RS-RVM-based medium and long term runoff information prediction method and system

Abstract

The invention relates to the technical field of medium-and-long-term runoff information prediction, and particularly discloses an RS-RVM-based medium-and-long-term runoff information prediction method and system, and the method comprises a rough set (RS) preprocessing stage and a relevance vector machine (RVM) prediction stage, and comprises the steps: firstly, carrying out the preprocessing of input data through employing an RS method, i.e., enabling an RS network to serve as a front-end system in advance, and carrying out the prediction of the RS-RVM-based medium-and-long-term runoff information; and according to the information structure after RS processing, performing information prediction by using the RVM. According to the prediction system and method, a rough set method is utilized to reduce the number of features expressed by information, remove redundant information and simplify a training sample set, so that the data volume of the input end of the relevance vector machine is greatly reduced, the operation speed of the model is improved, the relevance vector machine serves as a rear information processing system, the fault-tolerant and anti-interference capabilities are high, and the prediction accuracy is high. The method has good generalization ability.

Description

Technical Field

[0001] The present invention relates to the technical field, and in particular to a medium- and long-term runoff information prediction method and system based on RS-RVM. Background Art

[0002] Rough set theory (RST) is a new data analysis tool for dealing with fuzzy and uncertain information. RST can effectively analyze and process various incomplete information such as imprecise, inconsistent, and incomplete information, discover implicit knowledge from it, and reveal potential laws.

[0003] The Relevance Vector Machine (RVM) model is a new machine learning method based on the prediction formula of the support vector machine and the sparse Bayesian learning theory based on probability learning. Due to the high sparsity of the algorithm and the learning characteristics based on probability, the relevance vector machine not only obtains a higher prediction accuracy, but also greatly reduces the number of kernel functions involved in the prediction calculation compared with the support vector machine, reduces the prediction calculation time, and can provide probabilistic prediction. It has been initially applied to the fields of stock index prediction, nonlinear system identification, product quality inspection, etc.

[0004] From the principles and characteristics of rough set theory (RS) and relevance vector machine (RVM), it can be found that there is a strong complementarity between the two. Rough set method (RS) is a powerful data analysis tool with the advantages of processing incomplete information, simplifying massive data information and obtaining key knowledge expressions, identifying and evaluating the dependencies between data, and obtaining easily verifiable rule knowledge from empirical data. However, its fault tolerance and generalization capabilities are relatively limited, and it can only process quantitative data. The relevance vector machine (RVM) is based on the structural risk minimization (SRM) criterion. Due to the high sparsity and probability-based learning characteristics of the algorithm, it has good generalization ability. Compared with machine learning methods such as neural networks and support vector machines, it greatly reduces the number of free parameters and prediction calculations, and reduces the prediction calculation time. However, the relevance vector machine has deficiencies in information extraction of input data, and cannot remove the redundancy of information. When the input information space dimension is large, it will cause distortion in the training of the relevance vector machine. Summary of the invention

[0005] In order to overcome the deficiencies of the above-mentioned prior art, the present invention provides a method of correlation vector machines in rough set theory and statistical learning theory, explores a new way to mine key forecast information, and establishes a medium- and long-term hydrological forecast model that can fully absorb information from multiple forecast factors and adaptively eliminate information redundancy. That is, it utilizes RS theory to pre-process the input data, that is, the RS network is used as a front-end system in advance, and then according to the information structure after RS ​​processing, RVM is used to predict the information to obtain a medium- and long-term runoff information prediction system and method based on RS-RVM.

[0006] The technical solution adopted by the present invention is: a medium- and long-term runoff information prediction method based on RS-RVM, including RS preprocessing and RVM prediction stages, specifically including:

[0007] S100. Select a set of early impact prediction factors of runoff, and perform data preprocessing according to the determined prediction factors and object data to form a decision table with conditional attributes and decision attributes;

[0008] S200. Use the rough set reduction algorithm to simplify the attributes of the decision table and establish the minimum condition attribute set;

[0009] S300. Form a training sample set using the minimum conditional attribute set, and use the training sample set as an input of the relevant vector machine model and input it into the RVM prediction system;

[0010] S400. The RVM prediction system performs prediction based on a reserved test sample set and a determined related vector machine model to obtain medium- and long-term runoff information prediction results.

