River tidal bore forecasting method based on convolutional neural network

Abstract

The present invention discloses a river tidal bore prediction method based on a convolutional neural network. The method pre-processes river tidal bore data, uses a correlation analysis method to analyze the influence of various influencing factors on the tidal bore, selects the influencing factors of the lunar calendar date, meteorological conditions, upstream water, river terrain, and the previous day information of the prediction factor as feature vectors to construct a feature set, uses a two-dimensional convolutional neural network to extract effective features in the input feature set, uses a fully connected neural network to establish a river tidal bore prediction model, uses a particle swarm algorithm to optimize the model hyperparameters, obtains an optimal model, and predicts the river tidal bore feature information based on the extracted influencing factor features. The present invention comprehensively considers the influence of various major influencing factors on river tidal bores, has high prediction accuracy and reliability, and can provide timely, accurate, and reliable tidal flood services for tide prevention safety management, citizens and tourists watching the tide, etc.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a river tidal bore forecasting method based on a convolutional neural network. Background Art

[0002] Citizens and tourists watching tides, tide prevention and safety management, water resources development and utilization, transportation and shipping, and water-related engineering construction all require accurate river tidal bore forecasts as a reference. However, river tidal bores are affected by a variety of factors, including astronomical and meteorological conditions (wind force, wind direction, air pressure, etc.), upstream water, and river topography, making forecasts difficult. Traditional empirical forecasting methods based on similar tide analysis cannot accurately analyze the impact of short-term random factors such as wind force, wind direction, and air pressure, and the forecast accuracy cannot be guaranteed.

[0003] The prior art (CN201610981643.8) discloses a Qiantang River tide forecasting method based on a support vector machine, but this method only considers influencing factors such as astronomical tides and upstream water, and does not consider influencing factors such as meteorological conditions and river topography. The main influencing factors are not considered comprehensively, and there is a certain uncertainty in the forecast accuracy. Summary of the invention

[0004] The present invention aims at the problem that the existing empirical forecasting method and support vector machine forecasting method do not fully consider the impact of various influencing factors on river tidal bores, and there is uncertainty in the forecasting accuracy, and provides a river tidal bore forecasting method based on a convolutional neural network.

[0005] In order to solve the above technical problems, the present invention is solved by the following technical solutions:

[0006] A river tidal bore forecasting method based on a convolutional neural network includes:

[0007] Preprocessing of river tidal bore data: determining the influencing factors of river tidal bore through correlation analysis and constructing a feature set of river tidal bore data; the influencing factors of river tidal bore include the lunar calendar date, meteorological conditions, upstream water, river topography, and the previous day's information of tidal bore forecasting factors;

[0008] Establishment of river tidal bore forecasting model, training the river tidal bore forecasting model based on the river tidal bore data feature set; the forecasting model includes CNN network and fully connected network;

[0009] Optimization of hyperparameters of river tidal bore forecasting model. The model hyperparameters are optimized by particle swarm optimization algorithm to determine the prediction model of river tidal bore forecasting.

[0010] Preferably, the preprocessing of river tidal bore data includes:

[0011] Acquisition of historical data samples of river tidal bores;

[0012] Data preprocessing, normalization of feature set data, normalization of feature set according to formula 1;

[0013]

[0014] In formula 1, x′ is the normalized tidal bore data; x is the data before preprocessing; x max 、x min are the maximum and minimum values ​​of each feature in the training set respectively;

[0015] Preprocessing of meteorological data, where the meteorological data includes wind force and direction, and the wind characteristic data is processed by formula 2;

[0016] w e =-w*sinθ Formula 2

[0017] In formula 2, w e is the wind force in the east direction after conversion, w is the measured wind force, and θ is the measured wind direction;

[0018] Correlation analysis was conducted on the influencing factors of river tidal bores and the lunar calendar date, meteorological conditions, upstream water, river topography, and the previous day's information on tidal bore forecasting factors;

[0019] Feature set construction,selects influencing factors such as lunar calendar date, meteorological conditions, upstream water, river topography, and the previous day's information of tidal bore prediction factors as feature vectors to construct feature sets as inputs of the forecast model.

