A deep learning-based waterlogging dot water volume prediction method

Abstract

The application discloses a kind of based on deep learning's inner waterlogging point water-logging volume prediction method, utilize the automatic rainfall observation data or rainfall forecast data of inner waterlogging point, and drainage capacity data and water-logging volume data in the forming process of inner waterlogging point, the water-logging volume of inner waterlogging point future time is predicted by the neural network model established in advance.This application makes full use of the advantage of deep learning in data mining, analyzes historical data, sorts out the relationship between existing urban road waterlogging and precipitation, drainage, etc., combined with precipitation forecast information or monitoring information, makes up for the defects of traditional mechanism model modeling, rating, low calculation efficiency, can quickly give urban road waterlogging profile prediction, provides technical support for water affairs flood control department to quickly, timely respond to flood control task.

Description

TECHNICAL FIELD

[0001] The present application relates to a waterlogging point water accumulation prediction method. BACKGROUND

[0002] At present, waterlogging and water accumulation depth simulation prediction mainly adopts physical mechanism models based on hydrology and hydrodynamics. Such mechanism models require a large amount of data such as terrain, underlying surface, soil type, and drainage pipe network. Due to the complexity of the required data, model building is difficult and prediction is inaccurate. In view of the above problems, it is expected to realize rapid and accurate waterlogging water accumulation prediction by fully utilizing monitoring resources and adopting deep learning technology, to meet the timeliness requirements of water affairs flood control departments in coping with flood control tasks, and to solve the problem of too many data required for traditional mechanism model building, calibration and verification. SUMMARY

[0003] The present application aims to provide a waterlogging point water accumulation prediction method based on deep learning.

[0004] Technical scheme: A waterlogging point water accumulation prediction method based on deep learning, comprising: using automatic rainfall observation data or rainfall prediction data of the waterlogging point, and drainage volume data and water accumulation data in the formation process of the waterlogging point, and predicting the water accumulation of the waterlogging point at a future time through a pre-established neural network model.

[0005] Further, the establishment of the neural network model comprises the following steps:

[0006] S1: Obtain the historical data of the waterlogging point, including automatic rainfall observation data or rainfall prediction data, drainage volume data, and water accumulation data in the historical formation process of the waterlogging point; wherein the historical data are time series values;

[0007] S2: Preprocess the historical data, and then use a sliding time window method to select input vector x(t)={x a (t),x b (t),x c (t)} and output vector y(t)={{q hb+L},{q hb+L+1},…,{q t}} of the neural network model respectively, and divide them into training set samples and test set samples; wherein the width H of the sliding time window is ha+hb+hc, ha, hb, and hc represent the embedding dimension of rainfall time series x a (t), water accumulation time series x b (t), and drainage volume time series x c (t) respectively, {q t} represents the water accumulation prediction time series, and L represents the prediction step.

[0008] S3: training and testing the neural network model using the training set samples.

[0009] Further, when the automatic rainfall observation data or the rainfall forecast data is point rainfall data of points around the waterlogging point, a method of fitting a quadratic surface and then interpolating is used to calculate surface rainfall data covering positions corresponding to the point rainfall data, and then point rainfall data of the waterlogging point region is obtained according to the coordinates of the waterlogging point.

[0010] Further, in S2, the data preprocessing includes normalizing each type of data by using the maximum and minimum value method.

[0011] Further, in S3, during the training of the neural network model, the number of nodes of the input layer and the output layer, the number of hidden layer nodes, and the neural network training parameters of the neural network model are first set according to the embedding dimension of the training sample time series, then the training set samples are input into the neural network model, and the number of hidden layer nodes and the training parameters of the neural network model are adjusted during the training process until a termination condition is reached, and a function mapping relationship between the input vector and the output vector is obtained.

[0012] Further, according to the function mapping relationship between the input vector and the output vector, an iterative prediction method is used to obtain the waterlogging amount prediction value at the future time.

[0013] Further, the root mean square error is used as a prediction index to analyze and evaluate the model prediction results.

