecasting is realized, the forecasting time is greatly improved, the model only needs to pay attention to input data, and the method has the characteristics of simple and convenient operation, high operation speed and convenience for shifting values to other areas.
Description
Technical Field
The invention relates to the field of disaster early warning, in particular to a mountain torrent early warning method based on deep learning.
Background
The dynamic early warning of the torrential flood is realized by early warning sent out to the torrential flood danger area. By 2018, China has defined torrential flood dangerous areas nationwide, and more than 28000 torrential flood dangerous areas exist in Sichuan provinces. The early warning of the mountain torrents is carried out by setting early warning indexes of rainfall or water level at present. The specific method comprises the steps of setting rainfall or water level stations in the area, associating torrential flood dangerous areas (a plurality of torrential flood dangerous areas) to corresponding measuring stations, and carrying out early warning through actually measuring rainfall (water level). In order to prolong the early warning time and improve the early warning precision, a hydrological model is used for carrying out mountain torrent early warning in some areas, but the early warning mode has some limitations.
Models can be classified into predictive models and descriptive models by the function of the hydrological model. At present, a prediction type hydrological model is mostly adopted in the mountain torrent early warning. The predictive hydrological model can be classified into a lumped hydrological model for researching the global water circulation rule and a distributed hydrological model for considering the space-time difference of hydrological elements in the domain according to different application technologies, and can also be classified into a lumped model, a distributed hydrological model and a semi-distributed hydrological model. The lumped model comprises an inference formula method widely used in China, a Xinanjiang model, an American SAC-SMA model, a Swedish HBV model and the like. A distributed hydrological model is a model with strict physical significance in links such as runoff yield, confluence, base runoff and the like on the basis of obtaining high-precision DEM data, land utilization data, soil data, remote sensing data and the like.
The two hydrological models have strong physical significance, but have obvious limitation when being applied to mountain torrent early warning. First, theoretical structural imperfection, two kinds of models are to the high generalization of hydrology phenomenon, to actual precipitation and evaporation, precipitation and production flow, ground damps and oozes etc. physical process theoretical still has the defect. Secondly, the model parameters are too many, and in the actual parameter adjusting process, in order to better fit the actual situation, part of the parameters lose the physical significance after being adjusted. Thirdly, the fault tolerance is poor, and the model is difficult to construct for the regions with poor data or no data. And fourthly, the forecasting time is limited by the drainage basin converging time, the forecasting time for small drainage basins (mountain flood areas) is short, and even if forecasting is carried out, a disaster reduction department cannot respond in time. And fifthly, the operation process is complex, the speed is low, the operation difficulty is high, the model building process is a process of simulating the flood of the drainage basin, the model building process is very complex, although the model is generalized, still has numerous parameters and huge calculation amount, and some models need to use different parameters under different conditions, so that the generation result is slow, the operation is not easy, and the application is not easy to popularize.
Disclosure of Invention
The invention aims to overcome the defects of the traditional technology and provides a mountain torrent early warning method based on deep learning, which realizes early-stage prediction of mountain torrent early warning.
The aim of the invention is achieved by the following technical measures:
a mountain torrent early warning method based on deep learning comprises the following steps:
the first step is as follows: determining an early warning object: drawing a drainage basin where the early warning area is located on a GIS map, and determining a mountain torrent early warning area in the drainage basin and a preset early warning rainfall or water level, a distributed rainfall and a water level monitoring station thereof;
the second step is that: collecting the whole data: collecting historical data of rainfall, water level stations, regional evaporation capacity monitoring stations and county meteorological early warning in an early warning area, and calculating drainage basin convergence time;
the basin confluence time calculation method is inference formula calculation or Thiessen polygon calculation or the combination determination of the two.
The reasoning formula calculation method comprises the following steps: selecting a typical flood process in a historical period of a drainage basin water level control station, and calculating by using an inference formula, wherein the calculation formula is as follows:
in the formula, L is the distance from a watershed water distribution line to a control station, Q is the typical flood peak flow of the control station, J is the average watershed drop, and m is an empirical convergence coefficient.
The calculation method of the Thiessen polygon method comprises the following steps: converging the rainfall of the drainage basin surface and the drainage basin water level station on a time axis to form a surface rainfall and drainage basin control station flow timing diagram, and obtaining convergence time by adopting a secondary flood segmentation method.
The third step: monitoring station cross-correlation analysis: classifying the monitoring points according to rainfall and water level or flow, respectively calculating the cross-correlation coefficient of a certain station and the like according to a cross-correlation coefficient formula, and when the measured data of two stations cannot be obtained or the difference between the measured values of the two stations and the measured values of the related stations exceeds the maximum difference value N1 times of the historical series, calculating a result to replace the measured values through the measured data replacing formula;
the cross correlation coefficient formula calculation method comprises the following steps: let the site sequences be { x, respectivelyt1,2, …, n and ytI, t ═ 1,2, …, n }, and their cross-correlation coefficients are constructed as:
wherein, the closer the calculated correlation coefficient is to 1, the stronger the correlation between the two sites is represented. When the measured data of two stations cannot be obtained or the measured value of the two stations is too different from the measured value of the relevant station, if the maximum difference value N1 times of the historical series is taken, N1 is generally selected to be 1, and the measured data is considered to be possibly wrong.
The measured data substitution formula is as follows:
in the formula, A is the basin area controlled by the monitoring station, the rainfall station is the area cut by the Thiessen polygon, and the water level station is the basin area controlled.
Is the mean of the x samples and is,
is the y sample mean.
