Medium and long term runoff multi-forecast-period combined forecasting method based on seq2seqLSTM

Abstract

The invention discloses a seq2seqLSTM-based medium and long term runoff multi-forecast-period combined forecasting method. The method comprises the following steps: step 1, collecting and sorting monthly runoff, monthly rainfall and climate remote correlation factor data of a research basin; 2, calculating linear and nonlinear coefficients of runoff and remote correlation factors of early-stage runoff, early-stage rainfall and early-stage climate, and selecting a high-correlation factor as a runoff forecasting factor; 3, combining different types of forecasting factors, and constructing a runoff forecasting factor scheme; step 4, constructing a multi-forecast-period forecasting model based on a sequence-to-sequence long and short term memory neural network seq2seqLSTM; 5, selecting target function training model parameters by taking the time sequence of the first 70% length of the data as a calibration period; and step 6, taking a time sequence with the last 30% length of the data as a test period, and verifying a model effect. According to the method, the correlation of adjacent runoff sequences is fully utilized, and the seq2seqLSTM model is constructed, so that medium and long-term runoff in multiple prediction periods can be predicted at the same time, and the medium and long-term runoff prediction precision is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrological forecasting, and particularly relates to a method for multi-forecast period joint forecasting of medium and long-term runoff based on seq2seqLSTM. Background Art

[0002] Medium and long-term runoff forecasting can provide scientific decision-making support for basin water resources planning and management, reservoir optimal scheduling, and social sustainable development, help to understand the future water resources supply situation, ensure the sustainable utilization of water resources in the future for a long time, and is one of the basic bases for formulating the scheduling plan of water conservancy projects.

[0003] Due to the long forecast period and often lack of reliable meteorological forecast information, medium and long-term runoff forecasting is usually based on the variation law of the water cycle, using known information such as meteorology, climate, and hydrology as input conditions, and adopting methods such as physical cause analysis method and mathematical statistics method to predict future runoff. However, due to the action of the sea-land cycle, basin runoff may be affected by complex factors such as astronomical factors, atmospheric circulation, and ocean thermal conditions, and it is difficult to identify the key forecast factors of runoff; on the other hand, under climate change, the basin hydrological situation changes accordingly, increasing the complexity of runoff formation. How to improve the forecast period and forecast accuracy of medium and long-term runoff forecasting is a key problem faced by the development and utilization of basin water resources. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for multi-forecast period joint forecasting of medium and long-term runoff based on seq2seqLSTM, which realizes the simultaneous forecasting of multi-forecast period runoff and improves the runoff forecasting accuracy by fully mining the correlation of adjacent time runoff sequences.

[0005] To solve the above technical problem, the technical solution adopted by the present invention is as follows:

[0006] A method for multi-forecast period joint forecasting of medium and long-term runoff based on seq2seqLSTM, comprising the following steps: Step1, collection and arrangement of monthly runoff, monthly rainfall, and climate teleconnection factors in the basin;

[0007] Step2, screening of runoff forecast factors;

[0008] Step3, construction of runoff forecast factor schemes; constructing different forecast factor combination schemes based on the runoff factors, rainfall factors, and climate remote sensing related factors screened in Step2;

[0009] Step4, construction of a multi-forecast period forecasting model based on the sequence-to-sequence long short-term memory neural network seq2seqLSTM; Step5, taking the time series with a preset ratio length before the data as the calibration period, and selecting the objective function to train the model parameters;

[0010] Step 6: Use the data with a set ratio length time series as the test period to obtain monthly runoff simulation values for different lead time lengths and verify the model performance.

[0011] In Step 6 above, the performance of the model is evaluated. The indicators used for model performance evaluation are the Nash-Sutcliffe efficiency coefficient NSE and the Kling-Gupta efficiency coefficient KGE; and the best forecasting scheme for the test period is selected through performance evaluation.

[0012] The formulas for the above Nash-Sutcliffe efficiency coefficient NSE and Kling-Gupta efficiency coefficient KGE are as follows:

[0013]

[0014] In the formula, Q o,i (m 3 / s) is the observed runoff, Q s,i (m 3 / s) is the simulated runoff, is the average value of the observed runoff, r is the linear correlation coefficient between the observed and simulated runoff, μ o and μ s are the mean values of the observed runoff and the simulated runoff respectively, σ o and σ s are the standard deviations of the observed runoff and the simulated runoff respectively, and n is the length of the runoff series.

[0015] In Step 1 above, the climate teleconnection factors include atmospheric circulation factors, sea surface temperature factors, and climate monitoring factors of sunspot indices.

