Watershed runoff prediction method based on multi-dimensional hydrological information

Abstract

The invention belongs to the technical field of drainage basin runoff prediction, and particularly relates to a drainage basin runoff prediction method based on multi-dimensional hydrological information. Comprising the following steps: step 1, constructing a medium-and-long-term runoff forecasting model: for hydrological data and meteorological data, analyzing correlation between influence factors and runoff by adopting a correlation coefficient method, screening out high-quality forecasting factors, and performing multi-factor forecasting; and step 2, constructing an LSTM network model: simulating an information processing system of a biological nervous system function, processing information through a neural network to complete a training task, and constructing the LSTM network model by referring to a neuron structure. According to the method, a Markov chain correction model is introduced from the angle of error correction, the method is combined with an artificial intelligence deep learning neural network model, an LSTM annual runoff prediction model based on Markov chain correction is established, and it is verified that the model is superior in medium and long term runoff prediction in combination with evaluation indexes.

Description

Technical Field

[0001] The invention belongs to the technical field of watershed runoff prediction, and in particular relates to a watershed runoff prediction method based on multi-dimensional hydrological information. Background Art

[0002] In recent years, due to the combined effects of meteorology, human activities and other influencing factors, the generation and development of runoff processes in hydrological phenomena have been affected to a certain extent. The model prediction results obtained by the single-factor prediction scheme based on the time series analysis method have large errors and lack reliability. Therefore, many scholars are currently trying to use correlation analysis methods and other methods to perform multi-factor predictions to provide a reliable data basis for subsequent runoff forecast models. Dong Qianjin et al. proposed a remote sensing hydrological prediction model based on the correlation coefficient method and random forest. At the same time, the differences in model prediction accuracy were compared and the results were verified based on the differences in correlation coefficients and the number of factors. Zhu Chunmiao et al. used the correlation coefficient method to screen the prediction factors in the artificial intelligence runoff forecast model, and used high-quality prediction factors as data input for subsequent prediction models. Shao Shuangshuang et al. studied the relationship between basin runoff and climate change through gray correlation and correlation analysis methods, and analyzed the annual distribution and variation characteristics of basin runoff and the driving effect of temperature and precipitation on the annual variation of runoff. The results show that the three have their own annual variation characteristics and have a strong correlation. Nong Zhenxue et al. used traditional factors such as rainfall and runoff as well as related factors such as atmospheric circulation indicators as prediction factors, and screened out high-quality prediction factors using methods such as correlation coefficient method, stepwise regression method and principal component analysis. On this basis, they established a medium- and long-term runoff prediction model BP for multi-factor prediction. Liu et al. used correlation analysis and other methods to select the leading driving factors of runoff from hydrological and meteorological indicators such as precipitation, soil moisture content and evaporation as prediction factors, and then used them as inputs for subsequent models;

[0003] The complexity and uncertainty of the runoff prediction process of the existing runoff prediction methods lead to large errors in the prediction results. The random fluctuations in the runoff process and the large errors in the results of the traditional runoff prediction model make the prediction results lack accuracy and reliability and cannot achieve satisfactory prediction effects. Therefore, we propose a basin runoff prediction method based on multidimensional hydrological information to solve the above problems. Summary of the invention

[0004] The purpose of the present invention is to provide a basin runoff prediction method based on multidimensional hydrological information, which can introduce a Markov chain correction model from the perspective of error correction, combine this method with a deep learning neural network model of artificial intelligence, and establish an LSTM annual runoff prediction model based on Markov chain correction. Combined with various evaluation indicators, it is verified that the proposed model performs superiorly in medium- and long-term runoff forecasting.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A basin runoff prediction method based on multi-dimensional hydrological information comprises the following steps:

[0007] Step 1. Construct a medium- and long-term runoff forecast model: The correlation coefficient method is used to analyze the correlation between various influencing factors and runoff for hydrological data and meteorological data, and select high-quality forecasting factors for multi-factor prediction;

[0008] Step 2. Build an LSTM network model: an information processing system that imitates the functions of biological nervous systems. Information is processed by a neural network to complete the training task. The LSTM network model is built based on the neuron structure.