[0011] In the present technical solution, in step S200, key predictors are screened according to the importance of the attributes of each predictor to eliminate redundant information, so as to establish a minimum condition attribute set.

[0012] In the present technical solution, in step S300, the model parameters in the test sample set are input into the training sample set formed by the minimum conditional attribute set, and relevant data simulation training is performed to form the latest training sample set.

[0013] In this technical solution, the rough set reduction algorithm used in step S200 is a numerical attribute reduction based on a neighborhood rough set model, and the reduction steps are:

[0014] S201. Obtaining the decision table matrix and neighborhood radius;

[0015] S202. Granulate the decision table;

[0016] S203. Initialize the reduction matrix;

[0017] S204. Calculate the attribute importance of all remaining attributes;

[0018] S205. Select the attribute with the largest attribute importance value and add it to the reduction matrix;

[0019] S206. If the value of the reduced matrix dependency function does not change after the new attribute is added, proceed to S207; otherwise, proceed to S204;

[0020] S207. The procedure ends.

[0021] In this technical solution, when the basic factor set is reduced according to the rough set reduction algorithm, the input data needs to be normalized before the relevant vector machine is used to build the model:

[0022] (1)

[0023] In formula (1): and are the variables before and after the normalization of the basic factor set; and are the maximum and minimum values ​​of the basic factor set before normalization.

[0024] In the present technical solution, when the rough set theory is used to process the decision table in step S200, the data of each basic factor is granulated using the neighborhood relationship model, and a decision relationship table for the corresponding monthly forecast is formed based on the granulated data.

[0025] A medium- and long-term runoff information prediction system based on RS-RVM, wherein the information prediction system includes an RS preprocessing model and an RVM prediction model;

[0026] Wherein: the RS preprocessing model is used to select the pre-effect prediction factor set of runoff, and perform data preprocessing according to the determined prediction factors and object data to form a decision table with condition attributes and decision attributes, and use the rough set reduction algorithm to simplify the attributes of the decision table to establish the minimum condition attribute set;

[0027] The RVM prediction model is used to form a training sample set according to a minimum conditional attribute set, and use the training sample set as an input of a relevant vector machine model to input into an RVM prediction system; and is used to perform a prediction test on a relevant vector machine model determined according to a reserved test sample set to obtain a medium- and long-term runoff information prediction result.

[0028] In the present technical solution, the RS preprocessing model is also used to screen key predictors according to the importance of the attributes of each predictor to eliminate redundant information, so as to establish a minimum conditional attribute set.

[0029] In the present technical solution, the RVM prediction model is also used to select model parameters in the test sample set and input them into the training sample set formed by the minimum conditional attribute set, and perform relevant data simulation training.

[0030] In the technical solution, the RVM prediction model is also used to utilize the rough set reduction algorithm to reduce the numerical attributes based on the neighborhood rough set model, and to granulate the data of the simplified basic factors using the neighborhood relationship model.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. The rough set method is used to reduce the number of features of information expression, remove redundant information, simplify the training sample set, greatly reduce the amount of data at the input of the relevant vector machine, and improve the calculation speed of the model.

[0033] 2. Using the relevant vector machine as a post-information processing system has strong fault tolerance and anti-interference capabilities and good generalization ability.