[0020] As a preferred method, the establishment of a river tidal bore forecasting model includes:

[0021] Construct the input matrix. A two-dimensional convolutional neural network usually takes a three-dimensional matrix of H×W as an input parameter, where the height of H is the number of rows of the input matrix and the width of W is the number of columns of the input matrix.

[0022] Select the convolution kernel. For the CNN network, use 10 2×2 convolution kernels and pad the edges of the input matrix with zeros.

[0023] Select the activation function and use the ReLU activation function;

[0024] Model pooling, pooling the output of the global mean;

[0025] A prediction model is established based on the fully connected layer. Each node of the input layer of the fully connected layer is connected to all the nodes of the CNN output layer, and the input features are integrated to predict the characteristics of river tidal bores. The output of the convolutional layer is a three-dimensional matrix of 10×7×7, which can be flattened to obtain a one-dimensional vector of 1×490 as the input of the fully connected layer. The fully connected layer consists of two hidden layers. The first layer consists of 100 neurons, and the second layer, i.e., the output layer, consists of one neuron (one output), using the LINEAR activation function.

[0026] Preferably, the method for optimizing the hyperparameters of the river tidal bore forecast model includes:

[0027] The training of the prediction model uses the root mean square error (RMSE) as the cost function for model training;

[0028]

[0029] In formula 3, is the predicted output value of the model target quantity, and y is the true value of the target quantity;

[0030] Model parameter optimization algorithm, the convolutional neural network parameters are optimized by particle swarm optimization algorithm; the training sample set T is divided into two non-intersecting subsets Ttrain and Tvalid, Ttrain is used for model training, and Tvalid is used for model verification; the parameter optimization problem with model learning rate lr, number of training times Tm, momentum terms beta1 and beta2 as variables is recorded as: min lr,Tm,beta1,beta2 F = RMSE(T Valid ).

[0031] Preferably, the particle swarm optimization algorithm method includes:

[0032] S1, initialization parameters, initialization population size Q, maximum number of iterations J, Q particle positions x i and speed v i ;

[0033] v i =v i +c 1 ×rand()×(pbest i -x i )+c 2 ×rand()×(gbest i -x i ) Formula 4

[0034] x i =x i +v i Formula 5

[0035] Where i = 1, 2, ..., N, N is the total number of particle swarms; v i is the speed of the particle; rand() is a random number between (0,1); x i is the current position of the particle; c1, c2 are learning factors, v i The maximum value is V max (greater than 0), and greater than 0;

[0036] S2, evaluates the position of each particle based on the river tidal bore forecasting model;

[0037] S3, compare the current position of each particle with the optimal position P best If the current position is good, record the new position and update P best ;

[0038] S4, calculate and update the speed and position of each particle according to Formula 4 and Formula 5;

[0039] S5, determine the number of iterations. If the number of iterations is less than J, return to S2, otherwise end.

[0040] In order to solve the above technical problem, the present invention further provides an electronic device, characterized in that it includes: at least one processor and a memory;

[0041] The memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the river tidal forecasting method based on convolutional neural network as described.

[0042] In order to solve the above technical problems, the present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer execution instructions, and when the processor executes the computer execution instructions, the river tidal forecasting method based on convolutional neural network as described is implemented.

[0043] The present invention has significant technical effects due to the adoption of the above technical solution:

[0044] The present invention adopts the correlation analysis method to analyze the influence of various influencing factors on river tidal bore, and comprehensively considers the river tidal bore forecast under the influence of multiple factors such as astronomy, river topography, upstream water, and meteorological conditions, so the overall forecast accuracy is relatively high.

[0045] The present invention selects influencing factors such as lunar calendar date, meteorological conditions, upstream water, river topography, and the previous day information of prediction factors as feature vectors to construct different feature sets according to different tidal bore prediction factors such as the time difference of river tidal bore arrival every other day, propagation time, high tide level, low tide level, etc., uses convolutional neural network and fully connected neural network to establish river tidal bore prediction model, and uses particle swarm algorithm to optimize model hyperparameters, thereby further improving the accuracy of river tidal bore prediction;

[0046] The present invention overcomes the problem that conventional empirical forecasting methods cannot solve the impact of short-term random changes in influencing factors, solves the problem that traditional machine learning algorithms do not fully consider the impact of various influencing factors on river tidal bores, and there is uncertainty in forecast accuracy. It comprehensively considers the impact of various major influencing factors on river tidal bores. Compared with conventional forecasting methods and traditional machine learning algorithms, it has higher forecasting accuracy and reliability, and can provide timely, accurate and reliable tidal services for citizens and tourists to watch tides, tide prevention and safety management, water resources development and utilization, shipping and transportation, and construction of water-related projects along the river. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of the present invention.