[0009] Step 3. Data analysis preprocessing: In view of the problem that the runoff process is affected by multiple factors and has non-stationary and complex characteristics, a trend analysis of the runoff data in the study area is conducted, and a multi-factor prediction scheme is studied considering the influence of meteorological factors;

[0010] Step 4. Model parameter setting: Based on the collected basin hydrological data and meteorological data, prediction scheme 1 selects the runoff data of the first 17 and 19 time periods as the input of the artificial intelligence model to predict the flow value of the next time period and perform cyclic rolling prediction;

[0011] Step 5. Build an LSTM runoff prediction model: store information in memory cells one by one, and form a conveyor belt through a small amount of linear interactions between memory cells in different hidden layers;

[0012] Step 5. Construct an LSTM annual runoff prediction model based on Markov chain correction: predict the scenario data, use the artificial intelligence runoff prediction model to obtain the initial prediction results, and perform error correction through the Markov chain correction model.

[0013] In a preferred embodiment, the construction of the medium- and long-term runoff forecast model also includes an error correction model, which uses error correction technology and various available information to timely correct the calculated values ​​of the original model based on the error information generated by the measured values ​​and the model forecast values.

[0014] In a preferred embodiment, the construction of the LSTM network model also includes dividing the LSTM network into a fully connected layer and an LSTM cell layer, and the LSTM network includes four types of layers: (a) an input layer that accepts input sequence data; (b) a fully connected layer that converts the input data into the LSTM cell layer dimension and establishes a bridge between the input layer and the LSTM cell layer; (c) an LSTM cell layer that provides different memory capabilities; and (d) an output layer that converts the LSTM unit into an output.

[0015] In a preferred embodiment, the construction of the LSTM network model further includes using two or more influencing factors as independent variables to explain the changes in the dependent variable, and applying it to the statistical correlation analysis between the basin runoff and various influencing factors, and its mathematical model is:

[0016] y=β0+β1x1+β2x2+…+β n x n

[0017] In the formula: β0, β1, β2, ..β n There are n+1 unknown variables, x1, x2, ..x n is n sample values, and y is the statistical prediction value of the forecast object.

[0018] In a preferred embodiment, the data analysis preprocessing also includes predicting the correlation between the influencing factors and the runoff, and using Mann-Kendall significance and Spearman correlation to reflect the correlation coefficient, which is represented by the letter r:

[0019]

[0020] Where: r is the correlation coefficient between x and y;

[0021] x i is the i-th sample value of the impact factor;

[0022] y i is the i-th sample value of the forecast object;

[0023] is the sample average of the impact factor;

[0024] is the sample mean of the forecast object;

[0025] n is the number of samples, and the correlation coefficient r can be both positive and negative, and its value is between [-1,1], that is, positive correlation (r>0) and negative correlation (r<0).

[0026] In a preferred embodiment, the model parameter setting also includes selecting 7 factors, namely, maximum wind speed, average air pressure and maximum temperature, monthly sunshine percentage, precipitation, sunshine hours, evapotranspiration and number of days with daily precipitation>0.1 mm, average air pressure, average temperature, average 2-minute wind speed, average maximum temperature, precipitation, and maximum wind speed, as independent variables, predicting runoff as the dependent variable, constructing a BP neural network, setting the maximum number of selections to 1000 times, the target error to 0.000001, the learning rate to 0.01, using the default Sigmoid function as the network activation function, the Levenberg-Marquardt function as the network training function, and the default function MSE as the network performance function: the number of neurons in the input layer of multi-factor prediction is 7; the number of neurons in the output layer of the scheme is 1, that is, the runoff.

[0027] In a preferred embodiment, the construction of the LSTM runoff prediction model also includes introducing a "gate" structure to add or delete information in memory cells to control the flow of information. x is the multiplication of the corresponding elements in the operation matrix. The two multiplication matrices are required to be of the same type. The + sign represents matrix addition. Ct-1 is the output of the current neuron, Ht-1 is the parameter of the hidden layer, and the architecture diagram shows three gate units, the forget gate, the input gate and the output gate. The inputs of the forget gate and the input gate are the input Xt of the current time and the data of the previous hidden layer. The input of the output gate is the current output.