[0034] In summary, the medium- and long-term runoff information prediction and method based on RS-RVM of the present invention improves the calculation speed of the medium- and long-term runoff information prediction model, improves the fault tolerance and anti-interference ability of the prediction system, and has good generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is the flow chart of medium- and long-term runoff information prediction based on RS-RVM;

[0036] Figure 2 This is a flowchart of the medium- and long-term runoff information prediction process based on RS-RVM;

[0037] Figure 3 It is a medium- and long-term runoff information prediction system based on RS-RVM;

[0038] Figure 4 The flowchart of simplification using the rough set reduction algorithm;

[0039] Figure 5 This is the structure diagram of the medium- and long-term runoff forecast model based on the correlation vector machine;

[0040] Figure 6 It is a comparison between the simulated and measured values ​​of runoff in September from 1952 to 2000;

[0041] Figure 7 It is a comparison between the simulated and measured values ​​of runoff in October from 1952 to 2000;

[0042] Figure 8 It is a comparison between the predicted and measured values ​​of runoff in September from 2001 to 2008;

[0043] Fig. 9 It is a comparison between the predicted and measured values ​​of runoff in October from 2001 to 2008;

[0044] Among them: 100-RS pretreatment model, 200-RVM prediction model. DETAILED DESCRIPTION

[0045] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0046] like Figure 1 and Figure 2 As shown in FIG. 1 , the medium- and long-term runoff information prediction method based on RS-RVM includes an RS processing stage and an RVM prediction stage, which specifically includes the following steps:

[0047] The RS processing stage mainly completes the following steps: S100. Select the set of early impact prediction factors of runoff, and perform data preprocessing according to the determined prediction factors and object data to form a decision table with conditional attributes and decision attributes;

[0048] S200. Use the rough set reduction algorithm to simplify the attributes of the decision table and establish the minimum condition attribute set;

[0049] The RVM prediction stage mainly completes the following steps: S300. Use the minimum condition attribute set to form a training sample set, and use the training sample set as the input of the relevant vector machine model and input it into the RVM prediction system;

[0050] S400. The RVM prediction system performs prediction based on a reserved test sample set and a determined related vector machine model to obtain medium- and long-term runoff information prediction results.

[0051] like Figure 3As shown, the medium- and long-term runoff information prediction system based on RS-RVM includes an RS preprocessing model 100 and an RVM prediction model 200; wherein: the RS preprocessing model 100 is used to select a set of early-stage influencing prediction factors of runoff, and perform data preprocessing according to the determined prediction factors and object data to form a decision table with conditional attributes and decision attributes, and use the rough set reduction algorithm to simplify the attributes of the decision table to establish a minimum conditional attribute set; the RVM prediction model 200 is used to form a training sample set according to the minimum conditional attribute set, and use the training sample set as the input of the relevance vector machine model and input it into the RVM prediction system; and is used to perform prediction testing on the determined relevance vector machine model according to the reserved test sample set to obtain the medium- and long-term runoff information prediction results.

[0052] like Figure 4 As shown, the simplification algorithm using rough sets in step S200 is a numerical attribute simplification based on a neighborhood rough set model, and the simplification steps are:

[0053] S201. Obtaining the decision table matrix and neighborhood radius;

[0054] S202. Granulate the decision table;

[0055] S203. Initialize the reduction matrix;

[0056] S204. Calculate the attribute importance of all remaining attributes;

[0057] S205. Select the attribute with the largest attribute importance value and add it to the reduction matrix;

[0058] S206. If the value of the reduced matrix dependency function does not change after the new attribute is added, proceed to S207; otherwise, proceed to S204;

[0059] S207. The procedure ends.

[0060] Example

[0061] Taking a reservoir in the upper reaches of the Han River as an example, the present invention adopts a hydro-meteorological coupling approach, combines the medium- and long-term rainfall forecast model of each sub-region and the monthly water balance model, and establishes a medium- and long-term water inflow forecast system with a certain physical cause mechanism. Based on the analysis of the sea-atmosphere background affecting the runoff in the autumn flood season in the basin and the identification of forecast factors, the medium- and long-term forecast of the inflow runoff of a reservoir in the autumn flood season is focused on.

[0062] Based on the sea-atmosphere background that affects the runoff in the autumn flood season, the preliminary factors for the medium- and long-term forecast of the runoff into a reservoir in the autumn flood season were identified. The number of preliminary factors for the runoff forecast in September is 39, and the number of preliminary factors for the runoff forecast in October is 44. These factors have the characteristics of satisfying both mathematical statistics and certain physical causes.