[0048] Figure 2-1 This is a graph showing the time difference between two days and the lunar calendar date. Figure 2-2 This is a graph showing the variation of propagation time with lunar calendar date. Figure 2-3 This is a graph showing the high tide level changing with the lunar calendar date. Figure 2-4 It is a graph showing the low tide level changing with the lunar calendar date;

[0049] Figure 3-1 This is the distribution diagram of the time difference between two days and the wind force. Figure 3-2 is the distribution diagram of propagation time and wind speed. Figure 3-3 This is a distribution diagram of high tide level and wind force. Figure 3-4 It is a distribution map related to low tide level and wind force;

[0050] Figure 4-1 This is the relationship between the time difference between the next two days and the outflow of the upstream reservoir. Figure 4-2 is the relationship between propagation time and outflow from upstream reservoirs, Figure 4-3 This is the relationship between high tide level and outflow from upstream reservoirs. Figure 4-4 It is a graph showing the relationship between low tide level and outflow from upstream reservoirs;

[0051] Figure 5-1 It is the correlation diagram between the time difference between the current day and the time difference between the previous day and the previous day. Figure 5-2 It is the correlation diagram between the propagation time of the day and the propagation time of the previous day. Figure 5-3 This is the correlation diagram between the high tide level of the day and the high tide level of the previous day. Figure 5-4It is a correlation diagram between the low tide level of the day and the low tide level of the previous day;

[0052] Figure 6-1 It is a comparison chart of the measured data and the predicted data of the time difference of the next day. Figure 6-2 This is a comparison chart between the measured data and the predicted data of the propagation time. Figure 6-3 Comparison chart of high tide level measured data and predicted data. Figure 6-4 It is a comparison chart between the measured data and the predicted data of low tide level;

[0053] Figure 7 It is a forecast model structure diagram of the present invention;

[0054] Figure 8 It is the flow chart of the particle swarm algorithm of the present invention. DETAILED DESCRIPTION

[0055] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0056] Example 1

[0057] A river tidal bore forecasting method based on a convolutional neural network includes:

[0058] Preprocessing of river tidal bore data: determining the influencing factors of river tidal bore through correlation analysis and constructing a feature set of river tidal bore data; the influencing factors of river tidal bore include the lunar calendar date, meteorological conditions, upstream water, river topography, and the previous day's information of tidal bore forecasting factors;

[0059] Establishment of river tidal bore forecasting model, training the river tidal bore forecasting model based on the river tidal bore data feature set; the forecasting model includes CNN network and fully connected network;

[0060] Optimization of hyperparameters of river tidal bore forecasting model. The model hyperparameters are optimized by particle swarm optimization algorithm to determine the prediction model of river tidal bore forecasting.

[0061] The preprocessing of river tidal bore data includes:

[0062] Acquisition of historical data samples of river tidal bores;

[0063] Data preprocessing, normalization of feature set data, normalization of feature set according to formula 1;

[0064]

[0065] In formula 1, x′ is the normalized tidal bore data; x is the data before preprocessing; x max 、x min are the maximum and minimum values ​​of each feature in the training set respectively;

[0066] Preprocessing of meteorological data, where the meteorological data includes wind force and direction, and the wind characteristic data is processed by formula 2;

[0067] w e =-w*sinθ Formula 2

[0068] In formula 2, w e is the wind force in the east direction after conversion, w is the measured wind force, and θ is the measured wind direction;

[0069] Correlation analysis was conducted on river tidal bores and other influencing factors such as the lunar calendar date, meteorological conditions, upstream water, river topography, and the previous day's information on tidal bore forecasting factors.

[0070] Feature set construction,selects influencing factors such as lunar calendar date, meteorological conditions, upstream water, river topography, and the previous day's information of tidal bore prediction factors as feature vectors to construct feature sets as inputs of the forecast model.