[0016] In Step 2 above, the screening of runoff forecasting factors specifically includes:

[0017] The forecasting variable is the monthly runoff at different lags. The forecasting factors include the previous runoff, the previous rainfall, and the previous climate teleconnection factors. The Pearson correlation coefficient, Kendall correlation coefficient, and maximum mutual information coefficient between the forecasting factors and the forecasting variable are calculated respectively, and the coefficient threshold is set to screen the forecasting factors.

[0018] In Step 3 above, the different forecasting factor combination schemes specifically include:

[0019] Single type of factors: runoff factor, rainfall factor, climate teleconnection factor;

[0020] Combination of two factors: runoff factor and rainfall factor, rainfall factor and climate teleconnection factor, runoff factor and climate teleconnection factor;

[0021] Combination of three types of factors in total, 7 forecasting factor schemes;

[0022] In the above Step 3, the specific process of constructing the multi-forecast period prediction model based on the sequence-to-sequence long short-term memory neural network seq2seqLSTM is as follows:

[0023] Construct a multi-dimensional input matrix with the previous sequence of the forecasting factors and the monthly runoff sequence of future multi-forecast periods as the output vector, and construct a sequence-to-sequence long short-term memory neural network LSTM model, that is, the Seq2Seq-LSTM model. The Seq2Seq model consists of two recurrent neural networks: an encoder Encoder and a decoder Decoder.

[0024] In the above Step 4, the specific process of constructing a multi-dimensional input matrix with the previous sequence of the forecasting factors and the monthly runoff sequence of future multi-forecast periods as the output vector is as follows:

[0025] Suppose there are m forecasting factors collected, and the previous t time steps of data for each factor are taken as input; for the i-th sample, its input data is represented as a three-dimensional tensor X i ∈R t×m , where the j-th row of X i represents the values of m forecasting factors at the j-th time step, that is, X i,j =[x i,j,1 ,x i,j,2 ,…,x i,j,m ; the output is the monthly runoff sequence of future multi-forecast periods. The future multi-forecast periods are set to k time steps, and the output of the i-th sample is represented as a vector Y i ∈r k , that is, Y i =[y i,1 ,y i,2 ,…,y i,k .

[0026] In the above Step 4, the specific process of constructing the sequence-to-sequence long short-term memory neural network LSTM model is as follows:

[0027] To make the data of different forecasting factors comparable and avoid adverse effects on model training due to data scale differences, it is necessary to standardize the input data. The commonly used standardization method is Z-score standardization, and its formula is:

[0028]

[0029] where x is the original data, μ is the mean of the data, and σ is the standard deviation of the data; the same standardization process is also performed on the output monthly runoff data;

[0030] Construct input and output sequences:

[0031] Input sequence:

[0032] Suppose there are m predictors in total, and the data of the previous T time steps of each predictor are taken as the input. For a time series data with a length of N, the dimension of the input matrix X is (N - T, T, m); when constructing, starting from the (T + 1)-th time step, the m predictor data of the previous T time steps are successively intercepted as an input sample;

[0033] Output sequence:

[0034] Taking the monthly runoff sequence of future multiple prediction periods as the output vector; if the maximum prediction period is L months, the dimension of the output matrix Y is (N - T, L); also starting from the (T + 1)-th time step, the monthly runoff data of the future L time steps are selected as the corresponding output samples.

[0035] In the above Step4, the construction processes of the two recurrent neural networks, the encoder Encoder and the decoder Decoder, in the Seq2Seq model are as follows:

[0036] 1. Construction of the encoder Encoder:

[0037] 1.1 Initialization of parameters:

[0038] Define the hidden layer dimension d of the LSTM cell, and randomly initialize the weight matrix and bias vector of the LSTM cell in the encoder:

[0039] Input gate weight matrix W xi ∈R m×d , W hi ∈R d×d , bias vector b i ∈R d ;

[0040] Forget gate weight matrix W xf ∈R m×d , W hf ∈R d×d , bias vector b f ∈R d ;

[0041] Candidate cell state weight matrix W xc ∈R m×d , W hc ∈R d×d , bias vector b c ∈R d ;

[0042] Output gate weight matrix W xo ∈R m×d , W ho ∈r d×d , Wco ∈ ℝ d×d , the bias vector b o ∈ ℝ d ;

[0043] 1.2 Forward Propagation

[0044] For each sample Xi in the input sequence X i (i = 1, 2, …, N - T), the following calculations are performed at each time step t (t = 1, 2, …, T):

[0045] Input gate:

[0046] i t = σ(WiX xi + Wh i,t + b hi ) t-1 ; i )

[0047] Forget gate:

[0048] f t = σ(WfX xf + Wh i,t + b hf ) t-1 ; f )

[0049] Candidate cell state:

[0050]

[0051] Cell state update:

[0052]

[0053] Output gate:

[0054] o t = σ(WoX xo + Wh i,t + Wc ho h t-1 + b co c t ) o ;

[0055] Hidden state update:

[0056] h t = o t ⊙ tanh(c t )

[0057] where the initial hidden state h0 and cell state c0 are set to all-zero vectors, σ is the Sigmoid function, tanh is the hyperbolic tangent function, and ⊙ denotes element-wise multiplication;

[0058] The hidden state h of the encoder at the last time step T is used as the context vector C, i.e., C = h T ;

[0059] 2. Construction of the decoder Decoder

[0060] 2.1 Initialization of parameters

[0061] Randomly initialize the weight matrix and bias vector of the LSTM unit in the decoder:

[0062] Input gate weight matrix Bias vector

[0063] Forget gate weight matrix Bias vector

[0064] Candidate cell state weight matrix Bias vector

[0065] Output gate weight matrix Bias vector

[0066] At the same time, initialize the weight matrix W of the fully connected layer out ∈R d×1 and the bias vector b out ∈R 1 ;

[0067] 2.2 Forward propagation

[0068] For the context vector C corresponding to each input sample, the decoder performs the following calculations at each time step l (l = 1, 2,..., L):

[0069] Initial state setting: Set the input y0 at the first time step to 0;

[0070] Input gate:

[0071]

[0072] Forget gate:

[0073]

[0074] Candidate cell state:

[0075]

[0076] Cell state update:

[0077]

[0078] Output gate:

[0079]

[0080] Hidden state update:

[0081]

[0082] Predicted output:

[0083]

[0084] A medium and long-term runoff multi-forecast period joint forecasting method based on seq2seqLSTM provided by the present invention has the following beneficial effects:

[0085] 1. Existing technologies often separately model the runoff for a specific forecast period of a certain section. For runoff forecasting of different forecast periods, a model needs to be reconstructed. The present invention constructs a medium and long-term runoff forecasting model for multiple forecast periods, which can forecast the runoff of multiple forecast periods simultaneously.

[0086] 2. Existing technologies usually do not fully utilize the autocorrelation of the runoff sequence. Affected by the continuity of processes such as climate, meteorology, and hydrology, there is a certain autocorrelation in the runoff sequence over time. The present invention fully explores the correlation of the runoff sequence by jointly forecasting the runoff at adjacent times, which can improve the accuracy of runoff forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The present invention will be further described below with reference to the drawings and embodiments:

[0088] Figure 1 is a schematic flow chart of the present invention;

[0089] Figure 2 is a graph of runoff forecasting result indicators of a single-forecast-period LSTM model at different forecast periods;

[0090] Figure 3 is a graph of runoff forecasting result indicators of a seq2seqLSTM model constructed with rainfall factor p and climate teleconnection factor t as forecasting factors;

[0091] Figure 4 is a graph of runoff forecasting result indicators of a seq2seqLSTM model constructed with rainfall factor p as a forecasting factor. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0092] The technical solution of the present invention will be described in detail below with reference to the drawings and embodiments.

[0093] Embodiment 1:

[0094] AsFigure 1 As shown in Figure 1 , a medium- and long-term runoff multi-forecast period joint forecasting method based on seq2seqLSTM includes the following steps:

[0095] Step1. Collection and arrangement of monthly runoff, monthly rainfall, and climate teleconnection factors in the basin;

[0096] Step2. Screening of runoff forecasting factors;

[0097] Step3. Construction of runoff forecasting factor schemes; constructing different forecasting factor combination schemes based on the runoff factors, rainfall factors, and climate remote sensing related factors screened in Step2;

[0098] Step4. Construction of a multi-forecast period forecasting model based on the sequence-to-sequence long short-term memory neural network seq2seqLSTM; Step5. Using a time series with a preset ratio, such as 70% of the length, of the data as the calibration period, and selecting the objective function to train the model parameters;

[0099] Step6. Using a time series with a preset ratio, such as 30% of the length, of the data as the verification period, obtaining monthly runoff simulation values for different forecast period lengths, and verifying the model effect.

[0100] In the above Step6, the performance of the model is evaluated. The indicators used for model performance evaluation are the Nash-Sutcliffe efficiency coefficient NSE and the Kling-Gupta efficiency coefficient KGE; and the best forecasting scheme for the verification period is selected through performance evaluation.