[0014] Beneficial effects: The present application makes full use of the advantages of deep learning in data mining, analyzes historical data, sorts out the relationship between existing urban road waterlogging and precipitation, drainage, etc., combines precipitation forecast information or monitoring information, and makes up for the defects of traditional mechanism modeling, such as difficulty in rating and low calculation efficiency, and can quickly give a prediction of urban road waterlogging, providing technical support for water affairs flood control departments to quickly and timely respond to flood control tasks. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A flowchart is established for the neural network model of the present application. DETAILED DESCRIPTION

[0016] The present application will be further explained below in conjunction with the drawings.

[0017] A waterlogging point waterlogging amount prediction method based on deep learning, comprising: using automatic rainfall observation data or rainfall forecast data of a waterlogging point, and drainage amount data and waterlogging amount data in the formation process of the waterlogging point, to predict the waterlogging amount of the waterlogging point at a future time during rainfall through a pre-established neural network model.

[0018] As Figure 1 shown, the establishment of the neural network model includes the following steps:

[0019] S1: Obtain historical data of the waterlogging point, including automatic rainfall observation data or rainfall forecast data in the historical formation process of the waterlogging point, drainage capacity data, and water accumulation data, all of which are time series values.

[0020] S2: Preprocess the historical data, then respectively select input vector x(t) = {x a (t), x b (t), x c (t)} and output vector y(t) = {{q hb+L}, {q hb+L+1}, …, {q t}} of the neural network model by using a sliding time window method, and divide them into training set samples and test set samples. The width H of the sliding time window is ha+hb+hc, ha, hb, and hc respectively represent the embedding dimension of the rainfall time series x a (t), the water accumulation time series x b (t), and the drainage capacity time series x c (t), {q t} represents the water accumulation prediction time series, and L represents the prediction step.

[0021] S3: Train and test the neural network model using the training set samples.

[0022] In the method, when the automatic rainfall observation data or the rainfall forecast data is point rainfall data around the waterlogging point, a method of fitting a quadratic surface and then interpolating is used to calculate the surface rainfall data corresponding to the position of the point rainfall data, and then the point rainfall data of the waterlogging point region is obtained according to the coordinates of the waterlogging point.

[0023] The surface rainfall refers to the average precipitation per unit area when urban rainfall occurs, which can objectively and accurately show the level of rainfall in a certain region of the city, is a very important parameter index in the precipitation forecast process, and is also a very important link in the road water accumulation big data analysis model. Since the automatic rainfall observation station or the numerical prediction product of the meteorological department is discrete points, the rainfall directly estimated is point rainfall, which must be processed by some calculation method to obtain the surface rainfall information in the urban waterlogging prediction. The surface rainfall calculation method of fitting a quadratic surface and then interpolating is used to solve the surface rainfall in the present application. Assuming that the number of automatic rainfall monitoring stations is N, the point rainfall q(x, y) at the position (x, y) of the waterlogging point can be calculated by selecting the monitoring data of the nearest n (n≤N) automatic rainfall monitoring stations.n It is indicated that the more the number of selected automatic rainfall monitoring sites, the more accurate the fitted quadric surface.

[0024] The quadric surface equation is:

[0025] q(x, y) = a1x 2 +a2y 2 +a3xy+a4x+a5y+a6 (1)

[0026] The error sum of n points is:

[0027]

[0028] According to the principle of least squares, the error sum R of the point rainfall of n automatic rainfall monitoring sites fitted by the quadric surface is minimized:

[0029]

[0030] The coefficient a i is obtained from equation (2), which is substituted into equation (1). Finally, the point rainfall q(x, y) at position (x, y) is obtained by substituting the quadric surface fitted by the second difference.

[0031] The waterlogging point monitoring data is usually stored in the system database at a certain time interval in minutes. In this method, the unit of the collected rainfall data is usually millimeters / hour, the unit of the drainage volume data is usually cubic meters / second, and the unit of the water accumulation data is usually meters. In order to eliminate the influence of large data dimension difference on the convergence speed of neural network, and at the same time meet the value range requirement of hidden layer neuron activation function, the data is preprocessed, and the rainfall, drainage volume and water accumulation data are normalized to the interval [0, 1], and the final prediction result of the model is de-normalized for further error analysis.

[0032] Common data normalization methods include maximum and minimum value method, mean and variance method, etc. This method uses the maximum and minimum value method for data preprocessing, and the normalization formula can be expressed as:

[0033]

[0034] The de-normalization formula can be expressed as:

[0035] x = (x max -x min )x' + x min (4)

[0036] In the formula, x and x' represent the original value and the normalized value of the input sample respectively, x max and x minrespectively represent the maximum and minimum values of the input samples.