[0071] The establishment of river tidal bore forecasting model includes:

[0072] Construct the input matrix. A two-dimensional convolutional neural network usually takes a three-dimensional matrix of H×W as an input parameter, where the height of H is the number of rows of the input matrix and the width of W is the number of columns of the input matrix.

[0073] Select the convolution kernel. For the CNN network, use 10 2×2 convolution kernels and pad the edges of the input matrix with zeros.

[0074] Select the activation function and use the ReLU activation function;

[0075] Model pooling, pooling the output of the global mean;

[0076] A prediction model is established based on the fully connected layer. Each node of the input layer of the fully connected layer is connected to all the nodes of the CNN output layer, and the input features are integrated to predict the characteristics of river tidal bores. The output of the convolutional layer is a three-dimensional matrix of 10×7×7, which can be flattened to obtain a one-dimensional vector of 1×490 as the input of the fully connected layer. The fully connected layer consists of two hidden layers. The first layer consists of 100 neurons, and the second layer, i.e., the output layer, consists of one neuron (one output), using the LINEAR activation function.

[0077] The optimization methods of hyperparameters of river tidal bore forecasting model include:

[0078] The training of the prediction model uses the root mean square error (RMSE) as the cost function for model training;

[0079]

[0080] In formula 3, is the predicted output value of the model target quantity, and y is the true value of the target quantity;

[0081] Model parameter optimization algorithm, the convolutional neural network parameters are optimized by particle swarm optimization algorithm; the training sample set T is divided into two non-intersecting subsets Ttrain and Tvalid, Ttrain is used for model training, and Tvalid is used for model verification; the parameter optimization problem with model learning rate lr, number of training times Tm, momentum terms beta1 and beta2 as variables is recorded as: min lr,Tm,beta1,beta2 F = RMSE(T Valid ).

[0082] The particle swarm optimization algorithm methods include:

[0083] S1, initialization parameters, initialization population size Q, maximum number of iterations J, Q particle positions x i and speed v i ;

[0084] v i =v i +c 1 ×rand()×(pbest i -x i )+c 2 ×rand()×(gbest i -x i ) Formula 4

[0085] x i =x i +v i Formula 5

[0086] Where i = 1, 2, ..., N, N is the total number of particle swarms; v i is the speed of the particle; rand() is a random number between (0,1); x i is the current position of the particle; c1, c2 are learning factors, v i The maximum value is V max (greater than 0), and greater than 0;

[0087] S2, evaluates the position of each particle based on the river tidal bore forecasting model;

[0088] S3, compare the current position of each particle with the optimal position P best If the current position is good, record the new position and update P best ;

[0089] S4, calculate and update the speed and position of each particle according to Formula 4 and Formula 5;

[0090] S5, determine the number of iterations. If the number of iterations is less than J, return to S2, otherwise end.

[0091] Example 2

[0092] Based on Example 1, this example is a Qiantang River tidal bore forecasting method based on a convolutional neural network. The Qiantang River tidal bore is mainly affected by factors such as the lunar calendar date, wind force and direction, upstream water, river topography, and yesterday's information of forecast factors.

[0093] The time difference of the tidal bore arrival at Cangqian Station, the representative station of Qiantang River, the propagation time (Cangqian-Qibao), the high tide level and the low tide level at each other's day with the lunar calendar date are shown in the figure. Figures 2-1 to 2-4 As shown in the figure, the time difference of the tidal bore arrival at Cangqian Station every other day, the propagation time (Cangqian-Qibao), the high tide level and the low tide level all show a periodic change relationship with the change of the lunar calendar date, that is, the Qiantang River tidal bore prediction factor has a strong correlation with the lunar calendar date.

[0094] The distribution diagram of the time difference of the tidal bore arrival every other day at Cangqian Station, the representative station of Qiantang River, the propagation time (Cangqian-Qibao), the high tide level, the low tide level and the instantaneous wind force at Tanhudao Station, the representative meteorological station of Qiantang River, is shown in the figure below. Figures 3-1 to 3-4 As shown in the figure, although the Qiantang River tidal bore is greatly affected by wind force and direction, the time difference between the arrival of the tidal bore at Cangqian Station, the propagation time (Cangqian-Qibao), the high tide level, and the low tide level are not strongly correlated with the instantaneous wind force at Tanhudao Station, that is, the Qiantang River tidal bore prediction factor is weakly correlated with the instantaneous wind force.