The fourth step: early warning of rainfall: an RNN rainfall forecasting model with an LTSM function is established, and mountain torrent rainfall early warning forecasting is carried out by adopting integrated rainfall forecasting calculation;
the method for performing the mountain torrent rainfall early warning forecast by integrating the rainfall forecast calculation comprises the following steps:
s4-1, calculating forecast accuracy evaluation index through forecast evaluation formula,
the forecast evaluation formula is as follows:
in the formula, E is the evaluation accuracy, m is the number of times that the predicted value and the measured value are in the same interval, and n is the total number of times of prediction.
S4-2, determining input layer and output layer data:
the input layer includes four factors, respectively: the real-time station observation value at the time t, the forecast value of the station at the time t + i, the forecast value of the provincial meteorological department grid forecast at the time t + i converted to the station by the professional meteorological company and the rainstorm early warning value issued by the local meteorological department at the time t;
wherein, the rainstorm early warning value that local meteorological department's t instant was released respectively is: no 0 is issued, blue warning is 1, yellow is 2, orange is 3, red is 4.
The output layer is the measured value of the station at the time t + i;
s4-3, establishing an RNN rainfall forecasting model with the LTSM function:
constructing an RNN model with an LTSM function, wherein the RNN model comprises an input layer and a hidden layer, and the calculation formula is as follows:
ot+1＝V·st+1+bo
st+1＝f(U·xt+1+W·st+bs)
ot+1＝V·st+1+bo
in the above formula, U is the connection weight from the input layer to the hidden layer, V is the connection weight from the hidden layer to the output layer, W is the connection weight from the previous hidden layer to the current hidden layer, and xtFor inputting layer data, StHidden layer computation output data, otOutput layer calculates output data, bsTo hide the layer offset number, boFor the output layer bias number, f () is the excitation function,
wherein the hidden layer of the RNN comprises a short-term state s sensitive to short-term input and a structure c for storing a long-term state, the long-term state stored by the structure c is a unit state, the structure c comprises a forgetting gate, an input gate and an output gate for controlling the unit state,
wherein the content of the first and second substances,
the forgetting gate determines how much the unit state at the previous moment is reserved to the current moment ct;
the input gate determines the input of the network at the current moment and how many units are stored in the unit state;
how many output values of the states of the output door control unit are output to the hidden layer value of the current LSTM;
the calculation formula is as follows:
ft＝σ(Wt·[ht-1,xt]+bf)
it＝σ(Wi·[ht-1,xt]+bi)
ot＝σ(Wo·[ht-1,xt]+bo)
ht＝ot·tanh(ct)
in the above formula, ftIndicating forgetting gate, itRepresentation input gate, ctIndicating the state of the cell at time t, otDenotes the output gate, W denotes the weight, b denotes the offset, htExpressing the output at the time t, wherein sigma is an activation function, and tanh is a hyperbolic tangent activation function;
the hidden layer comprises two hidden layers which are respectively a first hidden layer and a second hidden layer, wherein the first hidden layer adopts the same number of units as the input layer, and the second hidden layer adopts half of the number of input factors as the number of hidden units for calculation;
s4-4, training and verifying: dividing collected and sorted historical measured rainfall and forecast data into a training set and a verification set, and training model parameters by using the training set so as to verify and calculate forecast accuracy by using the verification set;
wherein the data volume ratio of the training set to the validation set is 4: 1; the calculation forecasting precision requirement of the verification set reaches 80% or more.
S4-5, calculating real-time data: and substituting the real-time data into the calculation model to perform mountain torrent rainfall early warning and forecasting.
The fifth step: water level/flow early warning: and establishing an RNN water level forecasting model with an LTSM function, substituting real-time data into the model for calculation, and realizing mountain torrent water level early warning and forecasting.
The steps of mountain torrent water level early warning and forecasting are as follows:
s5-1, determining the early warning water level: determining an early warning water level according to a regional flood prevention plan or a hydrologic station water level amplitude variation correlation relation;
the method for determining the early warning water level through the hydrologic station water level amplitude variation correlation relationship comprises the following steps:
acquiring water level values of different flood frequencies of hydrological stations with long-sequence observed values in a flow area, and calculating water level amplitude difference delta H of simple water level stations set in a dangerous area in the same field1,ΔH2And drawing a water level amplitude variation correlation diagram of the two stations, thereby determining the early warning threshold values of the simple water level station corresponding to different torrential flood dangerous areas.
S5-2, determining an input layer and an output layer: the input layer and the output layer can be determined by adopting an actual measurement data method or a forecast time method;
the method for determining the input layer and the output layer by the actual measurement data method comprises the following steps: adopting measured data, wherein an input layer is a rainfall station measured value, an evaporation capacity measured value and a control station measured value above a control station, and an output layer is a control station water level value after confluence time tau;
the method for determining the input layer and the output layer by the forecast time method comprises the following steps: setting the rainfall forecast time as n times of the convergence time tau, setting the rainfall forecast values, the forecast air temperature values and the forecast values of the control station on the input layer, and setting the water level value of the post-control station when the output layer is n tau jf.
