Real-time runoff rolling forecasting method

Abstract

The invention provides a runoff real-time rolling forecasting method, which comprises the following steps of: performing runoff forecasting by constructing a distance weighted self-regularization neural network model, and inputting training set data into the distance weighted self-regularization neural network model after data preprocessing; according to the distance weighted self-regularization neural network model, multiple groups of LSTM and CNN neural networks are used for learning training set data from three directions of points far away from a mean value, points close to the mean value and index data, then an FC layer is used for data fusion, loss is calculated through multiple distance weighted loss functions, and finally a forecast result is obtained. In the forecasting process, a sequence of a time step length is forecasted backwards based on the current time, and the future runoff process can be forecasted day by day in a rolling manner by setting the forecasted time step length and continuously updating input data, so that real-time forecasting is realized; and collecting a first time period prediction result of each time sequence, and taking a merged result as a runoff prediction process of the whole prediction period, thereby realizing a real-time prediction effect.

Description

Technical Field

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

[0002] With the continuous growth of the global population and the acceleration of industrialization and urbanization, people's demand for freshwater resources has increased dramatically. In this context, how to scientifically predict and manage runoff resources to ensure the stability of life and production and food security has become a key issue that needs to be solved urgently. As the main source of runoff, rainfall has brought great challenges to runoff forecasting due to its uncertainty in temporal and spatial distribution and the complexity of being affected by climate change. In addition, the unique microclimate characteristics of different regions, the complex and changeable underlying surface conditions and the increasing interference of human activities make the runoff formation process full of variables. Incorporating relevant factors such as upstream rainfall into the forecasting system and analyzing and predicting them as important indicator sequences will help improve the accuracy and reliability of runoff forecasting. Such forecast results can not only provide timely and accurate water resources information for irrigation agriculture, support farmers to scientifically formulate irrigation plans, and reduce water waste, but also provide a scientific basis for government departments to formulate water resources management policies and optimize water resources allocation, and can also provide early warning risks in response to natural disasters such as droughts and floods, guide emergency responses, and protect people's lives and property safety and social stability.

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

[0004] At present, many common time series forecasting models perform poorly on data with high skewness and high kurtosis scores, and traditional methods often have deviations when faced with extreme events. The non-stationarity of runoff series and the existence of extreme events add a lot of difficulties to the forecast of runoff. The forecast model should be able to capture long-term dependencies and model rare but important extreme values, but existing methods often find it difficult to cope with such dual challenges at the same time. Summary of the invention

[0005] In view of the problem 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 the technical problem is: a real-time rolling forecast method for runoff, which forecasts runoff through a distance-weighted self-regularized neural network model, wherein the training set of the model is historical data of each representative year, and the test set is data of the representative year, wherein the training set data is input into the distance-weighted self-regularized neural network model after data preprocessing, and 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, and then uses the FC layer to fuse the data, and calculates the loss through multiple distance-weighted loss functions, and finally obtains the forecast result;

[0007] 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 to achieve real-time forecasting.

[0008] 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.

[0009] Furthermore, the training set data includes historical runoff sequences and related time series simultaneously with the historical runoff sequences, wherein the historical runoff sequence x1 is modeled as a common sequence, and the related time series x2 to x m The group model is an indicator sequence. The Gaussian mixture model is used to fit the historical runoff sequence x1. The Gaussian mixture model GMM indicator is generated based on x1. This indicator is used as the related time series x2. The formula of the Gaussian mixture model is as follows

[0010]

[0011] Where x is a d-dimensional sample vector (column vector); β i is the mixture 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 time series.

[0012] The Gaussian mixture model generates a probability for each point in the time series, which represents the possibility 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 indicator feature of each value in the time series, which provides a measure of the contribution of each component to the entire time series at a specific point.

[0013] 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.

[0014] Draw multiple random samples x of size t+h from the input sequence, split the sequence into k consecutive subsequences of equal size, and 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 rank sum of the jth group, n j 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.

[0017] Furthermore, 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 includes three parallel encoder-decoder blocks, and the three parallel encoder-decoder blocks are divided into f layers, i layers, and n layers.