[0095] The relationship between the time difference of the tidal bore arrival at Cangqian Station, the propagation time (Cangqian-Qibao), the high tide level, the low tide level and the sum of the daily average outflow from the Fenshuijiang Reservoir and the Fuchunjiang Reservoir upstream of the Qiantang River is shown in the figure below. Figures 4-1 to 4-4 As shown. Although when the upper reaches of the Qiantang River are flooded and the water volume is large, the flood will severely scour the river channel, deepen the river channel, reduce the low tide level, strengthen the upstream ability of the tide, and enhance the tidal bore. However, as can be seen from the figure, the time difference of the tidal bore arrival every other day at Cangqian Station, the propagation time (Cangqian-Qibao), the high tide level, and the low tide level are not strongly correlated with the sum of the daily average flow of the Fenshuijiang Reservoir and the Fuchunjiang Reservoir in the upper reaches of the Qiantang River, that is, the Qiantang River tidal bore prediction factor is weakly correlated with the daily average flow of the upstream reservoir;

[0096] The correlation diagram of the time difference between the arrival time of the tidal bore on the next day, the propagation time (Cangqian-Qibao), the high tide level, the low tide level and the data of the previous day at the representative station of Qiantang River Cangqian Station is shown in the figure below. Figures 5-1 to 5-4 As shown in the figure, the time difference between the arrival of the tidal bore on the next day, the propagation time (Cangqian-Qibao), the high tide level and the low tide level at Cangqian Station on the same day all have a strong linear relationship with the data of the previous day.

[0097] In summary, the time difference of tidal bore arrival every other day, propagation time (Cangqian-Qibao), high tide level and low tide level at Cangqian Station, the representative station of Qiantang River, all show a periodic change relationship with the lunar calendar date, and have a strong correlation with the lunar calendar date; have a strong linear relationship with the data of the previous day; have a weak correlation with the instantaneous wind force of the representative meteorological station at the trumpet mouth of Qiantang River; and have a weak correlation with the sum of the daily average flow of Fenshuijiang Reservoir and Fuchunjiang Reservoir in the upper reaches of Qiantang River.

[0098] A feature matrix consisting of 7*7 influencing factors is used as input.

[0099] There are 7 influencing factors for the Qiantang River tidal bore forecast, and the forecast feature set is shown in Table 1. The previous day’s next day’s time difference, lunar calendar date, previous day’s low tide, Fuchun River reservoir outflow, Fenshui River reservoir outflow, and easterly wind force at two meteorological stations (22 time points at each station, a total of 44 time points), a total of 49 input factors, constitute a 7×7 two-dimensional matrix as input.

[0100] Table 1. Feature set for prediction of time difference between two days

[0101]

[0102] For the Qiantang River tidal bore propagation time forecast, there are 7 influencing factors, and the prediction feature set is shown in Table 2.

[0103] Table 2. Propagation time prediction feature set

[0104]

[0105] For the Qiantang River tidal high water level forecast, there are 7 influencing factors, and the prediction feature set is shown in Table 3.

[0106] Table 3 High tide prediction feature set

[0107]

[0108] For the Qiantang River tidal low tide forecast, there are 6 influencing factors, and the prediction feature set is shown in Table 4.

[0109] Table 4 Low tide prediction feature set

[0110]

[0111]

[0112] The forecast model structure is as follows Figure 7The model consists of two parts: a convolutional neural network (CNN network) and a fully connected network. First, a layer of CNN network is used to extract effective feature information from the two-dimensional input feature matrix. Secondly, based on the extracted feature components, a Qiantang River tidal bore forecasting model is established through a fully connected network to forecast tidal bores.

[0113] The particle swarm algorithm is used to optimize the model parameter training times, learning rate and momentum term. Usually the learning rate lr, training times Tm, momentum terms beta1 and beta2 are set in the range of [0.0001, 0.01], [3000, 20000], [0.9, 1], [0.99, 1], so the particle swarm algorithm is used to optimize them within the above range.