[0101] The formulas for the above Nash-Sutcliffe efficiency coefficient NSE and Kling-Gupta efficiency coefficient KGE are:

[0102]

[0103]

[0104] In the formula, Q o,i (m 3 / s) is the observed runoff, Q s,i (m 3 / s) is the simulated runoff, is the average value of the observed runoff, r is the linear correlation coefficient between the observed and simulated runoff, μ o and μ s are the mean values of the observed runoff and the simulated runoff respectively, σ o and σ s are the standard deviations of the observed runoff and the simulated runoff respectively, and n is the length of the runoff sequence.

[0105] In the above Step 1, the climate teleconnection factors include 88 atmospheric circulation factors, 26 sea surface temperature factors, and 16 other factors including typhoons, sunspot activities, and sea surface temperature, for a total of 130 climate monitoring factors.

[0106] In the above Step 2, the screening of runoff forecasting factors specifically includes:

[0107] The forecast variable is the monthly runoff at different lags. The forecast factors include the previous runoff, previous rainfall, and previous climate teleconnection factors. The Pearson correlation coefficient, Kendall correlation coefficient, and maximum mutual information coefficient between the forecast factors and the forecast variable are calculated respectively, and a coefficient threshold is set to screen the forecast factors.

[0108] In the above Step 3, the specific combinations of different forecast factor schemes include:

[0109] Single type of factors: runoff factors, rainfall factors, climate teleconnection factors;

[0110] Combination of two factors: runoff factor and rainfall factor, rainfall factor and climate teleconnection factor, runoff factor and climate teleconnection factor;

[0111] Combination of all three types of factors, a total of 7 forecast factor schemes;

[0112] In the above Step 3, the specific process of constructing a multi-forecast period prediction model based on the sequence-to-sequence long short-term memory neural network seq2seqLSTM is as follows:

[0113] Construct a multi-dimensional input matrix with the previous sequence of forecast factors and an output vector of the monthly runoff sequence for future multi-forecast periods, and construct a sequence-to-sequence long short-term memory neural network LSTM model, that is, the Seq2Seq-LSTM model. The Seq2Seq model is usually composed of two recurrent neural networks, an encoder Encoder and a decoder Decoder. It is generally suitable for processing the mapping relationship between two sequences of different lengths and is suitable for multi-step prediction problems. The combination of LSTM and Seq2Seq models can achieve a long short-term memory model with different input and output step lengths.

[0114] In the above Step 4, the specific process of constructing a multi-dimensional input matrix with the previous sequence of forecast factors and an output vector of the monthly runoff sequence for future multi-forecast periods is as follows:

[0115] Suppose there are m forecast factors collected, and the previous t time step data of each factor is taken as the input; for the i-th sample, its input data is represented as a three-dimensional tensor X i ∈R t×m , where the j-th row of X i represents the m forecast factor values at the j-th time step, that is, X i,j = [xi,j,1 , x i,j,2 , …, x i,j,m ; The output is the monthly runoff sequence for multiple future prediction periods. The multiple future prediction periods are set to k time steps, and the output of the i-th sample is represented as a vector Y i ∈R k , that is, Y i = [y i,1 , y i,2 , …, y i,k .

[0116] In the above Step 4, the specific process of constructing the sequence-to-sequence long short-term memory neural network LSTM model is as follows:

[0117] To make the data of different prediction factors comparable and avoid adverse effects on model training due to data scale differences, it is necessary to standardize the input data. The commonly used standardization method is Z-score standardization, and its formula is:

[0118]

[0119] where x is the original data, μ is the mean of the data, and σ is the standard deviation of the data; the same standardization process is also applied to the output monthly runoff data;

[0120] Construct input and output sequences:

[0121] Input sequence:

[0122] Suppose there are m prediction factors, and the data of the previous T time steps of each factor are taken as input. For a time series data with a length of N, the dimension of the input matrix X is (N - T, T, m); when constructing, starting from the (T + 1)-th time step, the m prediction factor data of the previous T time steps are successively intercepted as an input sample;

[0123] Output sequence:

[0124] The monthly runoff sequence for multiple future prediction periods is used as the output vector; if the maximum prediction period is L months, the dimension of the output matrix Y is (N - T, L); starting from the (T + 1)-th time step, the monthly runoff data of the next L time steps are selected as the corresponding output samples.