[0063] When rough set theory is used to process decision tables, it is generally required that the values ​​in the decision tables be represented by symbolic data. The prediction factors and runoff data used in the present invention are both numerical, so the relevant data must be preprocessed first. When conventional discretization algorithms such as equal width and equal frequency are used to convert numerical attributes into symbolic attributes, information loss is inevitably caused, and the result of the calculation process depends to a large extent on the effect of discretization. In order to solve this problem, the present invention uses a neighborhood relationship model to granulate the data of each prediction factor.

[0064] Given a non-empty finite set on the real number space , for any object x on U i , its δ neighborhood is: , where δ ≥ 0, δ (x i ) is the object x i The neighborhood particles of δ are in the range of 0.05 to 0.5. In the present invention, δ is 0.125, and 39 factors in September and 44 factors in October are granulated. Based on the granulated data, a decision table for the corresponding monthly forecast is formed.

[0065] 1. Determination of the basic prediction factor set

[0066] On the basis of the determination of the decision table, in order to extract the key forecast information in the forecast factors and avoid high correlation (i.e., multicollinearity) between the forecast factors, the present invention needs to simplify the forecast factors to achieve the purpose of concentrating the forecast information and reducing the impact of multicollinearity noise, thereby improving the forecasting ability of the forecast model.

[0067] If the number of attributes in a decision table is N, it is found that all reductions of the decision system need to be tested When the number of attributes is too large, the amount of calculation is intolerable. Therefore, a forward greedy algorithm based on attribute importance is used here to simplify the granulated decision table.

[0068] The algorithm starts with the empty set, calculates the attribute importance of all remaining attributes each time, and selects the attribute with the largest attribute importance value to add to the reduced set until the importance of all remaining attributes is 0, that is, the value of the system's dependency function no longer changes when any new attribute is added. The forward search algorithm can ensure that important attributes are added to the simplification first, so that important features are not lost. However, the backward search algorithm is difficult to guarantee this result, because for decision systems with a large number of redundant features, the deletion of those important attributes does not necessarily reduce the discrimination ability of the entire system. Therefore, the system may eventually retain a large number of features with weak discrimination ability, but as a whole, they can still maintain the resolution ability of the original data, rather than a small number of features with strong discrimination ability.

[0069] The steps of the numerical attribute reduction algorithm based on the neighborhood rough set model are as follows:

[0070] Input: decision table matrix and neighborhood radius; Output: attribute reduction table and importance;

[0071] Step 1: Granulate the decision table;

[0072] Step 2: Initialize the reduction matrix;

[0073] Step 3: Calculate the attribute importance of all remaining attributes;

[0074] Step 4: Select the attribute with the largest attribute importance value and add it to the reduction matrix;

[0075] Step 5: If the value of the reduced matrix dependency function does not change after the new attribute is added, go to step 6; otherwise, go to step 3;

[0076] Step 6: End of the program.

[0077] According to calculations, the number of simplified forecast factors for the September runoff forecast is 7, and the number of simplified forecast factors for the October runoff forecast is 9. The basic forecast factor sets for September and October are shown in Tables 1 and 2 below.

[0078] Table 1 Basic prediction factor set of runoff in September

[0079]

[0080] Table 2 Basic prediction factor set of runoff in October

[0081]

[0082] From Table 1 and Table 2, it can be seen that the correlation of each simplified forecast factor has passed the bilateral test with a confidence level of 0.05 (the critical value of the correlation coefficient at this time is 0.273), indicating that each factor has a significant correlation with the forecast object. The multiple correlation coefficient of the forecast factor in September is 0.908, and the multiple correlation coefficient of the forecast factor in October is 0.929, indicating that the screened factors have high predictability. Similarly, according to the attribute importance of each forecast factor, the sum of the attribute importance of the forecast factor of September runoff reached 0.995, and the sum of the attribute importance of the forecast factor of October runoff was 0.992, indicating that the simplified basic forecast factor set basically contains all the forecast information and uses the least forecast factors.