[0037] In the present application, in order to reasonably select the rainfall, drainage and water accumulation time series for neural network training and prediction, a sliding time window method is used to select the input vector and output vector of the neural network. The sliding time window method is to set a fixed window width H, select H data from the first time series data, and move the window backward as the time series samples increase, discard the old samples, and select the time series data with a data length value of H again, and so on until all the time series data are selected by the sliding time window.

[0038] Assuming that the width of the sliding time window is set to H (H≥1) and the prediction step value is L (L≥1), the sample selection first selects ha length of rainfall data, hc length of drainage data and hb length of water accumulation data as the neural network input vector x(t), then selects the (hb+L)th water accumulation data as the neural network output vector y(t), and then moves backward by unit length to the last water accumulation data, and finally combines the above vectors to form the sample pair U(t). The specific expression is as follows:

[0039] The neural network input vector x(t) is expressed as:

[0040] x(t) = {x a (t), x b (t), x c (t)} (5)

[0041] The rainfall time series is expressed as:

[0042] x a (t) = { {p1, p2, …, p ha} T , {p2, p3, …, p ha+1} T , …, {p t-L-ha , p t-L-ha+1 , …, p t-L} T} (6)

[0043] where p i represents the ith rainfall data in the time series.

[0044] The drainage time series is expressed as:

[0045] x c (t) = { {r1, r2, …, r hc} T , {r2, r3, …, r hc+1}T ,…,{r t-L-hc ,r t-L-hc+1 ,…,r t-L} T} (7)

[0046] wherein r i represents the i-th water discharge data in the time series.

[0047] The water accumulation time series is represented as:

[0048] x b (t)={{q1,q2,…,q hb} T ,{q2,q3,…,q hb+1} T ,…,{q t-L-hb ,q t-L-hb+1 ,…,q t-L} T} (8)

[0049] wherein q i represents the i-th water accumulation data in the time series.

[0050] The output vector is represented as:

[0051] y(t)={{q hb+L},{q hb+L+1},…,{q t}} (9)

[0052] The set sample pair U(t) is represented as:

[0053] U(t)={x(t),y(t)} (10)

[0054] The method is based on the historical data collected at waterlogging points and the relationship between rainfall input and water accumulation output during the waterlogging formation process. The neural network is used as a medium for mapping the function relationship between the input and output of the waterlogging point. The neural network is trained on the time series samples of the collected data to determine the input and output function mapping relationship, and a corresponding model is constructed to predict the future water accumulation trend of the waterlogging monitoring point. The neural network is trained on the time series samples of the pre-rainfall, drainage and water accumulation of a single waterlogging monitoring point, so as to obtain the corresponding input and output function mapping relationship f(·). The waterlogging point rainfall and water accumulation prediction process based on the neural network can be represented as:

[0055]

[0056] In the formula, represents the predicted value of the water accumulation at the L-th moment in the future.

[0057] After the rainfall and water accumulation data are divided into sample pairs U(t) according to a sliding time window, the first 70% of data in the sample pairs U(t) are selected as the model training set samples U train (t) = {x train (t), y train (t)} according to a 7:3 ratio test (t) = {x test (t), y test (t)}.

[0058] In the neural network model training process of step S3, the number of nodes of the input layer and the output layer, the number of hidden layer nodes, and the neural network training parameters of the neural network model are first set according to the embedding dimension of the training sample time series, then the training set samples U train (t) are input into the neural network model, and the number of hidden layer nodes and the training parameters of the neural network model are adjusted in the training process until a termination condition is reached, to obtain the best function mapping relationship of the input vector and the output vector.