[0028] In a preferred embodiment, the construction of the LSTM runoff prediction model also includes adding a key variable, the cell state, to store long-term memory information, and adjusting the cell state by three gate structures, namely, the forget gate F, the input gate IN, and the output gate O, wherein the forget gate determines the information of the cell state at time t that needs to be removed from the cell state at time t-1, the input gate determines the new information that needs to be stored in the cell state at time t, the output gate determines the information that needs to be output from the cell state at time t, and the cell state at time t records the input at time t, the gate structure information, the hidden layer state at time t-1, and the cell state at time t-1. A hydrological model suitable for multi-year runoff simulation is established based on LSTM, and the specific calculation process of the model is as follows:

[0029] ① Standardize the precipitation P and temperature T at time t as model input;

[0030] I t =[f n (P t )f n (T t )]

[0031] ② Remove the information of the cell state at time t-1 through the forget gate (F);

[0032] F t=σ[W F (H t-1 , I t )+b F ]

[0033] ③ Determine the information Ct used to update the cell state through the input gate (IN);

[0034] IN t =σ[W UN (H t-1 , I t )+b IN ]

[0035]

[0036] ④ Calculate the output of the cell state and the generated hidden layer state variable H through the output gate (O);

[0037] O t =σ[W o (H t-1 , I t )+b o ] H t =O t tanh(C t )

[0038] ⑤ After de-standardizing the model output, the simulated runoff RM of the study area is obtained;

[0039] RM t =f d (O t )

[0040] Where: Pt is the precipitation at time t (mm); Tt is the temperature at time t (℃); It is the input of the model at time r (including precipitation and temperature at time t); R is the runoff at time t simulated by the model (m·s'); W F , b F are the weight matrix and bias of the forget gate (F), respectively; Ht-1 and Ht are the hidden layer state variables at time t-1 and t, respectively; W IN , b IN They are the weight matrix and bias of the input gate (IN) respectively; is the update vector of the cell state (C); W C , b C are the weight matrix and bias of the cell state update vector, respectively; W. and b. are the weight matrix and bias of the output gate, respectively.

[0041] In a preferred embodiment, the construction of the LSTM annual runoff prediction model based on Markov chain correction also includes using the state transition probability matrix to infer the state at a certain moment in the future based on the initial known state at the moment, wherein the specific model correction steps are as follows:

[0042] (1) According to the absolute value of the relative error between the predicted value data and the measured value data of the model training set samples, the appropriate sample space state is divided:

[0043] (2) Determine the state transition probability matrix P;

[0044] (3) Solve the probability distribution of future states based on the matrix P;

[0045] The key to reducing the prediction error by using the Markov chain correction model is to use Matlab software to determine the state transition probability matrix P, determine the error variation range based on the model training set data, and then determine the transition probability matrix of the next state and the sum of the probabilities of each sample space state. The prediction value error of the model test set is in the state with the largest sample space state probability. The Markov chain correction formula can be expressed as:

[0046]

[0047] Where: f(x) represents the initial prediction result of the model;

[0048] F(x) represents the prediction result after the Markov chain correction;

[0049] Δ represents the relative error between the initial predicted value and the measured value

[0050] ΔU represents the maximum value of the relative error interval between the predicted value and the actual value of the model training set:

[0051] ΔD represents the minimum value of the relative error interval between the predicted value and the actual value of the model training set.

[0052] In a preferred embodiment, the construction of the LSTM annual runoff prediction model based on Markov chain correction also includes introducing the Markov chain correction model into the annual runoff prediction. The key to the Markov chain correction model is the determination of the state probability matrix P. Based on the initial prediction result of the LSTM model, the Markov chain correction model is used to perform error correction to construct the LSTM model based on Markov chain correction. The specific modeling steps are as follows:

[0053] (1) The collected hydrological data sets of the research basin were normalized and divided into training sets and test sets;

[0054] (2) The mean square error of the training set and the test set is selected as the fitness function, and the LSTM model is constructed for simulation prediction;

[0055] (3) Calculate the absolute value of the relative error between the training set sample data obtained by simulation training and the predicted value to divide the sample space state;

[0056] (4) Determine the state transition probability matrix P using the Markov chain correction model;

[0057] (5) Based on the data of matrix P, the probability and sum of each sample space state are calculated. The state where the prediction value error is located is the state with the largest probability sum;

[0058] (6) Combine the Markov chain correction formula to obtain the final prediction value.