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

[0019] Beneficial effects of the present invention: The real-time rolling forecast method for runoff described in the present invention introduces a distance-weighted self-regularizing neural network (DAN), which is a new extreme adaptive model that enhances the prediction effect of runoff by adding polarity representation learning. The distance-weighted self-regularizing neural network utilizes a distance-weighted multiple loss mechanism and stackable blocks to dynamically refine the indicator sequence from exogenous data. At the same time, it can also process 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-regularizing neural network to process 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-regularizing neural network (DAN) model can better predict extreme events with low frequency but large impact in time series, and has excellent performance. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

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

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

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

[0025] Figure 6 This is a comparison chart of the prediction results of DAN and SARIMAX in the flood year. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be described in detail below.

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

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

[0029] The data preprocessing includes sequence processing, extreme event sampling, logarithmic transformation and standardization. The input information of the model includes historical runoff series and related series (such as upstream rainfall series) as special indicators. First, the historical runoff series x1 is modeled as a common series, and the related time series x2 to x m The task is defined by modeling the group as an indicator sequence. In order to better consider the timeliness of the impact of related sequences, the model allows users to set the lag t of each sequence and change the sequence input order internally to achieve better prediction results.

[0030] When no additional indicator series is provided, a Gaussian mixture model (GMM) is used to fit the historical runoff series x1, and a GMM indicator is generated based on x1, which is used as the related time series x2.

[0031] In this case, the problem reduces to a univariate time series forecasting problem.

[0032] The formula for the Gaussian mixture model is as follows

[0033]

[0034]

[0035] Where x is a d-dimensional sample vector (column vector); β i is the mixture 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 time series, which is set to 3.

[0036] In this embodiment, the expectation maximization (EM) algorithm is used to estimate the parameters of the GMM by iteratively maximizing the observed data. The GMM generates a probability for each point in the time series, indicating the possibility of the observed data point. 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. 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, Kruskal-Wallis sampling is used to evaluate the normality of the training samples and provide an oversampling strategy to sample oversampled areas with extreme events in the training set to improve the model's prediction ability for extreme events.

[0038] The model extracts multiple random samples x of size t+h from the input sequence, splits the sequence into k consecutive subsequences of equal size, and then calculates the Kruskal-Wallis test statistic H between the k subsequences.

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

[0040]

[0041] In the formula, n is the total number of samples in all groups, R j is the rank sum of the jth group, n j is the number of samples in the jth group.

[0042] In order to prevent the H statistic from being affected by small differences in the subsequence, it is necessary to round the value in x 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 of p<1. The threshold λ affects the relative change of the sample, making it more likely that the sample will 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 be preprocessed by logarithmic transformation and standardization. The prediction sequence output by the model can obtain the final prediction result through inverse standardization and logarithmic transformation operations.

[0044] like Figure 1As shown in the figure, the scalable end-to-end framework of the distance-weighted self-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 by the model, which can capture the intrinsic structure and relationship of the data. In the distance-weighted self-regularized neural network (DAN) model, RepGen consists of two stacks, RepGen (E) and RepMerg (D), which are responsible for generating polar representations of common sequences and refined indicators, where RepGen (E) is the encoder part and RepGen (D) is the decoder part; while RepMerg is responsible for feature fusion. The distance-weighted self-regularized neural network (DAN) model can iteratively update the indicators by repeatedly stacking RepGen (E) + RepGen (D), and can also repeatedly stack RepMerg to refine the indicators multiple times. According to the intrinsic relationship of the multivariate sequence, the optimal stack configuration for different data sets may be different. This embodiment adopts the stack configuration of RepGen(E)+RepGen(D)+RepMerg.

[0045] The key innovation of the distance-weighted self-regularized neural network (DAN) model is the use of a new mechanism for generating and exchanging information between near and far polar representations and indicators (ind), which can directly improve the prediction results. Since extreme values ​​usually follow different distributions, polar coordinate representation learning in RepGen encodes them separately, and then retains these points in RepMerg and obtains the final prediction results after feature fusion with the preliminary prediction results, ensuring that these representations are not affected by the dominance of normal values ​​during training.

[0046] RepGen contains three parallel encoder-decoder blocks, such as Figure 2 As shown in , they can transform the metrics refined after the normal series input into polar representations, which are further merged in the RepMerg stack.