[0114] According to the lunar calendar, the Qiantang River tidal bore is divided into large, medium and small tides. The large tide period is from the second to the fifth day and the seventeenth to the twentieth day of each lunar month. The medium tide period is two days before and after the large tide period, and the rest are small tides. The historical hydrological and meteorological data during the nine-year high tide period from 2009 to 2017 were used to calibrate and verify the tidal bore forecast model of Cangqian Station.

[0115] All Qiantang River tidal bore data were preprocessed and the input matrix was constructed. At the same time, all the data were divided into two independent groups, one for training samples and the other for verification samples.

[0116] The forecast of the time difference between the arrival time of the tidal bore and the arrival time of the tidal bore:

[0117] The model was calibrated and verified using the daily time difference data of 781 high tides at Cangqian Station, a representative tidal bore station of the Qiantang River, from 2009 to 2017. The first 500 tidal bore data were used for model calibration, and the last 281 tidal bore data were used for model verification.

[0118] The parameter optimization results show that when the learning rate lr is 1.11610457e-3, the number of training times Tm is 4.02440640e+03, the momentum term beta1 is 9.74464217e-1, and beta2 is 9.98240117e-1, the minimum objective function is 5.328.

[0119] Figure 6-1The prediction effect of the time difference of the spring tide at the Cangqian station was compared. The model prediction results show that in the rate period, the root mean square error of the results of the proposed Qiantang River tidal bore time difference prediction method is 9.25min, the average absolute error is 7.05min, the data with errors within 10min account for 76.0%, the data with errors within 20min account for 95.8%, the data with errors within 30min account for 99.0%, and the data with errors within 40min account for 100.0%. In the verification period, the root mean square error of the results of the proposed Qiantang River tidal bore time difference prediction method is 7.43min, the average absolute error is 5.34min, the data with errors within 10min account for 85.8%, the data with errors within 20min account for 97.9%, the data with errors within 30min account for 99.7%, the data with errors within 40min account for 99.7%, and the data with errors within 50min account for 100.0%.

[0120] Tidal bore propagation time (Cangqian-Qibao) forecast:

[0121] A total of 839 tidal propagation time data from Cangqian Station to Qibao Station on the Qiantang River from 2009 to 2017 were used to calibrate the model, of which the first 500 tidal bore data were used for model calibration, and the last 339 tidal bore data were used for model verification.

[0122] The parameter optimization results show that when the learning rate lr is 8.95067436e-03, the number of training times Tm is 7.49267233e+03, the momentum term beta1 is 9.11203271e-01, and beta2 is 9.92642971e-01, the minimum objective function is 2.66.

[0123] Figure 6-2 The prediction effect of the propagation time from Cangqian to Qibao Station during the high tide was compared. The model prediction results show that: in the rate period, the root mean square error of the results obtained by the Qiantang River tidal bore propagation time prediction method proposed in this paper is 3.45min, the average absolute error is 2.68min, the data with errors within 5min account for 85.0%, the data with errors within 10min account for 99.4%, and the data with errors within 15min account for 100.0%. In the verification period, the root mean square error of the results obtained by the Qiantang River tidal bore propagation time prediction method proposed in this paper is 4.43min, the average absolute error is 3.48min, the data with errors within 5min account for 75.5%, the data with errors within 10min account for 97.9%, the data with errors within 15min account for 99.4%, and the data with errors within 20min account for 100.0%.

[0124] Tidal high tide forecast:

[0125] A total of 839 high tide data from Cangqian Station, a representative tidal station of the Qiantang River, from 2009 to 2017 were used to calibrate the model. The first 500 tidal bore data were used for model calibration, and the last 339 tidal bore data were used for model verification.

[0126] The parameter optimization results show that when the learning rate lr is 6.06847982e-03, the number of training times Tm is 6.04312846e+03, the momentum term beta1 is 9.61614625e-01, and beta2 is 9.92211934e-01, the minimum objective function is 0.15.

[0127] Figure 6-3 The prediction effect of the high tide level of the Cangqian station during the high tide was compared. The model prediction results show that: in the regular period, the root mean square error of the results of the Qiantang River tidal high tide level prediction method proposed in this paper is 0.18m, the average absolute error is 0.14m, the data with an error within 0.10m accounts for 40.8%, the data with an error within 0.20m accounts for 73.8%, the data with an error within 0.30m accounts for 90.8%, the data with an error within 0.40m accounts for 97.0%, and the data with an error within 0.50m accounts for 98.6%. During the verification period, the root mean square error of the results obtained by the Qiantang River tidal high tide prediction method proposed in this paper was 0.31m, the average absolute error was 0.24m, 24.8% of the data had an error within 0.10m, 49.3% of the data had an error within 0.20m, 67.0% of the data had an error within 0.30m, 81.7% of the data had an error within 0.40m, and 90.9% of the data had an error within 0.50m.