[0059] The technical effects achieved by the present invention are:

[0060] Focusing on issues such as data analysis preprocessing, artificial intelligence forecasting models, and error correction models, this paper conducts exploratory research on annual runoff prediction for rivers. Based on the collected basin hydrological and meteorological data, a multi-factor prediction scheme is used. The data of the above forecasting scheme is used as the model input to establish an annual runoff prediction based on an LSTM neural network model. Based on the above research, the Markov chain correction model is introduced from the perspective of error correction. This method is combined with the deep learning neural network model of artificial intelligence to establish an LSTM annual runoff prediction model based on Markov chain correction. Combined with various evaluation indicators, it is verified that the proposed model performs superiorly in medium- and long-term runoff forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a technical route schematic diagram of a basin runoff prediction method based on multi-dimensional hydrological information of the present invention;

[0062] Figure 2 It is a schematic diagram of influencing factors of a basin runoff prediction method based on multi-dimensional hydrological information of the present invention;

[0063] Figure 3 It is a schematic diagram of the LSTM architecture of a basin runoff prediction method based on multi-dimensional hydrological information of the present invention;

[0064] Figure 4 It is a schematic diagram of the LSTM modeling principle of a basin runoff prediction method based on multi-dimensional hydrological information of the present invention;

[0065] Figure 5 It is a prediction flow chart of a basin runoff prediction method based on multi-dimensional hydrological information of the present invention, which is based on a LSTM model corrected by a Markov chain. DETAILED DESCRIPTION

[0066] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0067] See also Figure 1-5 As shown, the present invention provides a basin runoff prediction method based on multi-dimensional hydrological information, comprising the following steps:

[0068] Step 1. Construct a medium- and long-term runoff forecast model: The correlation coefficient method is used to analyze the correlation between various influencing factors and runoff for hydrological data and meteorological data, and select high-quality forecasting factors for multi-factor prediction;

[0069] Step 2. Build an LSTM network model: an information processing system that imitates the functions of biological nervous systems. Information is processed by a neural network to complete the training task. The LSTM network model is built based on the neuron structure.

[0070] Step 3. Data analysis preprocessing: In view of the problem that the runoff process is affected by multiple factors and has non-stationary and complex characteristics, a trend analysis of the runoff data in the study area is conducted, and a multi-factor prediction scheme is studied considering the influence of meteorological factors;

[0071] Step 4. Model parameter setting: Based on the collected basin hydrological data and meteorological data, prediction scheme 1 selects the runoff data of the first 17 and 19 time periods as the input of the artificial intelligence model to predict the flow value of the next time period and perform cyclic rolling prediction;

[0072] Step 5. Build an LSTM runoff prediction model: store information in memory cells one by one, and form a conveyor belt through a small amount of linear interactions between memory cells in different hidden layers;

[0073] Step 5. Construct an LSTM annual runoff prediction model based on Markov chain correction: predict the scenario data, use the artificial intelligence runoff prediction model to obtain the initial prediction results, and perform error correction through the Markov chain correction model.

[0074] The construction of medium- and long-term runoff forecast models also includes an error correction model, which uses error correction technology and various available information to timely correct the calculated values ​​of the original model based on the error information generated by the measured values ​​and the model forecast values.

[0075] Building an LSTM network model also includes dividing the LSTM network into a fully connected layer and an LSTM cell layer. The LSTM network contains four types of layers: (a) an input layer that accepts input sequence data; (b) a fully connected layer that converts the input data into the dimensions of the LSTM cell layer and establishes a bridge between the input layer and the LSTM cell layer; (c) an LSTM cell layer that provides different memory capabilities; and (d) an output layer that converts the LSTM unit into an output.

[0076] The construction of LSTM network model also includes using two or more influencing factors as independent variables to explain the changes of dependent variables, and applying it to the statistical correlation analysis between basin runoff and various influencing factors. Its mathematical model is:

[0077] y=β0+β1x1+β2x2+…+β n x n

[0078] In the formula: β0, β1, β2, ..β n There are n+1 unknown variables, x1, x2, ..x n is n sample values, and y is the statistical prediction value of the forecast object.