S5-3, establishing an RNN water level forecasting model with LTSM function, wherein the RNN model comprises an input layer and a hidden layer, and the calculation formula is as follows:
ot+1＝V·st+1+bo
st+1＝f(U·xt+1+W·st+bs)
ot+1＝V·st+1+bo
in the above formula, U is the connection weight from the input layer to the hidden layer, V is the connection weight from the hidden layer to the output layer, W is the connection weight from the previous hidden layer to the current hidden layer, and xtFor input layer data, stHidden layer computation output data, otOutput layerCalculating output data, bsTo hide the layer offset number, boFor the output layer offset number, f () is the excitation function,
wherein the hidden layer of the RNN comprises a short-term state s sensitive to short-term input and a structure c for storing a long-term state, the long-term state stored by the structure c is a unit state, the structure c comprises a forgetting gate, an input gate and an output gate for controlling the unit state,
wherein the content of the first and second substances,
the forgetting gate determines how much the unit state at the previous moment is reserved to the current moment ct;
the input gate determines the input of the network at the current moment and how many input nodes are stored in the unit state;
how many output values of the states of the output door control unit are output to the hidden layer value of the current LSTM;
the calculation formula is as follows:
ft＝σ(Wt·[ht-1,xt]+bf)
it＝σ(Wi·[ht-1,xt]+bi)
ot＝σ(Wo·[ht-1,xt]+bo)
ht＝ot·tanh(ct)
in the above formula, ftIndicating forgetting gate, itRepresentation input gate, ctIndicating the state of the cell at time t, otDenotes the output gate, W denotes the weight, b denotes the offset, htExpressing the output at the time t, wherein sigma is an activation function, and tanh is a hyperbolic tangent activation function;
the hidden layers comprise a plurality of hidden layers, the number of units of a first hidden layer in the plurality of hidden layers is the same as that of input layers, and the number of the input factors of the previous hidden layer is half that of the input factors of the second hidden layer and the hidden layers behind the second hidden layer in the plurality of hidden layers is used as the number of the hidden units for calculation until the number of the input factors of the last hidden layer is less than or equal to 4;
the number of the hidden layers is determined by halving the number of input factors of the next hidden layer until the number of input factors of the last hidden layer is less than or equal to 4.
S5-4, establishing a training set and a verification set, wherein the training precision takes the standard that the difference between the predicted value and the actually measured value in the verification set is qualified.
Because the water level series data are longer, the data collected and sorted in the invention are divided into two data sets of training and verification, and the data are divided into two data sets according to the series length of 5: 1 is divided.
The qualified standard of the training precision is as follows: the difference between the predicted value and the actual measured value is qualified when the difference is less than or equal to 20cm in a high water area, and is qualified when the difference is less than or equal to 50cm in a low water area.
And S5-5, substituting the real-time data into the model for calculation, and realizing the early warning and forecasting of the mountain torrent water level.
The prevention and control of the mountain torrent disasters are one of the difficulties in the field of disaster prevention and reduction in China, the difference of mountain torrent disaster areas in China is obvious, particularly the difference between the south and the north is obvious, and a reliable mountain torrent disaster early warning and forecasting technology is not available so far. Sichuan is one of the most provinces of mountain torrent disasters in China, more than 4000 counties (cities and districts) in total are threatened by the mountain torrent disasters, the total area of the mountain torrent disaster prevention and control district reaches 37.95 km2, and the area of the province area accounts for 78%. The research and development achievement of the project, namely the dynamic early warning technology of the mountain torrent disasters can effectively serve the national strategic requirements, so that the lives and properties threatened by the mountain torrent disasters in 21 grade cities in the whole province are effectively protected, and casualties are greatly reduced; the method can better meet the current key technical requirements of the future countries in the disaster prevention and reduction field and the public safety field, and improve the management efficiency; the technical bottleneck and the difficult problem of mountain torrent early warning in the data-lacking area can be better solved.
The method adopts two machine learning algorithms of BP neural network and SVR (support vector machine), combines hourly observation data of ground meteorological stations of Hanyuan and Dachuan provinces to optimize meteorological rainfall forecast data, uses the optimized rainfall data as input conditions of an early warning model, and accordingly achieves early warning of mountain torrent early warning.
The neural network algorithm in machine learning is introduced into the mountain torrent early warning to replace the traditional hydrological forecasting model, so that the limitation existing in the prior art can be effectively solved. One is to avoid the problem of insufficient theoretical research. The forecasting model constructed by the deep learning method is a data-driven black box model, mountain torrent early warning forecasting is carried out through the processes of input, calculation and output, each parameter has no actual physical significance in the core calculation process, and the output result only needs to be verified with the actual situation. And secondly, a proper forecasting model can be constructed in a mountain torrent area with missing data or low data quality by inputting sequence data internal relations in a simulation forecasting model by a Recurrent Neural Network (RNN), and the problems of gradient loss and gradient explosion in the RNN can be effectively solved by adding a long-short term memory (LTSM) method in an algorithm, so that small watershed (mountain torrent area) forecasting is more effectively carried out. And thirdly, the RNN can prolong the forecasting time to meet the requirement of mountain torrent early warning. The traditional hydrological forecast is usually based on the actual rainfall in a drainage basin, the forecast time is drainage basin convergence time, most of mountain torrent early warning areas are smaller drainage basins, the drainage basin convergence time is short, and although a hydrological meteorological department has forecast of rainstorm early warning on the areas, the small drainage basin data weather forecast cannot be carried out, so the traditional hydrological forecast cannot be used basically. The RNN can generalize the regional rainfall early warning forecast and then serve as an input factor, so that the forecast time is greatly prolonged. And fourthly, after the model is successfully developed, only input data needs to be concerned, and the method has the characteristics of simplicity and convenience in operation, high operation speed and convenience in value shifting in other areas. Along with the practical application of the model, data are accumulated continuously, manual secondary parameter adjustment is not needed, and the model can actively optimize and adjust parameters, so that the model is simpler and more convenient to apply.
Due to the adoption of the technical scheme, compared with the prior art, the invention has the advantages that:
the invention discloses a mountain torrent early warning method based on deep learning, which adopts an RNN model with an LTSM function, and adds a long-short term memory (LTSM) method into a model algorithm to effectively solve the problems of gradient dissipation and gradient explosion in the RNN, so that the forecasting of small watersheds or mountain torrent areas is more effectively carried out, the reliable mountain torrent disaster early warning forecasting is realized, the forecasting time is greatly improved, the model only needs to pay attention to input data, and the method has the characteristics of simple and convenient operation, high operation speed and convenience for shifting values to other areas.