[0047] In the RepGen(E) stack, the “fE” layer is responsible for learning representations for points far from the sequence mean. Including extreme values ​​that are 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 long-term, medium-term and short-term extreme value sequence dependencies respectively, and then fuses the captured relationships to improve the encoder's feature extraction ability and computational efficiency, and better predict distant points. Moreover, since the hidden state of extreme events may be updated multiple times in repeated blocks, combining convolution operations can shorten the input sequence and alleviate potential gradient explosion or vanishing problems. The "nE" layer uses LSTM to learn the hidden features of nearby points This includes most of the values ​​that conform to the normal distribution. The "iE" layer uses LSTM to learn the representation of the indicator 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: "fD", "nD", and "iD", which correspond to the RepGen(D) stack one by one, and each receives the unit state c and hidden state h output by the corresponding layer of RepGen(E). "fD" and "iD" are further optimized by LSTM combined with FC layer to obtain far-point Nearby It should be noted that Helps predict extreme events, accurately The prediction results are crucial to improving performance. Therefore, the model’s “iD” layer adds a refinement layer consisting of two one-dimensional convolutional neural network components (CNN1d), which first performs a localized The value is optimized to extract local time series features and improve the modeling ability of complex time series patterns to assist in generating expected indicators. Moreover, since RepGen is designed as an extensible stack that can be repeatedly stacked, the output of the refinement layer can cyclically update the input of the indicator ind. In addition, the RepGen (D) stack also adds two gating vectors to improve the prediction indicators and calculate the reflection and Proximity Variables And participate in the later representation fusion to achieve enhanced The effect of the discriminative ability of the RepGen(D) stack is to better reflect the extreme values ​​without sacrificing the overall normal value. The RepGen(D) stack will eventually output the preliminary prediction results. and variables reflecting proximity

[0049] The status output from the RepGen(E) stack and the preliminary predictions output from the RepGen(D) stack are expanded into the RepMerg stack. The feature fusion is performed with the prediction results output by the corresponding LSTM-FC layer in the RepMerg stack, and then the previous preliminary prediction results are overwritten and the overwritten results are output. At the same time, the near point and ind output (including the state of the far point) of the corresponding LSTM-FC layer are combined to form a two-dimensional vector, which is input to the next series FC layer to obtain the prediction result. The RepMerg stack will eventually output and

[0050] Output the RepMerg stack Output from the RepGen(D) stack The feature fusion is performed, and the fused result is subjected to the inverse normalization operation and logarithmic transformation operation to obtain the final prediction result. and It will be substituted into the loss function to calculate the overall loss, and the model parameters will be adjusted before entering the next cycle.

[0051] The prediction results should gradually approach the changing trend of the actual value during the training process to improve the accuracy and stability of the prediction. Converging to the first-order derivative of y, when the prediction result meets the set threshold or exceeds the maximum training cycle, the model training is completed.

[0052] because For prediction effects, two gating vectors are added to the RepGen(D) stack to improve the prediction indicators and use To enhance The gate vector calculation process is as follows: Figure 3 As shown. far accurately reflects Where is closer m near is m far The complement of Where is closer

[0053] m far 、m near and The calculation formula is as follows:

[0054]

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

[0056]

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

[0058] The distance-weighted self-regularized neural network (DAN) model uses multiple distance-weighted loss functions to calculate the loss after completing a cycle. The purpose is to force the model to learn more information representation and act as an effective regularizer to prevent the model from overfitting to the base normal value in the long-term runoff series prediction task.

[0059] Define w f To emphasize the accuracy of points far from the mean 0 (because the series is normalized to have a mean of 0), w n The definition of is the opposite. y is the sequence obtained after preprocessing the training set data. Then w f and w n The calculation formula is as follows:

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

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

[0062] Based on these weights, a multi-conditioning loss is established, which is calculated as follows:

[0063]

[0064] The overall loss is

[0065]

[0066] Where epoch is the number of the training cycle and λ is the multiplier of the operation.

[0067] The specific architecture of the distance-weighted self-regularized neural network (DAN) model is as follows: Figure 4 For ease of understanding, each neural network is represented by a combination of initials and numbers. The parameters of each neural network are shown in Tables 1, 2, and 3.

[0068] LSTM generally does not output results directly, and needs to be processed in conjunction with CNN or FC layers. It should be noted that the input sequence in the model will contain three dimensions (training batch; time step; feature dimension), among which the feature dimension of the sequence determines the input dimension of the neural network, and the two will remain consistent when data is transmitted.