[0079] Data analysis preprocessing also includes the correlation between the forecast influencing factors and runoff, and uses Mann-Kendall significance and Spearman correlation to reflect the correlation coefficient, which is represented by the letter r:

[0080]

[0081] Where: r is the correlation coefficient between x and y;

[0082] x i is the i-th sample value of the impact factor;

[0083] y i is the i-th sample value of the forecast object;

[0084] is the sample average of the impact factor;

[0085] is the sample mean of the forecast object;

[0086] n is the number of samples, and the correlation coefficient r can be both positive and negative, and its value is between [-1,1], that is, positive correlation (r>0) and negative correlation (r<0).

[0087] The model parameter setting also includes selecting 7 factors, namely, maximum wind speed, average air pressure and maximum temperature, monthly sunshine percentage, precipitation, sunshine hours, evapotranspiration, number of days with daily precipitation>0.1mm, average air pressure, average temperature, average 2-minute wind speed, average maximum temperature, precipitation, and maximum wind speed, as independent variables, and predicting runoff as the dependent variable, constructing a BP neural network, setting the maximum number of selections to 1000 times, the target error to 0.000001, the learning rate to 0.01, the default Sigmoid function as the network activation function, the Levenberg-Marquardt function as the network training function, and the default function MSE as the network performance function: the number of neurons in the input layer of multi-factor prediction is 7; the number of neurons in the output layer of the scheme is 1, that is, the runoff.

[0088] Building an LSTM runoff prediction model also includes introducing a "gate" structure to add or delete information in memory cells and control the flow of information. x is the multiplication of the corresponding elements in the operation matrix. The two multiplication matrices are required to be of the same type. The + sign represents matrix addition. Ct-1 is the output of the current neuron, Ht-1 is the parameter of the hidden layer, and the architecture diagram shows three gate units, the forget gate, the input gate, and the output gate. The inputs of the forget gate and the input gate are the input Xt of the current time and the data of the previous hidden layer. The input of the output gate is the current output.

[0089] The construction of LSTM runoff prediction model also includes adding the key variable of cell state to store long-term memory information, and adjusting the cell state by three gate structures: forget gate F, input gate IN and output gate 0. Among them, the forget gate determines the information of the cell state at time t that needs to be removed from the cell state at time t-1, the input gate determines the new information that needs to be stored in the cell state at time t, and the output gate determines the information that needs to be output by the cell state at time t. The cell state at time t records the input at time t, the gate structure information, the hidden layer state at time t-1, and the cell state at time t-1. Based on LSTM, a hydrological model suitable for multi-year runoff simulation is established. The specific calculation process of the model is as follows:

[0090] ① Standardize the precipitation P and temperature T at time t as model input;

[0091] I t =[f n (P t )f n (T t )]

[0092] ② Remove the information of the cell state at time t-1 through the forget gate (F);

[0093] F t =σ[W F (H t-1 , It )+b F ]

[0094] ③ Determine the information Ct used to update the cell state through the input gate (IN);

[0095] IN t =σ[W IN (H t-1 , I t )+b IN ]

[0096]

[0097] ④ Calculate the output of the cell state and the generated hidden layer state variable H through the output gate (O);

[0098] O t =σ[W o (H t-1 , I t )+b o ] H t =O t tanh(C t )

[0099] ⑤ After de-standardizing the model output, the simulated runoff RM of the study area is obtained;

[0100] RM t =f d (O t )

[0101] Where: Pt is the precipitation at time t (mm); Tt is the temperature at time t (℃); It is the input of the model at time r (including precipitation and temperature at time t); R is the runoff at time t simulated by the model (m·s'); W F , b F are the weight matrix and bias of the forget gate (F), respectively; Ht-1 and Ht are the hidden layer state variables at time t-1 and t, respectively; W IN , b IN They are the weight matrix and bias of the input gate (IN) respectively; is the update vector of the cell state (C); W C , b C are the weight matrix and bias of the cell state update vector, respectively; W. and b. are the weight matrix and bias of the output gate, respectively.