[0128] Tidal low tide forecast:

[0129] A total of 804 high tide and low tide data from Cangqian Station, a representative tidal bore station of the Qiantang River, from 2009 to 2017 were used to calibrate the model. The first 500 tidal bore data were used for model calibration, and the last 304 tidal bore data were used for model verification.

[0130] The parameter optimization results show that when the learning rate lr is 4.67495752e-03, the number of training times Tm is 6.26156595e+03, the momentum term beta1 is 9.47412254e-01, and beta2 is 9.95244827e-01, the objective function is minimized to 0.09.

[0131] Figure 6-4The prediction effect of the low tide level of Cangqian Station during the high tide was compared. The model prediction results show that: in the period of rate, the root mean square error of the results of the Qiantang River tidal low tide prediction method proposed in this paper is 0.12m, the average absolute error is 0.08m, the data with an error within 0.10m accounts for 71.8%, the data with an error within 0.20m accounts for 92.6%, the data with an error within 0.30m accounts for 97.0%, the data with an error within 0.40m accounts for 99.0%, and the data with an error within 0.50m accounts for 99.6%. During the verification period, the root mean square error of the results obtained by the Qiantang River tidal low tide prediction method proposed in this paper was 0.22m, the average absolute error was 0.15m, 50.3% of the data had an error within 0.10m, 76.0% of the data had an error within 0.20m, 87.8% of the data had an error within 0.30m, 92.4% of the data had an error within 0.40m, and 96.1% of the data had an error within 0.50m.

[0132] Example 3

[0133] Based on the above embodiments, this embodiment compares the prediction effects of three algorithms, namely, convolutional neural network CNN, support vector machine SVM, and BP neural network, during the verification period. The accuracy comparison of the three models is shown in Tables 5 to 8.

[0134] Table 5 Comparison results of the accuracy of three models of the time difference between two days at Cangqian Station

[0135]

[0136]

[0137] Table 6 Comparison results of the accuracy of three models for the propagation time from Cangqian to Qibao station

[0138]

[0139] Table 7 Comparison results of the accuracy of three high tide models at Cangqian station

[0140]

[0141] Table 8 Comparison results of the accuracy of three low tide models at Cangqian station

[0142]

[0143] From the table above, we can see that from the perspective of the daily time difference and propagation time prediction of the tidal bore arrival time, the root mean square error, mean absolute error and the proportion of errors within 10 minutes of the daily time difference are arranged in order from small to large: convolutional neural network CNN, support vector machine SVM, BP neural network, and convolutional neural network CNN has obvious advantages. From the perspective of tidal bore high tide and low tide prediction, the prediction accuracy of convolutional neural network CNN and support vector machine SVM is not much different, the prediction accuracy of support vector machine SVM is relatively good, and the prediction accuracy of BP neural network is relatively poor.

[0144] Example 4

[0145] Based on the above embodiments, this embodiment is an electronic device, characterized in that it includes: at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the river tidal forecasting method based on convolutional neural network as described.

[0146] Example 5

[0147] Based on the above embodiments, this embodiment is a computer-readable storage medium, characterized in that computer execution instructions are stored in the computer-readable storage medium. When the processor executes the computer execution instructions, the river tidal forecasting method based on convolutional neural network as described is implemented.

