A real-time rolling prediction method for runoff

Abstract

The present invention provides a real-time rolling forecasting method for runoff. The runoff is forecasted by constructing a distance-weighted self-regularized neural network model. Among them, the training set data is input into the distance-weighted self-regularized neural network model after data preprocessing. The distance-weighted self-regularized neural network model uses multiple groups of LSTM and CNN neural networks to learn the training set data from three directions: points far from the mean, points close to the mean, and index data, then uses the FC layer for data fusion, and calculates the loss through multiple distance-weighted loss functions. Finally, the forecasting result is obtained. During the forecasting process, a sequence with a time step is predicted backward based on the current time. By setting the forecasting time step, the input data is continuously updated, so as to be able to forecast the future runoff process day by day in a rolling manner and achieve real-time forecasting; the first-period forecasting results of each time series are collected, and the combined results are used as the forecasted runoff process for the entire forecasting period to achieve the real-time forecasting effect.

Description

Technical Field

[0001] The present invention relates to the technical field of runoff forecasting, and particularly to a real-time rolling runoff forecasting method. Background Art

[0002] With the continuous growth of the global population and the acceleration of the processes of industrialization and urbanization, the demand for fresh water resources by people has increased sharply. Under such circumstances, how to scientifically predict and manage runoff resources to ensure the stability of life and production and food security has become a key issue to be solved urgently. Rainfall, as the main source of runoff, brings great challenges to runoff forecasting due to the uncertainty of its spatio-temporal distribution and the complexity of being affected by climate change. Coupled with the unique microclimate characteristics of different regions, the complex and changeable underlying surface conditions, and the increasingly enhanced interference of human activities, the runoff formation process is full of variables. Incorporating relevant elements such as upstream rainfall into the forecasting system and analyzing and predicting them as important index sequences helps to improve the accuracy and reliability of runoff forecasting. Such forecasting results can not only provide timely and accurate water resource information for irrigated agriculture, support farmers in scientifically formulating irrigation plans, and reducing water resource waste, but also provide a scientific basis for government departments to formulate water resource management policies and optimize water resource allocation. Moreover, it can give early warnings of risks in dealing with natural disasters such as droughts and floods, guide emergency responses, and ensure the safety of people's lives and property and social stability.

[0003] Traditional runoff prediction methods include univariate autoregressive (AR), moving average (MA), simple exponential smoothing (SES), and extreme learning machine (ELM) algorithms, as well as the classical autoregressive integrated moving average model (ARIMA) method and several of its variants.

[0004] Currently, many general time series prediction models perform poorly on data with high skewness and kurtosis scores, and traditional methods often show biases when facing extreme events. The non-stationarity of runoff sequences and the existence of extreme events increase the difficulty of runoff forecasting. The forecasting model should be able to capture long-term dependence relationships and model rare but important extreme values, while existing methods often have difficulty in dealing with such dual challenges simultaneously. Summary of the Invention

[0005] Aiming at the problems that the non-stationarity of runoff sequences and extreme events increase the difficulty of runoff forecasting, the present invention provides an extreme adaptive model, a real-time rolling runoff forecasting method that enhances the runoff prediction effect by adding polarity learning with the assistance of multiple types of runoff-related sequences.

[0006] The solution adopted by the present invention to solve its technical problems is as follows: A real-time rolling runoff forecasting method, which conducts runoff forecasting through a distance-weighted self-regularized neural network model. The training set of this model is the historical data of each representative year, and the test set is the data of the current year of the representative year. Among them, the training set data is input into the distance-weighted self-regularized neural network model after data preprocessing. The distance-weighted self-regularized neural network model uses multiple groups of LSTM and CNN neural networks to learn the training set data from three directions: points far from the mean, points close to the mean, and index data, then uses the FC layer for data fusion, and calculates the loss through multiple distance-weighted loss functions, and finally obtains the forecasting result;

[0007] During the forecasting process, the distance-weighted self-regularized neural network model predicts a sequence of one time step backward based on the current time, and continuously updates the input data, so as to be able to forecast the future runoff process day by day and realize real-time forecasting;

[0008] Collect the first-period forecasting results of each time series, and the combined result is used as the forecast runoff process for the entire forecasting period, and the real-time forecasting effect of the distance-weighted self-regularized neural network model is evaluated through this runoff process.