[0102] The construction of the LSTM annual runoff prediction model based on Markov chain correction also includes using the state transition probability matrix to infer the state at a certain moment in the future based on the initial known state. The specific steps of the model correction are as follows:

[0103] (1) According to the absolute value of the relative error between the predicted value data and the measured value data of the model training set samples, the appropriate sample space state is divided:

[0104] (2) Determine the state transition probability matrix P;

[0105] (3) Solve the probability distribution of future states based on the matrix P;

[0106] The key to reducing the prediction error by using the Markov chain correction model is to use Matlab software to determine the state transition probability matrix P, determine the error variation range based on the model training set data, and then determine the transition probability matrix of the next state and the sum of the probabilities of each sample space state. The prediction value error of the model test set is in the state with the largest sample space state probability. The Markov chain correction formula can be expressed as:

[0107]

[0108] Where: f(x) represents the initial prediction result of the model;

[0109] F(x) represents the prediction result after the Markov chain correction;

[0110] Δ represents the relative error between the initial predicted value and the measured value

[0111] ΔU represents the maximum value of the relative error interval between the predicted value and the actual value of the model training set:

[0112] ΔD represents the minimum value of the relative error interval between the predicted value and the actual value of the model training set.

[0113] Building an LSTM annual runoff prediction model based on Markov chain correction also includes introducing the Markov chain correction model into the annual runoff prediction. The key to the Markov chain correction model is the determination of the state probability matrix P. Based on the initial prediction results of the LSTM model, the Markov chain correction model is used to perform error correction and build an LSTM model based on Markov chain correction. The specific modeling steps are as follows:

[0114] (1) The collected hydrological data sets of the research basin were normalized and divided into training sets and test sets;

[0115] (2) The mean square error of the training set and the test set is selected as the fitness function, and the LSTM model is constructed for simulation prediction;

[0116] (3) Calculate the absolute value of the relative error between the training set sample data obtained by simulation training and the predicted value to divide the sample space state;

[0117] (4) Determine the state transition probability matrix P using the Markov chain correction model;

[0118] (5) Based on the data of matrix P, the probability and sum of each sample space state are calculated. The state where the prediction value error is located is the state with the largest probability sum;

[0119] (6) Combine the Markov chain correction formula to obtain the final prediction value.

[0120] LSTM architecture parameters explained

[0121]

[0122]

[0123]

[0124]

[0125]

[0126] LSTM modeling steps

[0127] The following is the code part, the comments in it represent the specific functions and parameter comments of each function block

[0128]

[0129]

[0130]

[0131]

[0132]

[0133]

[0134]

[0135]

[0136] In the present invention, focusing on the problems of data analysis preprocessing, artificial intelligence forecasting model, error correction model, etc., the annual runoff forecasting exploratory research work is carried out in the Ganhezi basin research area of ​​Changji Prefecture, Xinjiang Uygur Autonomous Region. Based on the collected basin hydrological data and meteorological data, a multi-factor prediction scheme is used. The data of the above forecasting scheme is used as the model input to establish an LSTM neural network model for annual runoff forecasting. Based on the above research, a Markov chain correction model is introduced from the perspective of error correction. This method is combined with the deep learning neural network model of artificial intelligence to establish an LSTM annual runoff prediction model based on Markov chain correction. Combined with various evaluation indicators, it is verified that the proposed model performs superiorly in medium- and long-term runoff forecasting.

[0137] The above is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art unless otherwise specified and limited.

Claims

1. A basin runoff prediction method based on multi-dimensional hydrological information, characterized by: The following steps are involved: Step 1. Construct a medium- and long-term runoff forecast model: The correlation coefficient method is used to analyze the correlation between various influencing factors and runoff for hydrological data and meteorological data, and select high-quality forecasting factors for multi-factor prediction; Step 2. Build an LSTM network model: an information processing system that imitates the functions of biological nervous systems. Information is processed by a neural network to complete the training task. The LSTM network model is built based on the neuron structure. Step 3. Data analysis preprocessing: In view of the problem that the runoff process is affected by multiple factors and has non-stationary and complex characteristics, a trend analysis of the runoff data in the study area is conducted, and a multi-factor prediction scheme is studied considering the influence of meteorological factors; Step 4. Model parameter setting: Based on the collected basin hydrological data and meteorological data, prediction scheme 1 selects the runoff data of the first 17 and 19 time periods as the input of the artificial intelligence model to predict the flow value of the next time period and perform cyclic rolling prediction; Step 5. Build an LSTM runoff prediction model: store information in memory cells one by one, and form a conveyor belt through a small amount of linear interactions between memory cells in different hidden layers; Step 5. Construct an LSTM annual runoff prediction model based on Markov chain correction: predict the scenario data, use the artificial intelligence runoff prediction model to obtain the initial prediction results, and perform error correction through the Markov chain correction model.