The invention is further described with reference to the following figures and detailed description.
Drawings
FIG. 1 is a RNN model calculation flow chart.
FIG. 2 is a flow chart of RNN model calculation with LSTM algorithm in the present invention.
Detailed Description
Example (b): a mountain torrent early warning method based on deep learning comprises the following steps:
the first step is as follows: determining an early warning object: drawing a drainage basin where the early warning area is located on a GIS map, and determining a mountain torrent early warning area in the drainage basin and a preset early warning rainfall or water level, a distributed rainfall and a water level monitoring station thereof;
the second step is that: collecting the whole data: collecting historical data of rainfall, water level stations, regional evaporation capacity monitoring stations and county meteorological early warning in an early warning area, and calculating drainage basin convergence time;
the basin confluence time calculation method is inference formula calculation or Thiessen polygon calculation or the two are comprehensively determined.
The reasoning formula calculation method comprises the following steps: selecting a typical flood process in a historical period of a drainage basin water level control station, and calculating by using an inference formula, wherein the calculation formula is as follows:
in the formula, L is the distance from a watershed water distribution line to a control station, Q is the typical flood peak flow of the control station, J is the average watershed drop, and m is an empirical convergence coefficient.
The calculation method of the Thiessen polygon method comprises the following steps: converging the rainfall of the drainage basin surface and the drainage basin water level station on a time axis to form a surface rainfall and drainage basin control station flow timing diagram, and obtaining convergence time by adopting a secondary flood segmentation method.
The third step: monitoring station cross-correlation analysis: classifying the monitoring points according to rainfall and water level or flow, respectively calculating the cross-correlation coefficient of a certain station and the like according to a cross-correlation coefficient formula, and when the measured data of two stations cannot be obtained or the difference between the measured values of the two stations and the measured values of the related stations exceeds the maximum difference value N1 times of the historical series, calculating a result to replace the measured values through the measured data replacing formula;
the cross correlation coefficient formula calculation method comprises the following steps: let the site sequences be { x, respectivelytT is 1,2, …, n, and ytAnd t is 1,2, …, n, and the cross correlation coefficients are constructed as follows:
wherein, the closer the calculated correlation coefficient is to 1, the stronger the correlation between the two sites is represented. When the measured data of two stations cannot be obtained or the measured value of the two stations is too different from the measured value of the relevant station, if the maximum difference value N1 times of the historical series is taken, N1 is generally selected to be 1, and the measured data is considered to be possibly wrong.
The measured data substitution formula is as follows:
in the formula, A is the basin area controlled by the monitoring station, the rainfall station is the area cut by the Thiessen polygon, and the water level station is the basin area controlled.
Is the mean of the x samples and is,
is the y sample mean.
The fourth step: early warning of rainfall: an RNN rainfall forecasting model with an LTSM function is established, and mountain torrent rainfall early warning forecasting is carried out by adopting integrated rainfall forecasting calculation;
the method for performing the mountain torrent rainfall early warning forecast by integrating the rainfall forecast calculation comprises the following steps:
s4-1, calculating forecast accuracy evaluation index through forecast evaluation formula,
the forecast evaluation formula is as follows:
in the formula, E is the evaluation accuracy, m is the number of times that the predicted value and the measured value are in the same interval, and n is the total number of times of prediction.
S4-2, determining input layer and output layer data:
the input layer includes four factors, respectively: the real-time station observation value at the time t, the forecast value of the station at the time t + i, the forecast value of the provincial meteorological department grid forecast at the time t + i converted to the station by the professional meteorological company and the rainstorm early warning value issued by the local meteorological department at the time t;
wherein, the rainstorm early warning value that local meteorological department's t instant was released respectively is: no 0 is issued, blue warning is 1, yellow is 2, orange is 3, red is 4.
The output layer is the measured value of the station at the time t + i;
s4-3, establishing an RNN rainfall forecasting model with the LTSM function:
constructing an RNN model with an LTSM function, wherein the RNN model comprises an input layer and a hidden layer, an RNN calculation flow chart is shown in the attached figure 1, and a calculation formula is as follows:
ot+1＝V·st+1+bo
st+1＝f(U·xt+1+W·st+bs)
ot+1＝V·st+1+bo
in the above formula, U is the connection weight from the input layer to the hidden layer, V is the connection weight from the hidden layer to the output layer, W is the connection weight from the previous hidden layer to the current hidden layer, and xtFor inputting layer data, StHidden layer computation output data, otOutput layer calculates output data, bsTo hide the layer offset number, boFor the output layer offset number, f () is the excitation function,
wherein, the RNN model with LSTM algorithm in the invention is shown in figure 2, the hidden layer of RNN comprises a short-term state s sensitive to short-term input and a structure c for storing long-term state, the long-term state stored by the structure c is a unit state, the structure c comprises a forgetting gate, an input gate and an output gate for controlling the unit state,
wherein the content of the first and second substances,
the forgetting gate determines how much the unit state at the previous moment is reserved to the current moment ct;
the input gate determines the input of the network at the current moment and how many units are stored in the unit state;
how many output values of the states of the output door control unit are output to the hidden layer value of the current LSTM;
the calculation formula is as follows:
ft＝σ(Wt·[ht-1,xt]+bf)
it＝σ(Wi·[ht-1,xt]+bi)
ot＝σ(Wo·[ht-1,xt]+bo)
ht＝ot·tanh(ct)
in the above formula, ftIndicating forgetting gate, itRepresentation input gate, ctIndicating the state of the cell at time t, otDenotes the output gate, W denotes the weight, b denotes the offset, htExpressing the output at the time t, wherein sigma is an activation function, and tanh is a hyperbolic tangent activation function;
the hidden layer comprises two hidden layers which are respectively a first hidden layer and a second hidden layer, wherein the first hidden layer adopts the same number of units as the input layer, and the second hidden layer adopts half of the number of input factors as the number of hidden units for calculation;
s4-4, training and verifying: dividing collected and sorted historical measured rainfall and forecast data into a training set and a verification set, and training model parameters by using the training set so as to verify and calculate forecast accuracy by using the verification set;
wherein the data volume ratio of the training set to the validation set is 4: 1; the calculation forecasting precision requirement of the verification set reaches 80% or more.