[0125] In the above Step 4, the construction processes of the two recurrent neural networks, the encoder Encoder and the decoder Decoder, in the Seq2Seq model are as follows:

[0126] 1. Construction of the encoder Encoder:

[0127] 1.1 Initialization of parameters:

[0128] Define the hidden layer dimension \(d\) of the LSTM cell, and randomly initialize the weight matrix and bias vector of the LSTM cell in the encoder:

[0129] Input gate weight matrix \(W\) xi \(\in\mathbb{R}\) m×d \(W\) hi \(\in\mathbb{R}\) d×d , bias vector \(b\) i \(\in\mathbb{R}\) d ;

[0130] Forget gate weight matrix \(W\) xf \(\in\mathbb{R}\) m×d \(W\) hf \(\in\mathbb{R}\) d×d , bias vector \(b\) f \(\in\mathbb{R}\) d ;

[0131] Candidate cell state weight matrix \(W\) xc \(\in\mathbb{R}\) m×d \(W\) hc \(\in\mathbb{R}\) d×d , bias vector \(b\) c \(\in\mathbb{R}\) d ;

[0132] Output gate weight matrix \(W\) xo \(\in\mathbb{R}\) m×d \(W\) ho \(\in\mathbb{R}\) d×d , \(W\) co \(\in\mathbb{R}\) d×d , bias vector \(b\) o \(\in\mathbb{R}\) d ;

[0133] 1.2 Forward propagation

[0134] For each sample \(X\) in the input sequence \(X\) i \((i = 1, 2, \ldots, N - T)\), perform the following calculations at each time step \(t (t = 1, 2, \ldots, T)\):

[0135] Input gate:

[0136] \(i\) t \(=\sigma(W\) xi \(X\) i,t \(+W\) hi \(h\) t-1 \(+b\) i );

[0137] Forget gate:

[0138] \(f\) t \(=\sigma(W\) xf \(X\) i,t \(+W\) hf \(h\) t-1 \(+b\)f )

[0139] Candidate cell state:

[0140]

[0141] Cell state update:

[0142]

[0143] Output gate:

[0144] o t = σ(W xo X i,t + W ho h t-1 + W co c t + b o )

[0145] Hidden state update:

[0146] h t = o t ⊙ tanh(c t )

[0147] Wherein, the initial hidden state h0 and cell state C0 are usually set to all-zero vectors, σ is the Sigmoid function, tanh is the hyperbolic tangent function, and ⊙ represents element-wise multiplication;

[0148] The hidden state h T at the last time step of the encoder is used as the context vector C, i.e., C = h T ;

[0149] 2. Construction of the decoder Decoder

[0150] 2.1 Initialization of parameters

[0151] Randomly initialize the weight matrix and bias vector of the LSTM cell in the decoder:

[0152] Input gate weight matrix Bias vector

[0153] Forget gate weight matrix Bias vector

[0154] Candidate cell state weight matrix Bias vector

[0155] Output gate weight matrix Bias vector

[0156] Meanwhile, initialize the weight matrix W of the fully connected layer out ∈R d×1 and the bias vector b out ∈R 1 ;

[0157] 2.2 Forward Propagation

[0158] For the context vector C corresponding to each input sample, the decoder performs the following calculations at each time step l (l = 1, 2, …, L):

[0159] Initial state setting: Set the input y0 at the first time step to 0;

[0160] Input gate:

[0161]

[0162] Forget gate:

[0163]

[0164] Candidate cell state:

[0165]

[0166] Cell state update:

[0167]

[0168] Output gate:

[0169]

[0170] Hidden state update:

[0171]

[0172] Predicted output:

[0173]

[0174] 3. Model Training

[0175] 3.1 Define the loss function

[0176] The commonly used loss function is the mean squared error (MSE), and the formula is:

[0177]

[0178] where y i,l is the true monthly runoff value of the i-th sample at the l-th prediction period, is the corresponding predicted value.

[0179] 3.2 Selection of optimization algorithm

[0180] Optimization algorithms such as Adam and Stochastic Gradient Descent (SGD) can be selected to update the parameters of the model. Taking the Adam algorithm as an example, it dynamically adjusts the learning rate of each parameter by calculating the first-order moment estimate and second-order moment estimate of the gradient.

[0181] 3.3 Training process

[0182] Perform multiple iterative trainings on the calibration data. In each iteration:

[0183] Forward propagation: Pass the input data through the encoder and decoder to obtain the predicted output.

[0184] Calculate the loss: Calculate the value of the loss function based on the predicted output and the true output.

[0185] Backward propagation: Calculate the gradient of the loss function with respect to the model parameters.