[0009] Furthermore, the training set data includes historical runoff sequences and related time series simultaneous with the historical runoff sequences. Among them, the historical runoff sequence x 1 is modeled as an ordinary sequence, and the related time series x 2 to x m are modeled as index sequences. The Gaussian mixture model is used to fit the historical runoff sequence x 1 , and based on x 1 , the Gaussian mixture model GMM index is generated, and this index is used as the related time series x 2 . The formula of the Gaussian mixture model is as follows

[0010]

[0011] In the formula, x is a d-dimensional sample vector (column vector); β i is the mixing weight of the i-th single Gaussian model, and f i (x|μ i , ∑ i ) is the probability density function of the i-th single Gaussian model; μ i is the mathematical expectation of the i-th single Gaussian model; ∑ i is the covariance matrix of the i-th single Gaussian model, and M is a hyperparameter optimized for the time series.

[0012] Generate a probability for each point in the time series through a Gaussian mixture model. This probability represents the likelihood of the observed data point occurring. At the same time, according to the weights obtained during the Gaussian mixture model fitting process, the weighted sum of all component probabilities is used as the index feature for each value in the time series. Through this feature, a measure of the contribution of each component to the entire time series at a specific point is provided.

[0013] Use Kruskal-Wallis sampling to evaluate the normality of the training samples and provide an oversampling strategy. Sample the oversample region with extreme events in the training set to improve the prediction ability of the distance-weighted self-regularized neural network model for extreme events.

[0014] Extract multiple random samples x of size t + h from the input sequence, and split the sequence into k consecutive subsequences of equal size. Then calculate the Kruskal-Wallis test statistic H between the k subsequences. The calculation formula of H is as follows:

[0015]

[0016] In the formula, n is the total number of samples in all groups, R j is the sum of ranks of the j-th group, n j is the number of samples in the j-th group. The time series undergoes preprocessing of logarithmic transformation and standardization. The predicted sequence output by the model is obtained through the inverse operation of standardization and logarithmic transformation operations to get the final prediction result.

[0017] Furthermore, the distance-weighted self-regularized neural network (DAN) model includes a RepGen module and a RepMerg module. The RepGen module includes two stacks, RepGen(E) and RepGen(D), which are used to generate ordinary sequences and refine the polarity representation of the index. Among them, RepGen(E) is the encoder part, and RepGen(D) is the decoder part; RepMerg is responsible for feature fusion. The RepGen contains three parallel encoder-decoder blocks, and the three parallel encoder-decoder blocks are divided into f layer, i layer, and n layer.

[0018] Furthermore, the distance-weighted self-regularized neural network model calculates the loss using multiple distance-weighted loss functions after completing one cycle.

[0019] Advantages of the present invention: The real-time rolling runoff prediction method of the present invention introduces a distance-weighted self-regularized neural network (DAN). This model is a new extreme adaptive model that enhances the runoff prediction effect by adding polar representation learning. The distance-weighted self-regularized neural network uses a distance-weighted multi-loss mechanism and stackable blocks to dynamically refine the index sequence from exogenous data. At the same time, it can also handle univariate time series by adopting Gaussian mixture probability modeling to improve the robustness to severe events. In addition, Kruskal Wallis sampling and gate control vectors are added to the distance-weighted self-regularized neural network to handle unbalanced extreme data. A distance-weighted multi-loss regularization penalty is added to the loss function to update the training parameters. The distance-weighted self-regularized neural network (DAN) model can better predict extreme events with low occurrence frequency but great influence in time series and has excellent performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a framework diagram of the distance-weighted self-regularized neural network model.