S4-5, calculating real-time data: and substituting the real-time data into the calculation model to perform mountain torrent rainfall early warning and forecasting.
The fifth step: water level/flow early warning: and establishing an RNN water level forecasting model with an LTSM function, substituting real-time data into the model for calculation, and realizing mountain torrent water level early warning and forecasting.
The steps of carrying out mountain torrent water level early warning and forecasting are as follows:
s5-1, determining the early warning water level: determining an early warning water level according to a regional flood prevention plan or a hydrologic station water level amplitude variation correlation relation;
the method for determining the early warning water level through the hydrologic station water level amplitude variation correlation relationship comprises the following steps:
acquiring water level values of different flood frequencies of hydrological stations with long-sequence observation values in a flow field, and calculating water level amplitude difference delta H of simple water level stations established in a danger zone in the same field1,ΔH2Drawing a correlation diagram of water level amplitude variation of the two stations, thereby determining early warning that the water level of the simple water station corresponds to different torrential flood dangerous areasAnd (4) a threshold value.
S5-2, determining an input layer and an output layer: the input layer and the output layer can be determined by adopting an actual measurement data method or a forecast time method;
the method for determining the input layer and the output layer by the actual measurement data method comprises the following steps: adopting measured data, wherein an input layer is a rainfall station measured value, an evaporation capacity measured value and a control station measured value above a control station, and an output layer is a control station water level value after confluence time tau;
the method for determining the input layer and the output layer by the forecast time method comprises the following steps: setting the rainfall forecast time as n times of the convergence time tau, setting the rainfall forecast values, the forecast air temperature values and the forecast values of the control station on the input layer, and setting the water level value of the post-control station when the output layer is n tau jf.
S5-3, establishing an RNN water level forecasting model with LTSM function, wherein the RNN model comprises an input layer and a hidden layer, and the calculation formula is as follows:
ot+1＝V·st+1+bo
st+1＝f(U·xt+1+W·st+bs)
ot+1＝V·st+1+bo
in the above formula, U is the connection weight from the input layer to the hidden layer, V is the connection weight from the hidden layer to the output layer, W is the connection weight from the previous hidden layer to the current hidden layer, and xtFor input layer data, stHidden layer computation output data, otOutput layer calculates output data, bsTo hide the layer offset number, boFor the output layer offset number, f () is the excitation function,
wherein the hidden layer of the RNN comprises a short-term state s sensitive to short-term input and a structure c for storing a long-term state, the long-term state stored by the structure c is a unit state, the structure c comprises a forgetting gate, an input gate and an output gate for controlling the unit state,
wherein the content of the first and second substances,
the forgetting gate determines how much the unit state at the previous moment is reserved to the current moment ct;
the input gate determines the input of the network at the current moment and how many units are stored in the unit state;
how many output values of the states of the output door control unit are output to the hidden layer value of the current LSTM;
the calculation formula is as follows:
ft＝σ(Wt·[ht-1,xt]+bf)
it＝σ(Wi·[ht-1,xt]+bi)
ot＝σ(Wo·[ht-1,xt]+bo)
ht＝ot·tanh(ct)
in the above formula, ftIndicating forgetting gate, itRepresentation input gate, ctIndicating the state of the cell at time t, otDenotes the output gate, W denotes the weight, b denotes the offset, htExpressing the output at the time t, wherein sigma is an activation function, and tanh is a hyperbolic tangent activation function;
the hidden layers comprise a plurality of hidden layers, the number of units of a first hidden layer in the plurality of hidden layers is the same as that of input layers, and the number of the input factors of the previous hidden layer is half that of the input factors of the second hidden layer and the hidden layers behind the second hidden layer in the plurality of hidden layers is used as the number of the hidden units for calculation until the number of the input factors of the last hidden layer is less than or equal to 4;
the number of the hidden layers is determined by halving the number of input factors of the next hidden layer until the number of input factors of the last hidden layer is less than or equal to 4.
And S5-4, establishing a training set and a verification set, wherein the training precision takes the standard that the difference between the predicted value and the actually measured value in the verification set is qualified.
Because the water level series data are longer, the data collected and sorted in the invention are divided into two data sets of training and verification, and the data are divided into two data sets according to the series length of 5: 1 is divided.
The qualified standard of the training precision is as follows: the difference between the predicted value and the actual measured value is qualified when the difference is less than or equal to 20cm in a high water area, and is qualified when the difference is less than or equal to 50cm in a low water area.
And S5-5, substituting the real-time data into the model for calculation, and realizing the early warning and forecasting of the mountain torrent water level.
The prevention and control of the mountain torrent disasters are one of the difficulties in the field of disaster prevention and reduction in China, the difference of mountain torrent disaster areas in China is obvious, particularly the difference between the south and the north is obvious, and a reliable mountain torrent disaster early warning and forecasting technology is not available so far. Sichuan is one of the most provinces of mountain torrent disasters in China, more than 4000 counties (cities and districts) in total are threatened by the mountain torrent disasters, the total area of the mountain torrent disaster prevention and control district reaches 37.95 km2, and the area of the province area accounts for 78%. The research and development achievement of the project, namely the dynamic early warning technology of the mountain torrent disasters can effectively serve the national strategic requirements, so that the lives and properties threatened by the mountain torrent disasters in 21 grade cities in the whole province are effectively protected, and casualties are greatly reduced; the method can better meet the current key technical requirements of the future countries in the fields of disaster prevention and reduction and public safety, and improve the management efficiency; the technical bottleneck and the difficult problem of mountain torrent early warning in the data-lacking area can be better solved.