[0083] 2. Relevance Vector Machine Modeling

[0084] According to the basic factor set simplified by rough set, the following uses the relevance vector machine method to build a model. The structure of the model is as follows Figure 5 As shown, the basic factor set X is input into the RVM prediction system as input data, and after processing, the output runoff data Y is obtained.

[0085] The input of the RVM model is the basic factor set X simplified by the rough set in the previous section, and the output is the runoff Y in the autumn flood season (September and October). In order to avoid the magnitude differences between the basic factors and eliminate the influence of different dimensions and units of each factor, the input data needs to be normalized first:

[0086] (1)

[0087] In formula (1): x i and i are the variables before and after the basic factor normalization; x max and x min are the maximum and minimum values ​​of the basic factors before normalization.

[0088] In the correlation vector machine, different kernel functions will form different algorithms. Commonly used kernel functions include polynomial, Gaussian radial basis, B-spline kernel function and many other types. Experience shows that the Gaussian radial basis kernel function has good nonlinear processing capabilities, so the present invention uses the Gaussian radial basis kernel function. , the specific formula is shown in formula (2).

[0089] (2);

[0090] Through the cross-validation method, the width of the Gaussian kernel function for the September runoff forecast is determined to be σ=3.5; the width of the Gaussian kernel function for the October runoff forecast is σ=1.8. The data from 1952 to 2000 are used for modeling and simulation, and the data from 2001 to 2008 are reserved for trial forecasting. At the same time, the support vector machine method is used for comparison.

[0091] 3. Analysis of forecast results

[0092] Based on the established correlation vector machine model, the runoff in September and October from 1952 to 2000 was simulated, and the runoff from 2001 to 2008 was reported. At the same time, it was compared with the support vector machine prediction model. Figure 6 and Figure 7 The comparison between the measured and simulated runoff values ​​in September and October from 1952 to 2000 is shown in Figure 2. Figure 8 and Fig. 9 They are the comparison between the reported runoff values ​​and the actual values ​​from September and October 2001 to 2008.

[0093] Depend on Figures 6 to 9 It can be seen that for the medium- and long-term forecast of runoff in September and October during the autumn flood season in a certain place, both the correlation vector machine model and the support vector machine model can achieve good simulation, and the fitting curves of the simulated values ​​and the measured values ​​are relatively consistent.

[0094] In order to evaluate the prediction accuracy and performance of the forecast model, the present invention adopts the following error standards to analyze the prediction accuracy, including correlation coefficient (R), root mean square error (RMSE), and Nash efficiency coefficient (E), and combines the scheme for evaluating the medium- and long-term forecast accuracy in the "Hydrological Information Forecast Specification" (SL250-2000) as the evaluation standard for the forecast model forecast: for quantitative forecast, the water level (flow) is 10% of the multi-year fluctuation, other factors are 20% of the multi-year fluctuation, and the occurrence time of the extreme value of the factor is 30% of the multi-year fluctuation as the allowable error. It is expected to fully reflect the performance of the model in terms of accuracy, efficiency and error response. The formulas of correlation coefficient (R), root mean square error (RMSE), and Nash efficiency coefficient (E) are as follows.

[0095] Correlation coefficient (R):

[0096] (3);

[0097] Root Mean Square Error (RMSE):

[0098] (4);

[0099] Nash efficiency coefficient (E):

[0100] (5);

[0101] In the above formulas (3), (4) and (5), is the length of the time series, is the measured runoff, To predict runoff, and are the means of the measured runoff and the predicted runoff, respectively. The larger the correlation coefficient (R) and the Nash efficiency coefficient (E), the smaller the root mean square error (RMSE), indicating that the prediction effect is better.

[0102] Tables 3 and 4 below show the accuracy evaluation results of the correlation coefficient (R), root mean square error (RMSE), and Nash efficiency coefficient (E). Table 5 shows the accuracy evaluation results according to the Hydrological Information Forecast Specifications.