[0021] Figure 2 is a structural diagram of the distance-weighted self-regularized neural network model.

[0022] Figure 3 is a schematic diagram of the gate vector.

[0023] Figure 4 is a structural diagram of the distance-weighted self-regularized neural network model.

[0024] Figure 5 is a comparison chart of the prediction results of DAN and SARIMAX in dry years.

[0025] Figure 6 is a comparison chart of the prediction results of DAN and SARIMAX in wet years. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below.

[0027] Embodiment 1: To better handle extreme events in runoff prediction, the present invention provides a real-time rolling runoff prediction method that introduces a distance-weighted self-regularized neural network (DAN) to enhance the runoff prediction effect by adding polar representation learning.

[0028] Specifically, first, a distance-weighted self-regularized neural network (DAN) model is constructed. The method for constructing the distance-weighted self-regularized neural network (DAN) model includes data preprocessing, model construction, and model architecture and parameter setting.

[0029] Among them, data preprocessing includes sequence processing, extreme event sampling, and logarithmic transformation and standardization processing. The input information of the model includes historical runoff sequences and simultaneous relevant sequences as special indicators (such as upstream rainfall sequences). First, by modeling the historical runoff sequence x 1 as an ordinary sequence, and modeling the relevant time series x 2 to x m as indicator sequences to define this task. To better consider the influence time effect of relevant sequences, the model allows users to set the lag time t of each sequence and change the sequence input order internally to achieve better prediction results.

[0030] When no additional indicator sequences are provided, the Gaussian Mixture Model (GMM) is used to fit the historical runoff sequence x 1 , and based on x 1 generate GMM indicators, and this indicator is used as the relevant time series x 2 .

[0031] In this case, the problem is simplified to a univariate time series prediction problem.

[0032] The formula of the Gaussian Mixture Model is as follows

[0033]

[0034]

[0035] In the formula, x is a d-dimensional sample vector (column vector); β i is the mixing weight of the i-th single Gaussian model, and f i (x|μ i , ∑ i ) is the probability density function of the i-th single Gaussian model; μ i is the mathematical expectation of the i-th single Gaussian model; ∑ i is the covariance matrix of the i-th single Gaussian model, and M is a hyperparameter tuned for the time series, set to 3.

[0036] In this embodiment, the Expectation-Maximization (EM) algorithm is adopted to estimate the parameters of the GMM by maximizing the observed data through likelihood iteration. The GMM generates a probability for each point in the time series, indicating the likelihood of the observed data point occurring. At the same time, according to the weights obtained during the GMM model fitting process, the weighted sum of all component probabilities is used as the "indicator" feature of each value in the time series, and this feature provides a measure of the contribution of each component to the entire time series at a specific point.

[0037] Considering that extreme events are rare in the data, the normality of the training samples is evaluated using Kruskal-Wallis sampling and an oversampling strategy is provided to sample the hypersample regions with extreme events in the training set, improving the model's prediction ability for extreme events.

[0038] The model will extract multiple random samples x of size t + h from the input sequence, split the sequence into k equally sized consecutive subsequences, and then calculate the Kruskal-Wallis test statistic H between the k subsequences.

[0039] The calculation formula for H is as follows:

[0040]

[0041] In the formula, n is the total number of samples in all groups, R j is the sum of ranks of the j-th group, and n j is the number of samples in the j-th group.

[0042] To avoid the H statistic being affected by minor differences in the subsequences, the values in x need to be rounded to the nearest integer before calculating H. If H > λ (λ is the set threshold), the sample is included in the training set, or the sample is included with a probability p < 1. The threshold λ affects the relative variation of the samples, making the samples more likely to contain extreme events, while the probability p affects the number of normal samples included in the training set.

[0043] Before training the model, all time series need to undergo preprocessing of logarithmic transformation and standardization. The predicted sequence output by the model can obtain the final prediction result through inverse standardization operation and logarithmic transformation operation.