The method adopts two machine learning algorithms of BP neural network and SVR (support vector machine), combines hourly observation data of ground meteorological stations of Hanyuan and Dachuan provinces to optimize meteorological rainfall forecast data, uses the optimized rainfall data as input conditions of an early warning model, and accordingly achieves early warning of mountain torrent early warning.
Claims (5)
1. A mountain torrent early warning method based on deep learning comprises the following steps:
s1: determining an early warning object: drawing a drainage basin where the early warning area is located on a GIS map, and determining a mountain torrent early warning area in the drainage basin and a preset early warning rainfall or water level, a distributed rainfall and a water level monitoring station thereof;
s2: collecting the whole data: collecting historical data of rainfall, water level stations, regional evaporation capacity monitoring stations and county meteorological early warning in an early warning area, and calculating basin convergence time, wherein the basin convergence time is calculated by an inference formula or a Thiessen polygon method or comprehensively determined by the two methods;
the reasoning formula calculation method comprises the following steps: selecting a typical flood process in a historical period of a drainage basin water level control station, and calculating by using an inference formula, wherein the calculation formula is as follows:
in the formula, L is the distance from a watershed water distribution line to a control station, Q is the typical flood peak flow of the control station, J is the average watershed drop, and m is an empirical convergence coefficient;
the Thiessen polygon method comprises the following steps: converging the rainfall of the drainage basin surface and the drainage basin water level station on a time axis to form a surface rainfall and drainage basin control station flow timing diagram, and obtaining convergence time by adopting a secondary flood segmentation method;
s3: monitoring station cross-correlation analysis: classifying the monitoring points according to rainfall and water level or flow, respectively calculating the cross-correlation coefficient of a certain station and the like according to a cross-correlation coefficient formula, and when the measured data of two stations cannot be obtained or the difference between the measured values of the two stations and the measured values of the related stations exceeds the maximum difference value N1 times of the historical series, calculating a result to replace the measured values through the measured data replacing formula;
the cross correlation coefficient formula calculation method comprises the following steps: set the site sequences as
And { ytI, t ═ 1,2, …, n }, and their cross-correlation coefficients are constructed as:
the measured data substitution formula is as follows:
in the formula, A is the basin area controlled by a monitoring station, a rainfall station is the area cut by a Thiessen polygon, and a water level station is the basin area controlled;
is the mean of the x samples and is,
is the y sample mean;
s4: early warning of rainfall: an RNN rainfall forecasting model with an LTSM function is established, and mountain torrent rainfall early warning forecasting is carried out by adopting integrated rainfall forecasting calculation;
the method for performing mountain torrent rainfall early warning and forecasting by integrating rainfall forecasting calculation comprises the following steps:
s4-1, calculating forecast accuracy evaluation index through forecast evaluation formula,
the forecast evaluation formula is as follows:
in the formula, E is the evaluation precision, m is the number of times that the predicted value and the measured value are in the same interval, and n is the total number of times of prediction;
s4-2, determining input layer and output layer data:
the input layer includes four factors, respectively: the real-time station observation value at the time t, the forecast value of the station at the time t + i, the forecast value of the provincial meteorological department grid forecast at the time t + i converted to the station by the professional meteorological company and the rainstorm early warning value issued by the local meteorological department at the time t;
the output layer is the measured value of the station at the time t + i;
the values of the rainstorm early warning values issued by the local meteorological department at the moment t are respectively as follows: 0 is not released, blue early warning is 1, yellow is 2, orange is 3, and red is 4;
s4-3, establishing an RNN rainfall forecasting model with the LTSM function;
the method for establishing the RNN rainfall forecasting model with the LTSM function comprises the following steps:
constructing an RNN model with an LTSM function, wherein the RNN model comprises an input layer and a hidden layer, and the calculation formula is as follows:
ot+1＝V·st+1+bo，
st+1＝f(U·xt+1+W·st+bs)，
ot+1＝V·st+1+bo，
in the above formula, U is the connection weight from the input layer to the hidden layer, V is the connection weight from the hidden layer to the output layer, W is the connection weight from the previous hidden layer to the current hidden layer, and xtFor inputting layer data, StHidden layer computation output data, otOutput layer calculates output data, bsTo hide the layer offset number, boF () is the excitation function for the output layer offset number;
wherein the hidden layer of the RNN comprises a short-term state s sensitive to short-term input and a structure c for storing a long-term state, the long-term state stored by the structure c is a unit state, the structure c comprises a forgetting gate, an input gate and an output gate for controlling the unit state,
wherein the content of the first and second substances,
the forgetting gate determines how much the unit state at the previous moment is reserved to the current moment ct;
the input gate determines the input of the network at the current moment and how many units are stored in the unit state;
how many output values of the states of the output door control unit are output to the hidden layer value of the current LSTM;
the calculation formula is as follows:
ft＝σ(Wt·[ht-1,xt]+bf)，
it＝σ(Wi·[ht-1,xt]+bi)，
ot＝σ(Wo·[ht-1,xt]+bo)，
ht＝ot·tanh(ct)，
in the above formula, ftIndicating forgetting gate, itRepresentation input gate, ctIndicating the state of the cell at time t, otDenotes the output gate, W denotes the weight, b denotes the offset, htExpressing the output at the time t, wherein sigma is an activation function, and tanh is a hyperbolic tangent activation function;
the hidden layer comprises two hidden layers which are respectively a first hidden layer and a second hidden layer, wherein the first hidden layer adopts the same number of units as the input layer, and the second hidden layer adopts half of the number of input factors as the number of hidden units for calculation;
s4-4, training and verifying: dividing collected and sorted historical measured rainfall and forecast data into a training set and a verification set, and training model parameters by using the training set so as to verify and calculate forecast accuracy by using the verification set;
s4-5, calculating real-time data: substituting real-time data into the calculation model to perform mountain torrent rainfall early warning and forecasting;
s5: water level/flow early warning: and establishing an RNN water level forecasting model with an LTSM function, substituting real-time data into the model for calculation, and realizing mountain torrent water level early warning and forecasting.