2. A method for predicting watershed runoff based on multidimensional hydrological information according to claim 1, characterized in that: The construction of the medium- and long-term runoff forecast model also includes an error correction model, which uses error correction technology and various available information to timely correct the calculated values ​​of the original model based on the error information generated by the measured values ​​and the model forecast values.

3. The method for predicting watershed runoff based on multidimensional hydrological information according to claim 1, characterized in that: The construction of the LSTM network model also includes dividing the LSTM network into a fully connected layer and an LSTM cell layer. The LSTM network includes four types of layers: (a) an input layer that accepts input sequence data; (b) a fully connected layer that converts the input data into the LSTM cell layer dimension and establishes a bridge between the input layer and the LSTM cell layer; (c) an LSTM cell layer that provides different memory capabilities; and (d) an output layer that converts the LSTM unit into an output.

4. The method for predicting watershed runoff based on multidimensional hydrological information according to claim 1, characterized in that: The construction of the LSTM network model also includes using two or more influencing factors as independent variables to explain the changes in the dependent variable, and applying it to the statistical correlation analysis between the basin runoff and various influencing factors. The mathematical model is: y=β0+β1x1+β2x2+…+β n x n In the formula: β0, β1, β2, ..β n There are n+1 unknown variables, x1, x2, ..x n is n sample values, and y is the statistical prediction value of the forecast object.

5. The method for predicting watershed runoff based on multidimensional hydrological information according to claim 1, characterized in that: The data analysis preprocessing also includes the correlation between the forecast influencing factors and runoff, and the Mann-Kendall significance and Spearman correlation are used to reflect the correlation coefficient, which is represented by the letter r: Where: r is the correlation coefficient between x and y; x i is the i-th sample value of the impact factor; y i is the i-th sample value of the forecast object; is the sample average of the impact factor; is the sample mean of the forecast object; n is the number of samples, and the correlation coefficient r can be both positive and negative, and its value is between [-1,1], that is, positive correlation (r>0) and negative correlation (r<0).

6. The method for predicting watershed runoff based on multidimensional hydrological information according to claim 1, characterized in that: The model parameter setting also includes selecting 7 factors, namely, maximum wind speed, average air pressure and maximum temperature, monthly sunshine percentage, precipitation, sunshine hours, evapotranspiration and number of days with daily precipitation>0.1mm, average air pressure, average temperature, average 2-minute wind speed, average maximum temperature, precipitation, and maximum wind speed, as independent variables, predicting runoff as the dependent variable, constructing a BP neural network, setting the maximum number of selections to 1000 times, the target error to 0.000001, the learning rate to 0.01, using the default Sigmoid function as the network activation function, the Levenberg-Marquardt function as the network training function, and the default function MSE as the network performance function: the number of neurons in the input layer of multi-factor prediction is 7; the number of neurons in the output layer of the scheme is 1, that is, the runoff.

7. The method for predicting watershed runoff based on multidimensional hydrological information according to claim 1, characterized in that: The construction of the LSTM runoff prediction model also includes introducing a "gate" structure to add or delete information in memory cells and control the flow of information. x is the multiplication of the corresponding elements in the operation matrix. The two multiplication matrices are required to be of the same type. The + sign represents matrix addition. Ct-1 is the output of the current neuron, Ht-1 is the parameter of the hidden layer, and the architecture diagram shows three gate units, the forget gate, the input gate and the output gate. The inputs of the forget gate and the input gate are the input Xt of the current time and the data of the previous hidden layer. The input of the output gate is the current output.