[0186] Parameter update: Use the optimization algorithm to update the parameters of the model according to the gradient (including the weight matrices and bias vectors of the encoder and decoder).

[0187] Example 2:

[0188] Above the dam site of the Three Gorges Reservoir is the upper reaches of the Yangtze River Basin. The main stream of the upper reaches of the Yangtze River is 4,504 km long, and the basin area is 1 million km 2 .

[0189] Step 1. Collect and organize the monthly runoff, monthly rainfall, and climate teleconnection factors in the upper reaches of the Yangtze River Basin. The runoff data is the monthly natural runoff sequence data of the Three Gorges Reservoir dam site from 1959 to 2022, including the monthly observed runoff at Yichang Station from 1959 to 2002 and the restored runoff into the Three Gorges Reservoir from 2003 to 2022; the rainfall data includes the monthly average surface rainfall data of 11 first-level sub-regions in the upper reaches of the Yangtze River from 1959 to 2022. The 11 first-level sub-regions are the upper reaches of the Jinsha River, the middle reaches of the Jinsha River, the lower reaches of the Jinsha River, the Yalong River, the Hengjiang River, the Minjiang River, the Tuojiang River, the Jialing River, the Wujiang River, Xiangjiaba-Cuntan, and Cuntan-Three Gorges; the climate teleconnection factors include 130 climate monitoring factors, including 88 atmospheric circulation factors, 26 sea surface temperature factors, and 16 other factors including typhoons, sunspot activities, and sea surface temperature from 1959 to 2022 (http: / / cmdp.ncc-cma.net / cn / monitoring.htm).

[0190] Step 2. Screening runoff prediction factors. The prediction variable is the monthly runoff at different lags. The alternative prediction factors include antecedent runoff, antecedent rainfall, and antecedent climate teleconnection factors. Calculate the Pearson correlation coefficient, Kendall correlation coefficient, and maximum mutual information coefficient between the alternative prediction factors and the prediction variable to characterize the linear and non-linear relationships between the alternative prediction factors and the prediction variable. Select those with the absolute value of the coefficient greater than 0.5 as the prediction factors for subsequent model construction. The selected runoff prediction factors are shown in Table 1.

[0191] Table 1. Selection of runoff prediction factors

[0192]

[0193]

[0194] Step 3. Construction of runoff prediction factor schemes. Based on the runoff factors, rainfall factors, and climate teleconnection factors screened in Step 2, seven prediction factor schemes are constructed. The prediction schemes are shown in Table 2. Specifically, they are single-type prediction factor schemes, pairwise combinations of different types of prediction factors, and combined prediction schemes of three types of prediction factors. The single-type prediction factor schemes include using rainfall factor (precipitation, p), runoff factor (streamflow, q), and climate teleconnection factor (teleconnection factor, t) as prediction factors respectively. The pairwise combination schemes of different types of prediction factors include using rainfall factor and runoff factor (pq), rainfall factor and climate teleconnection factor (pt), and runoff factor and climate teleconnection factor (qt) as prediction factors respectively. The combined prediction scheme of three types of prediction factors is using rainfall factor, runoff factor, and climate teleconnection factor (pqt) as prediction factors together.

[0195] Table 2. Prediction factor schemes

[0196]

[0197] Step 4. Construction of a multi-forecast period prediction model based on a sequence-to-sequence long short-term memory neural network (seq2seq LSTM) model. For the forecast factor scheme in Step 3, a multi-dimensional input matrix is constructed using the previous twelve-month sequences of the forecast factors, and the monthly runoff sequences for multiple future forecast periods (maximum forecast period = 1, 2,..., 12 months) are used as the output vectors to construct a sequence-to-sequence long short-term memory neural network LSTM (Seq2Seq-LSTM) model. The Seq2Seq model usually consists of two recurrent neural networks, an encoder and a decoder, and is generally suitable for processing the mapping relationship between two sequences of different lengths and is suitable for multi-step prediction problems. The combination of LSTM and Seq2Seq models can achieve a long short-term memory model with different input and output step lengths. At the same time, single-forecast period LSTM models (forecast period = 1, 2,..., 12 months) are constructed for the monthly runoff of different forecast periods, and the forecast results are used for comparison.

[0198] Step 5. Use the first 70% length of the time series of the data as the calibration period, that is, from January 1959 to July 2001, and select the mean square error as the loss function to train the model parameters.

[0199] Step 6. Use the last 30% length of the time series of the data as the verification period, that is, from July 2001 to December 2022, to obtain the simulated values of the monthly runoff for different forecast period lengths and verify the model performance. The model performance evaluation indicators are the Nash-Sutcliffe efficiency coefficient (NSE) and the Kling-Gupta efficiency coefficient (KGE). Select the best forecast scheme for the verification period.