[0069] The RepGen(E) stack is responsible for encoding the far point, near point and indicator sequence. The convolution kernel size of each CNN in "fE" is set to 11, 7, and 3, corresponding to the three local modes 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 perform feature fusion. This 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 is only the dimension of the input vector. L0, L1, and L2 are connected to CNN, and their input dimensions are equal to the CNN output dimensions. L3 and L4 are both directly input sequences, and the input dimensions are consistent with the feature dimensions of the sequence, both of which are 1. The dimensions of the vectors output by each LSTM are the same to facilitate feature fusion later.

[0070] The RepGen (D) stack is responsible for decoding the far point, near point and indicator sequence. Although L3 is "connected" to L6 in the figure, the input dimension of L6 is different from the output dimension of L3. This is because L3 only outputs long short-term memory (h2, c2) to L6, but they do not affect the feature dimension of the input sequence, so the output dimension of L3 is independent of the input dimension of L6. There is also a difference between the input dimension of C3 and the input dimension of L6. This is because the LSTM in this embodiment is bidirectional, so 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 upper-layer output dimension seem complicated, but in fact they are determined by whether the data transmitted between neural networks is related to the input sequence.

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

[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, which adopts the MSE criterion, MAE criterion, RMSE criterion, R 2Criteria to evaluate the forecast effect, if x i Indicates the measured value, represents the mean of the measured values, Represents the predicted value, then the calculation formulas for each criterion are:

[0079] (1) MSE (mean square error)

[0080]

[0081] (2) MAE (mean absolute error)

[0082]

[0083] (3) RMSE (root mean square error)

[0084]

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

[0086]

[0087] Example 3: This example takes the daily forecast process of the inflow runoff of the Linjia Village Reservoir as an example to compare and verify the runoff forecast method described in this application.

[0088] (1) Parameter setting

[0089] The historical runoff process of each representative year was used as the training set, and the actual runoff process of the year was used as the validation set, which were input into the constructed DAN model. The prediction step was set to 3, so that the model could predict and output the runoff for the next 3 days at a time, and the rainfall lag was set to 2. The maximum training cycle of the model was set to 20, the training data batch was set to 48, the model learning rate was set to 0.001, the training set size was set to 7200 days, and the validation set size was set to 120 days.

[0090] Table 4 Model parameters

[0091]

[0092]

[0093] In order to provide a reference and enhance the reliability of the research results, the seasonal autoregressive differential moving average model (SARIMAX) model was introduced. By comparing the two, the effectiveness of the DAN model was verified.

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

[0095] (2) Runoff forecast results and analysis

[0096] The daily runoff forecast results of 6 representative years are as follows: Figure 5 and Figure 6 shown.

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

[0098] Table 5 Comparison of forecast result evaluation indicators

[0099]

[0100] From the comprehensive comparison of the data in the table, it can be found that the overall fluctuation of the SARIMAX prediction sequence is large, which leads to an increase in the forecast error, while 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, and the results of various evaluation indicators are better than SARIMAX, especially the determination coefficient is above 0.75, indicating that its prediction effect is better. Therefore, on the whole, the prediction effect of the DAN model is significantly better than that of the SARIMAX model.

[0101] Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other implementation methods obtained by ordinary technicians in this field without creative work are 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.

2. The method for real-time rolling runoff forecasting according to claim 1 is characterized in that: The training set data includes historical runoff sequences and related time series simultaneously with the historical runoff sequences, wherein the historical runoff sequence x1 is modeled as a common sequence, and the related time series x2 to x m The group model is an indicator sequence. When the data only has the historical runoff sequence, the Gaussian mixture model is used to fit the historical runoff sequence x1. Based on x1, the Gaussian mixture model GMM indicator is generated. This indicator is used as the related time series x2. The formula of the Gaussian mixture model is as follows Where x is a d-dimensional sample vector (column vector); β i is the mixture 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 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. Draw 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: In the formula, n is the total number of samples in all groups, R j is the rank sum of the jth group, n j 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 (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 polar representations of common sequences and refined indicators, where RepGen(E) is the encoder part and RepGen(D) is the decoder part; RepMerg is responsible for feature fusion.

4. The method for real-time rolling runoff forecasting according to claim 3 is characterized in that: The RepGen includes three parallel encoder-decoder blocks, which are divided into f-layer, i-layer and n-layer.

5. 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.

6. The method for real-time rolling runoff forecast according to claim 1, 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