[0103] Table 3 Results and accuracy analysis of runoff prediction in September during the autumn flood season

[0104]

[0105] Table 4 Results and accuracy analysis of runoff prediction in October during autumn flood season

[0106]

[0107] Table 5 Statistics of runoff simulation and forecast accuracy during the autumn flood season from 1952 to 2000

[0108]

[0109] Judging from the evaluation results of the three forecast accuracy indicators, for the medium- and long-term forecast of runoff in September and October during the autumn flood season in a certain place, the relevance vector machine model is better than the support vector machine model, and the results of the three evaluation indicators are consistent. Moreover, for the runoff in September, the relevance vector machine model only used 5 relevance vectors, while the support vector machine model used 16 support vectors. For the runoff in October, the relevance vector machine model only used 6 relevance vectors, while the support vector machine model used 19 support vectors, indicating that the relevance vector machine has better prediction performance under the same circumstances.

[0110] In addition, from the accuracy evaluation results of the "Hydrological Information Forecast Specifications", the advantages of the correlation vector machine model are also reflected. According to the standard of 10% multi-year variation, the RVM model fitting pass rate of the September runoff historical samples reached more than 83%, and the SVM model fitting pass rate reached about 77%. During the 8-year forecast period, the RVM forecast was qualified for 5 years and the SVM forecast was qualified for 4 years. The RVM model fitting pass rate of the October runoff historical samples reached more than 81%, and the SVM model fitting pass rate reached about 73%. During the 8-year forecast period, the RVM The forecast is qualified for 6 years, and the SVM forecast is qualified for 4 years; according to the standard of 20% multi-year variation, the RVM model fitting pass rate of the September runoff historical samples is above 93.9%, and the SVM model fitting pass rate is about 87.8%. During the 8-year forecast period, the RVM forecast is all qualified, and the SVM forecast is qualified for 6 years. The RVM model fitting pass rate of the October runoff historical samples is above 95.9%, and the SVM model fitting pass rate is about 91.8%. During the 8-year forecast period, the RVM forecast is all qualified, and the SVM forecast is qualified for 7 years.

[0111] In general, the RVM model has good robustness and generalization performance. Both simulation and test accuracy are satisfactory, especially for the characteristics of flood and drought in the forecast year.

[0112] In summary, we can draw the following conclusions:

[0113] (1) The selection of prediction factors is to extract effective information from a large amount of prediction information as the input of the prediction model. Traditional statistical methods have great defects in dealing with such problems. Rough set theory has the advantages of processing incomplete information, simplifying massive data information and obtaining key knowledge expressions. The attribute reduction algorithm of rough set theory can directly obtain the basic prediction factor set with the highest prediction effect, which can provide an effective prediction information source for the establishment of the later model. The example results of a reservoir show that the simplified basic factor set basically contains all the prediction information, and uses the least prediction factors, and the prediction effect is more accurate.

[0114] (2) The medium- and long-term hydrological process is a highly nonlinear process. Traditional forecasting methods only consider it from a linear perspective, which will inevitably lead to forecast deviations. As a new machine learning method emerging from statistical learning theory, the correlation vector machine can effectively deal with the nonlinear relationship between the forecast factor and the forecast object, as well as between the forecast factors. The results of the simulation experiment show that for the sinc(x) function under the same noise conditions, the fitting effect of the correlation vector machine is close to that of the support vector machine, but the number of correlation vectors in the correlation vector machine is significantly smaller than the corresponding number of support vectors, which shows that the correlation vector machine has better robustness and generalization ability.

[0115] (3) Based on the basic factor set simplified by rough set, the runoff of a reservoir from September to October 1952 to 2000 was simulated by using the relevance vector machine method, and the runoff from 2001 to 2008 was tested. At the same time, it was compared with the support vector machine prediction model. The results of four evaluation criteria showed that for the medium- and long-term forecast of runoff in September and October during the autumn flood season in a certain place, the relevance vector machine model had better forecasting performance than the support vector machine model under the same circumstances.

[0116] The embodiments of the present invention disclose preferred embodiments, but are not limited thereto. A person skilled in the art can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. However, as long as they do not deviate from the spirit of the present invention, they are all within the protection scope of the present invention.