[0200]

[0201] Where Q o,i (m 3 / s) is the observed runoff, Q s,i (m 3 / s) is the simulated runoff, is the average value of the observed runoff, r is the linear correlation coefficient between the observed and simulated runoff, μ o and μ s are the mean values of the observed and simulated runoff respectively, σ o and σ s are the standard deviations of the observed and simulated runoff respectively, and n is the length of the runoff sequence.

[0202] The single-forecast period LSTM model forecasts the monthly runoff for different forecast period lengths respectively. The model forecast period lengths are 1, 2,..., 12 months, and the model is evaluated during the verification period. Figure 2The runoff forecast result index diagrams of the single-horizon LSTM model with different horizons are given. The lines of different colors represent the models constructed based on different forecast factor schemes (p, q, t, pq, pt, qt, pqt). Table 3 statistically shows the average index of the runoff forecast results of the model at each horizon. The evaluation index results of NSE and KGE show that the scheme with rainfall factors and climate teleconnection factors (pt) as forecast factors has an average NSE of 0.81 and an average KGE of 0.86 at each horizon, with the best comprehensive performance. The scheme with rainfall factor (p) as the forecast factor ranks second.

[0203] The multi-horizon seq2seqLSTM model can simultaneously achieve monthly runoff forecasts for multiple horizons. The maximum horizon lengths of the model are 1, 2,..., 12 months, and the model is evaluated during the test period. Figure 3 It is the multi-horizon forecast result index diagram of the seq2seqLSTM model constructed with rainfall factors and climate teleconnection factors (pt) as forecast factors. Figure 4 It is the multi-horizon forecast result index diagram of the seq2seqLSTM model constructed with rainfall factor (p) as the forecast factor. The lines of different colors represent the forecast models with maximum horizons of 1, 2,..., 12 months. Taking the multi-horizon seq2seqLSTM model with a maximum horizon of 12 months as an example, Table 3 statistically shows the average index of the forecast results of the model at each horizon. When each forecast factor scheme is converted from a single-horizon LSTM model to a multi-horizon joint forecast seq2seqLSTM model, the forecast indicators are basically improved. In the embodiment, the medium- and long-term runoff joint forecast method based on the seq2seqLSTM model can simultaneously forecast the runoff for multiple horizons and improve the medium- and long-term runoff forecast accuracy at the same time.

[0204] Table 3. Average index of runoff forecast results of the LSTM model at each horizon (test period)

[0205]

Claims

1. A medium- and long-term runoff multi-forecast period joint forecasting method based on seq2seqLSTM, characterized in that: The following steps are involved: Step 1: Collection and collation of monthly runoff, monthly rainfall, and climate teleconnection factors in the basin; Step 2, screening of runoff prediction factors; Step 3: Construct runoff prediction factor schemes; construct different prediction factor combination schemes based on the runoff factors, rainfall factors and climate remote sensing related factors selected in Step 2; Step 4: Construct a multi-forecast period forecast model based on the sequence-to-sequence long short-term memory neural network seq2seqLSTM; Step 5. Set the time series with the ratio length set before the data as the rate period and select the objective function training model parameters; Step 6. Set the ratio length time series after the data as the test period, obtain the monthly runoff simulation values ​​of different forecast period lengths, and verify the model effect.

2. A medium- and long-term runoff multi-forecast period joint forecasting method based on seq2seqLSTM according to claim 1, characterized in that: In the Step 6, the performance of the model is evaluated, and the indicators used for the model performance evaluation are the Nash-Sutcliffe efficiency coefficient NSE and the Kling-Gupta efficiency coefficient KGE; and the best forecasting scheme in the test period is selected through the performance evaluation.

3. A medium- and long-term runoff multi-forecast period joint forecasting method based on seq2seqLSTM according to claim 2, characterized in that: The Nash-Sutcliffe efficiency coefficient NSE and Kling-Gupta efficiency coefficient KGE formulas are: In the formula, Q o,i (m 3 / s) is the observed runoff, Q s,i (m 3 / s) is the simulated runoff, is the average value of observed runoff, r is the linear correlation coefficient between observed and simulated runoff, μ o and μ s are the means of observed runoff and simulated runoff, σ o and σ s are the standard deviations of observed and simulated runoff, respectively, and n is the length of the runoff series.