[0059] The model test set samples U test (t) are input into the trained prediction model, and the water accumulation prediction value is obtained according to the function mapping relationship of the rainfall and water accumulation time series contained in the obtained model. Generally, time series prediction includes direct prediction and iterative prediction. Considering the characteristics of dense waterlogging data collection and short rainfall process, the iterative prediction method is adopted. Direct prediction is to input the sample U(t) before t time into the neural network, and the prediction value at (t+1) time s (t+1) can be directly obtained by constructing a single-step predictor f The single-step predictor can be represented as:

[0060]

[0061] The prediction result is input into the model again to obtain the prediction value at (t+2) time and the process is repeated m times until the mth iterative prediction result of the time series is obtained. The iterative process can be realized by constructing a multi-step predictor f D (·). The iterative predictor can be represented as:

[0062]

[0063] Model evaluation is an important step in the process of neural network modeling and prediction, and selecting reasonable evaluation indicators can quantitatively analyze and evaluate the prediction performance of the waterlogging point rainfall water accumulation prediction model based on neural network. Common prediction model evaluation indicators mainly include mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) and the like. Since the waterlogging data usually contains values of zero, the value of the mean absolute percentage error is closer to zero, indicating that the prediction accuracy of the model is higher. When the real data contains values of zero, the mean absolute percentage error formula cannot be used. Considering the calculation complexity and error value domain expression convenience, the root mean square error is used as the prediction evaluation indicator to analyze and evaluate the model prediction results. The root mean square error is expressed as:

[0064]

[0065] In the formula, n represents the length of the time series to be evaluated, y(t) and y'(t) represent the true value and the predicted value respectively.

[0066] The present application makes full use of the advantages of deep learning in data mining, analyzes historical data, sorts out the relationship between existing urban road waterlogging and precipitation, drainage and the like, combines precipitation forecast information or monitoring information, makes up for the defects of traditional mechanism model modeling, rating and low calculation efficiency, and can quickly give urban road waterlogging prediction.

[0067] The method can also be applied to other environmental surface rainfall prediction applications, and is particularly suitable for use in smart cities and smart water, and can improve the scientificity of early flood control and drainage scheme formulation and late drainage engineering optimization, and lay a technical foundation for big data and intelligent analysis application for industry function departments decision-making.

[0068] The above only describes the preferred embodiments of the present application, and it should be noted that for ordinary skilled persons in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A method for predicting water accumulation at urban flooding points based on deep learning, characterized in that, include: Using automatic rainfall observation data or rainfall forecast data of the waterlogged points, as well as drainage and water accumulation data during the formation process of the waterlogged points, a pre-established neural network model is used to predict the water accumulation at the waterlogged points in the future. The establishment of the neural network model includes the following steps: S1: Obtain historical data of the waterlogged area, including automatic rainfall observation data or rainfall forecast data, drainage data, and water accumulation data during the historical formation process of the waterlogged area; wherein, all historical data are time-series values; S2: Preprocess the historical data, and then use the sliding time window method to select the input vector x(t) = {x} of the neural network model. a (t),x b (t),x c The output vector y(t) = {{q} and the output vector y(t) = {{q} hb+L },{q hb+L+1 },…,{q t The data is divided into training and testing sets; where the width of the sliding time window H = ha + hb + hc, and ha, hb, and hc represent the rainfall time series x, respectively. a (t), water accumulation time series x b (t), Discharge time series x c The embedding dimension of (t), {q t } represents the time series of water accumulation prediction, and L represents the prediction step size; S3: Train and test the neural network model using the training set samples; In S3, during the training process of the neural network model, the number of nodes in the input and output layers, the number of hidden layer nodes, and the neural network training parameters of the neural network model are first set according to the embedding dimension of the training sample time series. Then, the training set samples are input into the neural network model, and the number of hidden layer nodes and training parameters of the neural network model are adjusted during the training process until the termination condition is reached, so as to obtain the function mapping relationship between the input vector and the output vector. Based on the functional mapping relationship between the input vector and the output vector, the predicted value of the water accumulation at future times is obtained by using an iterative prediction method.

2. The method for predicting water accumulation at urban flooding points based on deep learning according to claim 1, characterized in that, When the automatic rainfall observation data or rainfall forecast data is the point rainfall data around the flooded point, the method of fitting a quadratic surface and then interpolating is used to calculate the areal rainfall data covering the corresponding position of the point rainfall data, and then the point rainfall data of the flooded point area is obtained according to the coordinates of the flooded point.

3. The method for predicting water accumulation at urban flooding points based on deep learning according to claim 1, characterized in that, In step S2, data preprocessing includes normalizing each type of data using the maximum-minimum method.

4. The method for predicting water accumulation at urban flooding points based on deep learning according to claim 1, characterized in that, The root mean square error is used as the prediction benchmark to analyze and evaluate the model prediction results.

Images