Claims

1. The medium- and long-term runoff information prediction method based on RS-RVM is characterized by: The prediction method includes RS preprocessing and RVM prediction stages, specifically including: S100. Select a set of early impact prediction factors of runoff, and perform data preprocessing according to the determined prediction factors and object data to form a decision table with conditional attributes and decision attributes; S200. Use the rough set reduction algorithm to simplify the attributes of the decision table and establish the minimum condition attribute set; S300. Form a training sample set using the minimum conditional attribute set, and use the training sample set as an input of the relevant vector machine model and input it into the RVM prediction system; S400. The RVM prediction system performs prediction based on a reserved test sample set and a determined related vector machine model to obtain medium- and long-term runoff information prediction results.

2. The medium- and long-term runoff information prediction method based on RS-RVM according to claim 1 is characterized in that: In step S200, key predictors are screened according to the importance of the attributes of each predictor to eliminate redundant information, so as to establish a minimum condition attribute set.

3. The medium- and long-term runoff information prediction method based on RS-RVM according to claim 2 is characterized in that: In step S300, the model parameters in the test sample set are input into the training sample set formed by the minimum condition attribute set, and relevant data simulation training is performed to form the latest training sample set.

4. The method for predicting medium- and long-term runoff information based on RS-RVM according to any one of claims 1 to 3, characterized in that: The reduction algorithm using rough sets in step S200 is a numerical attribute reduction based on a neighborhood rough set model. The reduction steps are: S201. Obtaining the decision table matrix and neighborhood radius; S202. Granulate the decision table; S203. Initialize the reduction matrix; S204. Calculate the attribute importance of all remaining attributes; S205. Select the attribute with the largest attribute importance value and add it to the reduction matrix; S206. If the value of the reduced matrix dependency function does not change after the new attribute is added, proceed to S207; otherwise, proceed to S204; S207. The procedure ends.

5. The method for predicting medium- and long-term runoff information based on RS-RVM according to claim 4 is characterized in that: When reducing the basic factor set according to the rough set reduction algorithm, the input data needs to be normalized before using the relevant vector machine to build the model: Where: x i and i are the variables before and after the normalization of the basic factor set; x max and x min are the maximum and minimum values ​​of the basic factor set before normalization.

6. The method for predicting medium- and long-term runoff information based on RS-RVM according to claim 5 is characterized in that: When the rough set theory is used to process the decision table in step S200, the data of each basic factor is granulated using the neighborhood relationship model, and a decision relationship table for the corresponding monthly forecast is formed based on the granulated data.

7. The medium- and long-term runoff information prediction system based on RS-RVM is characterized by: The information prediction system includes an RS preprocessing model and an RVM prediction model; Wherein: the RS preprocessing model is used to select the pre-effect prediction factor set of runoff, and perform data preprocessing according to the determined prediction factors and object data to form a decision table with condition attributes and decision attributes, and use the rough set reduction algorithm to simplify the attributes of the decision table to establish the minimum condition attribute set; The RVM prediction model is used to form a training sample set based on a minimum conditional attribute set, and use the training sample set as an input of a relevant vector machine model to input into an RVM prediction system; and is used to perform a prediction test on the determined relevant vector machine model based on a reserved test sample set to obtain medium- and long-term runoff information prediction results.

8. The information prediction system according to claim 7, characterized in that: The RS preprocessing model is also used to screen key predictors according to the importance of predictor attributes and eliminate redundant information to establish a minimum conditional attribute set.

9. The information prediction system according to claim 7, characterized in that: The RVM prediction model is also used to select model parameters in the test sample set and input them into the training sample set formed by the minimum condition attribute set, and perform relevant data simulation training.

10. The information prediction system according to any one of claims 7 to 9, characterized in that: The RVM prediction model is also used to simplify the numerical attributes based on the neighborhood rough set model by using the rough set reduction algorithm, and to granulate the data of the simplified basic factors by using the neighborhood relationship model.

Images