4. A medium- and long-term runoff multi-forecast period joint forecasting method based on seq2seqLSTM according to claim 3, characterized in that: The climate teleconnection factors in Step 1 include climate monitoring factors of atmospheric circulation factors, sea temperature factors and sunspot index.

5. A medium- and long-term runoff multi-forecast period joint forecasting method based on seq2seqLSTM according to claim 4, characterized in that: The screening of runoff prediction factors in Step 2 specifically includes: The prediction variables are monthly runoff with different lags, and the prediction factors include previous runoff, previous rainfall and previous climate teleconnection factors. The Pearson correlation coefficient, Kendall correlation coefficient and maximum mutual information coefficient between the prediction factors and the prediction variables are calculated respectively, and the coefficient threshold is set to screen the prediction factors.

6. A medium- and long-term runoff multi-forecast period joint forecasting method based on seq2seqLSTM according to claim 5, characterized in that: The different prediction factor combination schemes in Step 3 specifically include: Single factor: runoff factor, rainfall factor, climate teleconnection factor; Combination of two factors: runoff factor and rainfall factor, rainfall factor and climate teleconnection factor, runoff factor and climate teleconnection factor; There are a total of 7 prediction factor schemes combined with the three types of factors.

7. A medium- and long-term runoff multi-forecast period joint forecasting method based on seq2seqLSTM according to claim 6, characterized in that: The specific process of constructing the multi-forecast period forecast model based on the sequence-to-sequence long short-term memory neural network seq2seqLS TM in Step 4 is as follows: A multi-dimensional input matrix is ​​constructed using the previous sequence of forecast factors, and the monthly runoff sequence for multiple forecast periods in the future is used as the output vector. A sequence-to-sequence long short-term memory neural network LSTM model, namely the Seq2Seq-LSTM model, is constructed. The Seq2Seq model consists of two recurrent neural networks: an encoder and a decoder.

8. The method for medium- and long-term runoff multi-forecast period joint forecasting based on seq2seqLSTM according to claim 7, characterized in that: In the above-mentioned Step 4, the specific process of constructing a multi-dimensional input matrix with the previous sequence of forecast factors and the monthly runoff sequence of multiple forecast periods in the future as the output vector is as follows: Assume that there are m prediction factors collected, and each factor takes the data of the previous t time steps as input; for the i-th sample, its input data is represented as a three-dimensional tensor X i ∈R t×m , where X i The j-th row of represents the m prediction factor values ​​at the j-th time step, that is, X i,j =[x i,j,1 ,x i,j,2 ,…,x i,j,m ]; the output is the monthly runoff sequence for the future multi-forecast period, the future multi-forecast period is set to k time steps, and the output of the i-th sample is represented by the vector Y i =[y i,1 ,y i,2 ,…,y i,k ].

9. A medium- and long-term runoff multi-forecast period joint forecasting method based on seq2seqLSTM according to claim 8, characterized in that: In the above Step 4, the specific process of constructing a sequence-to-sequence long short-term memory neural network LSTM model is as follows: In order to make the data of different predictors comparable and avoid adverse effects on model training due to differences in data scales, the input data needs to be standardized; the standardization method is Z-score standardization, and its formula is: Among them, x is the original data, μ is the mean of the data, and σ is the standard deviation of the data; the same standardization process is also performed on the output monthly runoff data; Construct the input and output sequence: Input sequence: Assume that there are m prediction factors in total, and each factor takes the data of the previous T time steps as input. For a time series data with a length of N, the dimension of the input matrix X is (NT, T, m); when constructing, starting from the (T+1)th time step, the m prediction factor data of the previous T time steps are sequentially intercepted as an input sample; Output sequence: The monthly runoff sequence of multiple forecast periods in the future is used as the output vector; if the maximum forecast period is L months, the dimension of the output matrix Y is (NT, L); similarly, starting from the T+1th time step, the monthly runoff data of the next L time steps are selected as the corresponding output samples.

10. The method for medium- and long-term runoff multi-forecast period joint forecasting based on seq2seqLSTM according to claim 9, characterized in that: In the above Step 4, the construction process of the two recurrent neural networks, the encoder Encoder and the decoder Decoder in the Seq2Seq model is as follows:

1. Encoder construction; 1.1 Initialize parameters; define the hidden layer dimension d of the LSTM unit, and randomly initialize the weight matrix and bias vector of the LSTM unit in the encoder; 1.2 Forward propagation; 2. Decoder construction; 2.1 Initialization parameters; Randomly initialize the weight matrix and bias vector of the LSTM unit in the decoder; 2.2 Forward propagation.

Images