[0044] Such as Figure 1As shown, the scalable end-to-end framework of the Distance-Weighted Auto-Regularized Neural Network (DAN) model mainly consists of two parts, namely the RepGen (Representation Generating) module and the RepMerg (Representation Merging) module. In the field of neural networks, "representation" usually refers to the internal expression of data or features learned through the model, which can capture the internal structure and relationships of the data. In the Distance-Weighted Auto-Regularized Neural Network (DAN) model, RepGen consists of two stacks, RepGen(E) and RepMerg(D), which are responsible for generating the polar representations of ordinary sequences and refined metrics. Among them, RepGen(E) is the encoder part, and RepGen(D) is the decoder part; while RepMerg is responsible for feature fusion. The Distance-Weighted Auto-Regularized Neural Network (DAN) model can iteratively update the metrics by repeatedly stacking RepGen(E)+RepGen(D), and can also refine the metrics multiple times by repeatedly stacking RepMerg. According to the internal relationships of multivariate sequences, the optimal stack configuration may be different for different datasets. In this embodiment, the stack configuration of RepGen(E)+RepGen(D)+RepMerg is adopted.

[0045] The key innovation of the Distance-Weighted Auto-Regularized Neural Network (DAN) model is the adoption of a new mechanism in generating and exchanging information between the far and near polar representations and the metric (ind), which can directly improve the prediction results. Since extreme values usually follow different distributions, the polar coordinate representation learning performed in RepGen encodes them separately, and then these points are retained in RepMerg. After feature fusion with the preliminary prediction results, the final prediction results are obtained, ensuring that these representations are not affected by the dominance of normal values during the training process.

[0046] RepGen contains three parallel encoder-decoder blocks, as Figure 2 shown, which can convert the refined metrics after the ordinary series input into polar representations, and these representations will be further merged in the RepMerg stack.

[0047] In the RepGen(E) stack, the "f-E" layer is responsible for the representation learning of points far from the sequence mean. including extreme values sparsely distributed in the prediction interval. This layer uses three sets of parallel one-dimensional convolutional neural network layers (CNN1d) and long short-term memory network (LSTM) layers to capture the long-term, medium-term, and short-term extreme value sequence dependencies respectively, and then fuses the relationships they capture to improve the feature extraction ability and computational efficiency of the encoder, better predicting the far point. Moreover, since the hidden state of extreme events may be updated multiple times in the repeating block, combining convolutional operations can shorten the input sequence and alleviate potential gradient explosion or disappearance problems. The "n-E" layer uses LSTM to learn the hidden features of the near point which includes most values that conform to the normal distribution. The "i-E" layer uses LSTM to learn the representation of the metric sequence After learning, RepGen(E) will output the cell state c (cell state) and hidden state h (hidden state) of the near point, far point, and ind, as the input of the other two stacks

[0048] The RepGen(D) stack is divided into three layers: "f-D", "n-D", and "i-D", corresponding one-to-one with the RepGen(D) stack, each receiving the cell state c and hidden state h output by the corresponding layer of RepGen(E). "f-D" and "i-D" further optimize to obtain the far point through LSTM combined with the FC layer Near point The prediction result. It should be noted that is helpful for predicting extreme events, and accurate prediction results are crucial for performance improvement. Therefore, a refinement layer composed of two one-dimensional convolutional neural network components (CNN1d) is added to the "i-D" layer of the model. It will first optimize the values within a local range, extract local temporal features, improve the ability to model complex temporal patterns, and assist in generating the expected metrics. Moreover, since RepGen is designed as an extensible stack and can be stacked repeatedly, the output of the refinement layer can cyclically update the input of the metric ind. In addition, two gating vectors are added to the RepGen(D) stack to improve the prediction metrics, calculating a variable that reflects the and degree of proximity and participating in the later representation fusion to achieve the effect of enhancing the discriminative ability, in order to better reflect extreme values without sacrificing the overall normal values. The RepGen(D) stack will finally output the preliminary prediction result as well as the variable reflecting the degree of proximity