2. The mountain torrent early warning method based on deep learning of claim 1, wherein:
the steps of mountain torrent water level early warning and forecasting are as follows:
s5-1, determining the early warning water level: passing through a regional flood prevention plan or a hydrological station water level variation correlation relation;
s5-2, determining an input layer and an output layer: determining an input layer and an output layer by adopting an actual measurement data method or a forecast time method;
s5-3, establishing an RNN water level forecasting model with the LTSM function;
s5-4, establishing a training set and a verification set, wherein the training precision takes the standard that the difference between the predicted value and the actually measured value in the verification set is qualified;
and S5-5, substituting the real-time data into the model for calculation, and realizing the early warning and forecasting of the mountain torrent water level.
3. The mountain torrent early warning method based on deep learning of claim 2, wherein:
in S5-1, the method for establishing the correlation between the simple water level station and the water level amplitude of the actually measured hydrological station comprises the following steps:
acquiring water level values of different flood frequencies of hydrological stations with long-sequence observed values in a flow area, and calculating water level amplitude difference delta H of simple water level stations set in a dangerous area in the same field1,ΔH2And drawing a water level amplitude variation correlation diagram of the two stations, thereby determining the early warning threshold values of the simple water level station corresponding to different torrential flood dangerous areas.
4. The mountain torrent early warning method based on deep learning of claim 3, wherein:
in S5-2, the method for determining the input layer and the output layer by the measured data method comprises the following steps: adopting measured data, wherein an input layer is a rainfall station measured value, an evaporation capacity measured value and a control station measured value above a control station, and an output layer is a control station water level value after confluence time tau;
the method for determining the input layer and the output layer by the forecast time method comprises the following steps: setting the rainfall forecast time as n times of the convergence time tau, setting the rainfall forecast values, the forecast air temperature values and the forecast values of the control station on the input layer, and setting the water level value of the post-control station when the output layer is n tau jf.
5. The mountain torrent early warning method based on deep learning of claim 2, wherein:
the method for establishing the RNN water level forecasting model with the LTSM function in the S5-3 comprises the following steps:
constructing an RNN model with an LTSM function, wherein the RNN model comprises an input layer and a hidden layer, and the calculation formula is as follows:
ot+1＝V·st+1+bo，
st+1＝f(U·xt+1+W·st+bs)，
ot+1＝V·st+1+bo，
in the above formula, U is the connection weight from the input layer to the hidden layer, V is the connection weight from the hidden layer to the output layer, W is the connection weight from the previous hidden layer to the current hidden layer, and xtFor input layer data, stHidden layer computation output data, otOutput layer calculates output data, bsTo hide the layer offset number, boF () is the excitation function for the output layer offset number;
wherein the hidden layer of the RNN comprises a short-term state s sensitive to short-term input and a structure c for storing a long-term state, the long-term state stored by the structure c is a unit state, the structure c comprises a forgetting gate, an input gate and an output gate for controlling the unit state,
wherein the content of the first and second substances,
the forgetting gate determines how much the unit state at the previous moment is reserved to the current moment ct;
the input gate determines the input of the network at the current moment and how many units are stored in the unit state;
how many output values of the states of the output door control unit are output to the hidden layer value of the current LSTM;
the calculation formula is as follows:
ft＝σ(Wt·[ht-1,xt]+bf)，
it＝σ(Wi·[ht-1,xt]+bi)，
ot＝σ(Wo·[ht-1,xt]+bo)，
ht＝ot·tanh(ct)，
in the above formula, ftIndicating forgetting gate, itRepresentation input gate, ctIndicating the state of the cell at time t, otDenotes the output gate, W denotes the weight, b denotes the offset, htExpressing the output at the time t, wherein sigma is an activation function, and tanh is a hyperbolic tangent activation function;
the hidden layers comprise a plurality of hidden layers, the number of units of a first hidden layer in the hidden layers is the same as that of input layers, and the number of the input factors of the previous hidden layer is half that of the input factors of the second hidden layer and the hidden layers behind the second hidden layer in the hidden layers is used as the number of the hidden units for calculation until the number of the input factors of the last hidden layer is less than or equal to 4;
the number of the hidden layers is determined by halving the number of input factors of the next hidden layer until the number of input factors of the last hidden layer is less than or equal to 4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910846384.1A CN110459036B (en) | 2019-09-09 | 2019-09-09 | Mountain torrent early warning method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910846384.1A CN110459036B (en) | 2019-09-09 | 2019-09-09 | Mountain torrent early warning method based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110459036A CN110459036A (en) | 2019-11-15 |
| CN110459036B true CN110459036B (en) | 2022-05-17 |
Family
ID=68491190