8. The method for predicting watershed runoff based on multidimensional hydrological information according to claim 1, characterized in that: The construction of the LSTM runoff prediction model also includes adding a key variable, the cell state, to store long-term memory information, and adjusting the cell state by three gate structures: the forget gate F, the input gate IN, and the output gate O. Among them, the forget gate determines the information of the cell state at time t-1 that needs to be removed from the cell state at time t, the input gate determines the new information that needs to be stored in the cell state at time t, and the output gate determines the information that needs to be output by the cell state at time t. The cell state at time t records the input at time t, the gate structure information, the hidden layer state at time t-1, and the cell state at time t-1. Based on LSTM, a hydrological model suitable for multi-year runoff simulation is established. The specific calculation process of the model is as follows: ① Standardize the precipitation P and temperature T at time t as model input; I t =[f n (P t )f n (T t )] ② Remove the information of the cell state at time t-1 through the forget gate (F); F t =σ[W F (H t-1 ,I t )+b F ] ③ Determine the information Ct used to update the cell state through the input gate (IN); IN t =σ[W IN (H t-1 ,I t )+b IN ] ④ Calculate the output of the cell state and the generated hidden layer state variable H through the output gate (O); O t =σ[W o (H t-1 ,I t )+b o ] H t =O t tanh(C t ) ⑤ After de-standardizing the model output, the simulated runoff RM of the study area is obtained; RM t =f d (O t ) Where: Pt is the precipitation at time t (mm); Tt is the temperature at time t (℃); It is the input of the model at time r (including precipitation and temperature at time t); R is the runoff at time t simulated by the model (m·s'); W F 、b F are the weight matrix and bias of the forget gate (F), respectively; Ht-1 and Ht are the hidden layer state variables at time t-1 and t, respectively; W IN 、b IN They are the weight matrix and bias of the input gate (IN) respectively; is the update vector of the cell state (C); W C 、b C are the weight matrix and bias of the cell state update vector, respectively; W. and b. are the weight matrix and bias of the output gate, respectively.

9. The method for predicting watershed runoff based on multidimensional hydrological information according to claim 1, characterized in that: The construction of the LSTM annual runoff prediction model based on Markov chain correction also includes using the state transition probability matrix to infer the state at a certain moment in the future according to the initial known state at the moment, wherein the specific model correction steps are as follows: (1) According to the absolute value of the relative error between the predicted value data and the measured value data of the model training set samples, the appropriate sample space state is divided: (2) Determine the state transition probability matrix P; (3) Solve the probability distribution of future states based on the matrix P; The key to reducing the prediction error by using the Markov chain correction model is to use Matlab software to determine the state transition probability matrix P, determine the error variation range based on the model training set data, and then determine the transition probability matrix of the next state and the sum of the probabilities of each sample space state. The prediction value error of the model test set is in the state with the largest sample space state probability. The Markov chain correction formula can be expressed as: Where: f(x) represents the initial prediction result of the model; F(x) represents the prediction result after the Markov chain correction; Δ represents the relative error between the initial predicted value and the measured value ΔU represents the maximum value of the relative error interval between the predicted value and the actual value of the model training set: ΔD represents the minimum value of the relative error interval between the predicted value and the actual value of the model training set.

10. The method for predicting watershed runoff based on multi-dimensional hydrological information according to claim 1, characterized in that: The construction of the LSTM annual runoff prediction model based on Markov chain correction also includes introducing the Markov chain correction model into the annual runoff prediction. The key to the Markov chain correction model is the determination of the state probability matrix P. Based on the initial prediction result of the LSTM model, the Markov chain correction model is used to perform error correction to construct the LSTM model based on Markov chain correction. The specific modeling steps are as follows: (1) The collected hydrological data sets of the research basin were normalized and divided into training sets and test sets; (2) The mean square error of the training set and the test set is selected as the fitness function, and the LSTM model is constructed for simulation prediction; (3) Calculate the absolute value of the relative error between the training set sample data obtained by simulation training and the predicted value to divide the sample space state; (4) Determine the state transition probability matrix P using the Markov chain correction model; (5) Based on the data of matrix P, the probability and sum of each sample space state are calculated. The state where the prediction value error is located is the state with the largest probability sum; (6) Combine the Markov chain correction formula to obtain the final prediction value.

Images