[0049] The status output by the RepGen(E) stack and the preliminary prediction results output by the RepGen(D) stack will be extended to the RepMerg stack. The preliminary prediction results will be feature-fused with the prediction results output by the corresponding LSTM-FC layer in the RepMerg stack, then overwrite the previous preliminary prediction results, and output the overwritten At the same time, the near points of the corresponding LSTM-FC layer and the ind output (including the status of the far point) will be combined to form a two-dimensional vector, which is input into the next cascaded FC layer to obtain the prediction results The RepMerg stack will finally output and

[0050] The output by the RepMerg stack will be feature-fused with the output by the RepGen(D) stack. After the fused result undergoes inverse normalization operation and logarithmic transformation operation, the final prediction result can be obtained, and then and will be substituted into the loss function to calculate the overall loss. After adjusting the model parameters, it enters the next cycle.

[0051] The prediction results should gradually approach the change trend of the actual values during the training process to improve the accuracy and stability of the prediction. Therefore, it is expected that converges to the first derivative of y. When the prediction results meet the set threshold or exceed the maximum training cycle, the model training is completed.

[0052] Since is relatively important for the prediction effect, two gating vectors are added to the RepGen(D) stack to improve the prediction index, and is used to enhance the discriminant ability, in order to better reflect extreme values without sacrificing the overall normal values. The calculation process of the gating vector is as shown in Figure 3 m far accurately reflects where it is closer to m near is the complement of m far and reflects where it is closer to

[0053] m far 、m near and The calculation formulas are as follows:

[0054]

[0055] mnear = 1 - m far (V)

[0056]

[0057] In the formula, sigmoid(·) is the activation function, and ⊙ is the component multiplication.

[0058] After one cycle, the Distance-Weighted Auto-Regularized Neural Network (DAN) model calculates the loss using multiple distance-weighted loss functions. The purpose is to force the model to learn more information representations and serve as an effective regularizer to prevent the model from overfitting to the base normal values in the long-term runoff sequence prediction task.

[0059] Define w f as the weight that emphasizes the accuracy of points far from the average value of 0 (since the series is normalized to have an average value of 0), and the definition of w n is the opposite. If y is the sequence obtained after preprocessing the training set data, then w f and w n are calculated as follows:

[0060] w f = (tanh(y)) 2 (VII)

[0061] w n = (1 - abs(tanh(y))) 2 (VIII)

[0062] Based on these weights, a multi-regulation loss is established, and the calculation formula is as follows:

[0063]

[0064] The overall loss is

[0065]

[0066] In the formula, epoch is the number of the training cycle, and λ is the multiplier for the operation.

[0067] The specific architecture of the Distance-Weighted Auto-Regularized Neural Network (DAN) model is as Figure 4 shown. For ease of understanding, each neural network is represented by a combination of its first letter and a number. The parameters of each neural network are shown in Tables 1, 2, and 3.

[0068] LSTM generally does not directly output results and needs to be combined with the CNN or FC layer for processing. It should be noted that the input sequence will have three dimensions within the model (training batch; time step; feature dimension). Among them, the feature dimension of the sequence determines the input dimension of the neural network, and the two will be consistent during data transmission.

[0069] The RepGen(E) stack is responsible for encoding the far point, near point, and index sequence. The convolutional kernel sizes of each CNN in "f-E" are set to 11, 7, and 3 in sequence, corresponding to three local patterns of far, medium, and near, which are used to capture the dependencies of long-term, medium-term, and short-term extreme value sequences, and then feature fusion is performed. Such a setting can improve the feature extraction ability and computational efficiency of the encoder, so as to better predict the far point. In the RepGen(E) stack, the difference between each LSTM lies only in the dimension of the input vector. L0, L1, and L2 are connected to the CNN, and their input dimensions are equal to the output dimension of the CNN. L3 and L4 both directly input the sequence, and the input dimension is consistent with the feature dimension of the sequence, both being 1. The dimensions of the vectors output by each LSTM are the same to facilitate later feature fusion.