Claims

1. A river tidal bore forecasting method based on a convolutional neural network, the method comprising: Preprocessing of river tidal bore data, determining the influencing factors of river tidal bore through correlation analysis, and constructing a feature set of river tidal bore data; Factors influencing river tidal bores include the lunar calendar date, meteorological conditions, upstream water, river topography, and the previous day's information on tidal bore forecasting factors; Establishment of river tidal bore forecasting model, training the river tidal bore forecasting model based on the river tidal bore data feature set; the forecasting model includes CNN network and fully connected network; Optimization of hyperparameters of river tidal bore forecasting model: The model hyperparameters are optimized by particle swarm optimization algorithm to determine the prediction model of river tidal bore forecasting; The establishment of river tidal bore forecasting model includes: Construct the input matrix. The two-dimensional convolutional neural network takes a three-dimensional matrix of H×W as the input parameter, where the height of H is the number of rows of the input matrix and the width of W is the number of columns of the input matrix; Select the convolution kernel. For the CNN network, use 10 2×2 convolution kernels and pad the edges of the input matrix with zeros. Select the activation function and use the ReLU activation function; Model pooling, pooling the output of the global mean; A prediction model is established based on the fully connected layer. Each node of the input layer of the fully connected layer is connected to all the nodes of the CNN output layer, and the input features are integrated to predict the characteristics of river tidal bores. The output of the convolutional layer is a three-dimensional matrix of 10×7×7, which can be flattened to obtain a one-dimensional vector of 1×490 as the input of the fully connected layer. The fully connected layer consists of two hidden layers. The first layer consists of 100 neurons, and the second layer, i.e., the output layer, consists of one neuron for one output, using the LINEAR activation function.

2. The method for predicting river tidal bore based on convolutional neural network according to claim 1, characterized in that: The preprocessing of river tidal bore data includes: Acquisition of historical data samples of river tidal bores; Data preprocessing, normalization of feature set data, normalization of feature set according to formula 1; In formula 1, x′ is the normalized tidal bore data; x is the data before preprocessing; x max 、x min are the maximum and minimum values ​​of each feature in the training set respectively; Preprocessing of meteorological data, including wind force and direction, and processing of wind characteristic data using formula 2; w e = -w*sinθ Formula 2 In formula 2, w e is the wind force in the east direction after conversion, w is the measured wind force, and θ is the measured wind direction; Correlation analysis was conducted on the factors affecting river tidal bores and the previous day's information on lunar calendar date, meteorological conditions, upstream water, river topography, and tidal bore forecasting factors; The construction of the feature set selects the influencing factors of the previous day's information such as lunar calendar date, meteorological conditions, upstream water, river topography, and tidal bore forecasting factors as feature vectors to construct the feature set as the input of the forecast model.

3. The method for predicting river tidal bore based on convolutional neural network according to claim 1, characterized in that: The optimization methods of hyperparameters of river tidal bore forecasting model include: The training of the prediction model uses the root mean square error (RMSE) as the cost function for model training; In formula 3, is the predicted output value of the model target quantity, and y is the true value of the target quantity; Model parameter optimization algorithm, the convolutional neural network parameters are optimized by particle swarm optimization algorithm; the training sample set T is divided into two non-intersecting subsets Ttrain and Tvalid, Ttrain is used for model training, and Tvalid is used for model validation; the parameter optimization problem with model learning rate lr, number of training times Tm, momentum terms beta1 and beta2 as variables is recorded as: min lr,Tm,beta1,beta2 F=RMSE(T Valid )。 4. The method for predicting river tidal bore based on convolutional neural network according to claim 1, characterized in that: The particle swarm optimization algorithm methods include: S1, initialization parameters, initialization population size Q, maximum number of iterations J, Q particle positions x i and speed v i ; v i =v i +c1×rand()×(pbest i -x i )+c2×rand()×(gbest i -x i )Formula 4 x i =x i +v i Formula 5 Where i = 1, 2, ..., N, N is the total number of particle swarms; v i is the speed of the particle; rand() is a random number between (0,1); x i is the current position of the particle; c1, c2 are learning factors, v i The maximum value is V max , and greater than 0; S2, evaluates the position of each particle based on the river tidal bore forecasting model; S3, compare the current position of each particle with the optimal position P best If the current position is good, record the new position and update P best ; S4, calculate and update the speed and position of each particle according to Formula 4 and Formula 5; S5, determine the number of iterations. If the number of iterations is less than J, return to S2, otherwise end.

5. An electronic device, characterized in that: include: at least one processor and memory; The memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the river tidal forecasting method based on convolutional neural network as described in any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions. When the processor executes the computer-executable instructions, the river tidal bore forecasting method based on a convolutional neural network as described in any one of claims 1 to 4 is implemented.

Images