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910846384.1A Active CN110459036B (en) | 2019-09-09 | 2019-09-09 | Mountain torrent early warning method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110459036B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111639748B (en) * | 2020-05-15 | 2022-10-11 | 武汉大学 | Watershed pollutant flux prediction method based on LSTM-BP space-time combination model |
| CN111753965A (en) * | 2020-06-30 | 2020-10-09 | 长江水利委员会水文局 | Deep learning-based river flow automatic editing method and system |
| CN112418542A (en) * | 2020-12-03 | 2021-02-26 | 浙江知水信息技术有限公司 | Method for realizing early warning of flood conditions by machine deep learning based on meteorological data |
| CN112799156B (en) * | 2021-03-30 | 2021-07-20 | 中数通信息有限公司 | Meteorological integrated emergency early warning issuing method and device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105512767A (en) * | 2015-12-15 | 2016-04-20 | 武汉大学 | Flood forecasting method of multiple forecast periods |
| KR101695802B1 (en) * | 2015-07-16 | 2017-01-13 | 한국외국어대학교 연구산학협력단 | Flash flood monitoring method using a real time soil moisture data |
| CN106884405A (en) * | 2017-03-08 | 2017-06-23 | 中国水利水电科学研究院 | Inrush type mountain flood assay method for a kind of Cross Some Region Without Data |
| CN107992961A (en) * | 2017-11-21 | 2018-05-04 | 中国水利水电科学研究院 | A kind of adaptive basin Medium-and Long-Term Runoff Forecasting model framework method |
| CN109615011A (en) * | 2018-12-14 | 2019-04-12 | 河海大学 | A kind of middle and small river short time flood forecast method based on LSTM |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| AU2012101347A4 (en) * | 2011-09-06 | 2012-10-04 | Microair Technologies Pty Ltd | Flood warning system and method |
| CN105608840A (en) * | 2016-03-09 | 2016-05-25 | 长江水利委员会水文局 | Mountain torrents early warning platform based on fused quantitative rainfall forecast algorithm, and early warning method thereof |
| ES2735016T3 (en) * | 2016-06-29 | 2019-12-13 | Ontech Security Sl | Device and method of flood detection |
| CN207233166U (en) * | 2017-09-26 | 2018-04-13 | 成都远向电子有限公司 | Current early warning system based on Lora |
| CN109979172A (en) * | 2019-04-09 | 2019-07-05 | 南京信息工程大学 | A kind of dynamic mountain torrents Critical Rainfall forecasting procedure based on Xinanjiang model |
-
2019
- 2019-09-09 CN CN201910846384.1A patent/CN110459036B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101695802B1 (en) * | 2015-07-16 | 2017-01-13 | 한국외국어대학교 연구산학협력단 | Flash flood monitoring method using a real time soil moisture data |
| CN105512767A (en) * | 2015-12-15 | 2016-04-20 | 武汉大学 | Flood forecasting method of multiple forecast periods |
| CN106884405A (en) * | 2017-03-08 | 2017-06-23 | 中国水利水电科学研究院 | Inrush type mountain flood assay method for a kind of Cross Some Region Without Data |
| CN107992961A (en) * | 2017-11-21 | 2018-05-04 | 中国水利水电科学研究院 | A kind of adaptive basin Medium-and Long-Term Runoff Forecasting model framework method |
| CN109615011A (en) * | 2018-12-14 | 2019-04-12 | 河海大学 | A kind of middle and small river short time flood forecast method based on LSTM |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110459036A (en) | 2019-11-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110459036B (en) | Mountain torrent early warning method based on deep learning | |
| CN109978235B (en) | Waterlogging water level prediction method based on sample learning | |
| CN102289570B (en) | Flood forecast method based on rainfall-runoff-flood routing calculation | |
| CN111651885A (en) | Intelligent sponge urban flood forecasting method | |
| Azamathulla et al. | An ANFIS-based approach for predicting the bed load for moderately sized rivers | |
| CN104392111B (en) | Flood Forecasting Method based on water level sample | |
| CN113610264B (en) | Refined power grid typhoon flood disaster prediction system | |
| CN104298841A (en) | Flood forecasting method and system based on historical data | |
| Che et al. | Application of an optimization/simulation model for real-time flood-control operation of river-reservoirs systems | |
| Joshi et al. | Rainfall-Runoff Simulation in Cache River Basin, Illinois, Using HEC-HMS | |
| Mantilla et al. | The hydrological hillslope-link model for space-time prediction of streamflow: Insights and applications at the Iowa Flood Center | |
| Qiu et al. | Seepage monitoring models study of earth-rock dams influenced by rainstorms | |
| Wu et al. | Local and global Bayesian network based model for flood prediction | |
| CN114936505A (en) | Method for rapidly forecasting multi-point water depth of urban rainwater well | |
| Lai et al. | Waterlogging risk assessment based on self-organizing map (SOM) artificial neural networks: a case study of an urban storm in Beijing | |
| CN112528563B (en) | Urban waterlogging early warning method based on SVM algorithm | |
| Yuan et al. | Comprehensive assessment and rechecking of rainfall threshold for flash floods based on the disaster information | |
| Zarafshan et al. | Artificial Intelligence Hybrid Deep Learning Model for Groundwater Level Prediction Using MLP-ADAM | |
| Xu et al. | Exploration of Flood Prediction in Watersheds Based on the Fusion Analysis of Remote Sensing Big Data with Multiple Physical Fields | |
| Wang et al. | Study on city rainstorm waterlogging warning system based on historical data | |
| Fava et al. | Integration of information technology systems for flood forecasting with hybrid data sources | |
| Gao et al. | Regional flood risk analysis for Huaihong South Flood Control Protected Area in China using an integrated method. | |
| Wang et al. | Determination of dynamic critical rainfall based on geomorphological instantaneous unit hydrograph and radial basis function neural network. | |
| CN117313510A (en) | River channel runoff prediction method and system based on hydrologic model and neural network | |
| Li et al. | Research on Reservoir Safety Risk Based on BP Neural Network |
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 |