[0070] The RepGen(D) stack is responsible for decoding the far point, near point, and index sequence. Although L3 and L6 are "connected" on the graph, the input dimension of L6 is different from the output dimension of L3. This is because L3 only outputs the long short-term memory (h2, c2) to L6, but they do not affect the feature dimension of the input sequence. Therefore, the output dimension of L3 has nothing to do with the input dimension of L6. There is also a difference in the input dimension between C3 and the input dimension of L6. This is because the LSTM in this embodiment is bidirectional. Therefore, the input dimension of C3 is equal to twice the output dimension of L6. C4 is directly connected to C3, so the input dimension of C4 is equal to the output dimension of C3. The input dimension and the output dimension of the upper layer seem complex, but in fact, it is determined by whether the data transmitted between neural networks is related to the input sequence.

[0071] The RepMerg stack mainly performs final trimming on the output features and prediction results of the two RepGen stacks, and outputs various sequences used to calculate the final prediction result and loss.

[0072] Table 1 CNN Parameter Table

[0073]

[0074] Table 2 LSTM Parameter Table

[0075]

[0076] Table 3 FC Layer Parameter Table

[0077]

[0078] Embodiment 2: The present invention also provides a method for evaluating the prediction effect of a distance-weighted self-regularized neural network (DAN) model. This method uses the MSE criterion, MAE criterion, RMSE criterion, R 2Criteria are used to evaluate the prediction effect. If x i represents the measured value, represents the mean of the measured values, represents the predicted value, then the calculation formulas for each criterion are as follows:

[0079] (1) MSE (Mean Squared Error)

[0080]

[0081] (2) MAE (Mean Absolute Error)

[0082]

[0083] (3) RMSE (Root Mean Squared Error)

[0084]

[0085] (4) R 2 (Coefficient of Determination)

[0086]

[0087] Example 3: In this example, the daily runoff prediction process for the annual inflow of Linjiacun Reservoir is taken as an example to compare and verify the runoff prediction method described in this application.

[0088] (1) Parameter setting

[0089] The historical runoff processes of each representative year are used as the training set, and the actual runoff process of the current year is used as the verification set, which are input into the constructed DAN model. The prediction step is set to 3, allowing the model to predict and output the runoff for the next 3 days at once. The rainfall lag time is set to 2. The maximum training cycle of the model is set to 20, the training data batch is set to 48, the model learning rate is set to 0.001, the size of the training set is set to 7200 days, and the size of the verification set is set to 120 days.

[0090] Table 4 Model parameter table

[0091]

[0092]

[0093] To provide a reference and enhance the reliability of the research results, the Seasonal Autoregressive Integrated Moving Average with Exogenous Regressors (SARIMAX) model is introduced, and the effectiveness of the DAN model is verified through the comparison between the two.

[0094] The non-seasonal part parameters of the SARIMAX model include the autoregressive (AR) order, the differencing (d) order, and the moving average (MA) order, which are set to (4, 1, 0) respectively in this embodiment. The seasonal part parameters include the seasonal autoregressive (SAR) order, the seasonal differencing (s) order, the seasonal moving average (SMA) order, and the seasonal period, which are set to (1, 1, 0, 12) respectively in this embodiment.

[0095] (2) Runoff forecast results and analysis

[0096] The daily runoff forecasts for 6 representative years are as Figure 5 and Figure 6 shown.

[0097] Calculate the indicators for testing the prediction effect: mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R2). The results are shown in Table 5.

[0098] Table 5 Comparison table of forecast result evaluation indicators

[0099]

[0100] It can be found from the comprehensive comparison of the data in the table that the overall fluctuation of the SARIMAX prediction sequence is relatively large, resulting in an increase in the forecast error. The DAN model not only better judges the possible extreme situations, but also the prediction sequence is more consistent with the measured sequence, with a smaller overall error. The results of each evaluation index are better than those of SARIMAX. In particular, the coefficient of determination is above 0.75, indicating that its prediction effect is better. Therefore, overall, the prediction effect of the DAN model is significantly better than that of the SARIMAX model.

[0101] Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other implementation manners obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

Claims

1. A method for real-time rolling runoff forecasting, characterized in that: Runoff forecasting is performed by constructing a distance-weighted self-regularized neural network model. The training set of the model is the historical data of each representative year, and the test set is the data of the representative year. The training set data is input into the distance-weighted self-regularized neural network model after data preprocessing. The distance-weighted self-regularized neural network model uses multiple groups of LSTM and CNN neural networks to learn the training set data from three directions: points far from the mean, points close to the mean, and indicator data. The FC layer is then used for data fusion, and the loss is calculated through multiple distance-weighted loss functions to finally obtain the forecast result. In the forecasting process, the distance-weighted self-regularized neural network model predicts a sequence of time steps backward based on the current time, and continuously updates the input data, so that the future runoff process can be forecasted on a daily basis, realizing real-time forecasting. The prediction results of the first period of each time series are collected, and the combined results are used as the forecast runoff process of the entire forecast period. The real-time forecasting effect of the distance-weighted self-regularized neural network model is evaluated through the runoff process. The distance-weighted self-regularized neural network (DAN) model includes a RepGen module and a RepMerg module, wherein the RepGen module includes two stacks, RepGen (E) and RepGen (D), for generating polar representations of common sequences and refined indicators, wherein RepGen (E) is the encoder part and RepGen (D) is the decoder part; RepMerg is responsible for feature fusion; the RepGen contains three parallel encoder-decoder blocks, and the three parallel encoder-decoder blocks are divided into f layers, i layers and n layers.

2. The method for real-time rolling runoff forecasting according to claim 1 is characterized in that: The training set data includes historical runoff series and related time series simultaneously with the historical runoff series. Modeled as a normal series, correlated time series to The group model is an indicator sequence. When the data only has a historical runoff sequence, a Gaussian mixture model is used to fit the historical runoff sequence. ,based on Generates a Gaussian mixture model GMM indicator, which is used as the correlation time series , the formula of Gaussian mixture model is as follows (one) (two) In the formula, yes d -dimensional sample vector (column vector); is the mixture weight of the i-th single Gaussian model, and ; is the probability density function of the i-th single Gaussian model; is the mathematical expectation of the i-th single Gaussian model; is the covariance matrix of the i-th single Gaussian model, and M is a hyperparameter tuned for time series; A probability is generated for each point in the time series through the Gaussian mixture model. The probability represents the possibility of the observed data point occurring. At the same time, according to the weights obtained in the Gaussian mixture model fitting process, the weighted sum of all component probabilities is used as the indicator feature of each value in the time series. This feature provides a measure of the contribution of each component to the entire time series at a specific point. Kruskal-Wallis sampling is used to evaluate the normality of training samples and provide an oversampling strategy to sample oversampled areas with extreme events in the training set, thereby improving the prediction ability of the distance-weighted self-regularized neural network model for extreme events. Extract multiple sequences of size t+h A random sample of x , and split the sequence into k contiguous subsequences of equal size, and then calculate k Kruskal-Wallis test statistic between subsequences H ; H The calculation formula is as follows: (three) In the formula, n is the total number of samples in all groups, is the rank sum of the jth group, is the number of samples in the jth group. The time series is preprocessed by logarithmic transformation and standardization. The prediction sequence output by the model is subjected to inverse standardization and logarithmic transformation operations to obtain the final prediction result.

3. The method for real-time rolling runoff forecasting according to claim 1 is characterized in that: The distance-weighted self-regularized neural network model calculates the loss using multiple distance-weighted loss functions after completing one cycle.

4. The method for real-time rolling runoff forecasting according to claim 1 is characterized in that: The distance-weighted self-regularized neural network model also takes into account the time lag effect, allowing users to set the lag of each related sequence on the predicted sequence.

Images