Long-term-ultra-long-term runoff prediction method based on dynamic prediction factors and artificial intelligence
By dynamically selecting predictive factors and using the GRU model, the problem of climate factor dependence and variation in long-term and ultra-long-term runoff forecasting has been solved, improving the prediction accuracy and lead time of reservoir inflow in the upper reaches of the Yangtze River, and providing an effective basis for hydropower scheduling.
Patent Information
- Application Number
- CN202510982462.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-16
- Publication Date
- 2025-11-18
Smart Images
Figure CN120975291A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to the technical field of hydrology, meteorology and artificial intelligence, and particularly relates to a long-term-ultra-long-term runoff prediction method based on dynamic prediction factors and artificial intelligence. BACKGROUND
[0002] Long-term-ultra-long-term runoff prediction aims to predict the runoff in the future of more than half a year to one year. Accurate long-term-ultra-long-term runoff prediction is an important guarantee for water power scheduling and management, water conservancy project operation and maintenance, and flood control. Compared with short-term-medium-long-term runoff prediction, long-term-ultra-long-term runoff prediction lacks reliable precipitation prediction, and traditional physical / statistical prediction methods relying on precipitation prediction have problems such as prediction error accumulation effect, low prediction accuracy, and short effective prediction period, and there is still a large gap with the actual production demand. The upper reaches of the Yangtze River are located in a monsoon climate zone, and the influencing factors are complex and changeable, so long-term-ultra-long-term runoff prediction faces great challenges. Research shows that the introduction of climate teleconnection factors plays a key role in the medium-long-term runoff prediction of the upper reaches of the Yangtze River. Although some scholars have tried to combine climate factors and machine learning models to construct runoff prediction models, the existing research still has the following limitations: (1) The prediction time is limited to the intra-annual scale, and there is a lack of ultra-long-term runoff prediction on the inter-annual scale; (2) The interdependence between teleconnection factors is not fully considered, which may easily lead to overfitting of the model; (3) Traditional methods usually select fixed and unchanging prediction factors based on historical data, but under the background of climate change, the key prediction factors and their importance will dynamically evolve over time, and the prediction effect of the fixed factor set will be significantly reduced with climate change. Therefore, in order to cope with the challenges brought by climate change, it is necessary to explore more potential predictability sources and new prediction methods, fully consider the dynamic changes of prediction factors and their dependence, and significantly improve the accuracy and reliability of long-term-ultra-long-term runoff prediction. SUMMARY
[0003] The application aims to overcome the above-mentioned deficiencies, and provides a long-term-ultra-long-term runoff prediction method based on dynamic prediction factors and artificial intelligence. The method dynamically finds key prediction factors by considering the non-stationarity of the teleconnection effect of large-scale climate factors, combines the advantages of machine learning in nonlinear fitting, constructs a long-term-ultra-long-term runoff prediction model for the upper reaches of the Yangtze River, improves the prediction skill and prediction period of the inflow of key reservoirs in the upper reaches of the Yangtze River, and provides an effective basis for water power scheduling and management.
[0004] To solve the above technical problems, the technical scheme adopted by the application is as follows: a long-term-ultra-long-term runoff prediction method based on dynamic prediction factors and artificial intelligence, which comprises the following steps:
[0005] S1, data collection and preprocessing;
[0006] S2, determining the month to be predicted and the prediction time limit, setting the training set and the test set, and normalizing the data;
[0007] S3, taking the inflow of the key reservoir in the month to be predicted as the prediction object, taking the atmospheric circulation index and the sea temperature index of the previous 24 months as the candidate prediction factors according to the prediction time limit, and screening the optimal prediction factor combination by combining the correlation analysis and the stepwise regression method;
[0008] S4, construction and training of the GRU runoff prediction model of the upper reaches of the Yangtze River;
[0009] S5, real-time prediction of the inflow of the upper reaches of the Yangtze River based on dynamic prediction factors and the GRU prediction model.
[0010] Preferably, the step S1 comprises the following process:
[0011] S1.1, collecting and organizing the daily inflow data of the key reservoirs in the upper reaches of the Yangtze River in the historical period, and processing them into monthly runoff data;
[0012] S1.2, collecting and organizing the monthly atmospheric circulation index and the monthly sea temperature index provided by the National Meteorological Center as candidate prediction factors.
[0013] Preferably, the step S2 comprises the following process:
[0014] S2.1, determining the reporting date according to the month to be predicted and the prediction time limit;
[0015] S2.2, determining the training period and the test period according to the reporting date; wherein the training period data are used for prediction factor screening and prediction model training, and the test period is used for model verification;
[0016] S2.3, normalizing the inflow data and the atmospheric circulation and sea temperature index data of the key reservoirs in the training period and the test period, respectively.
[0017] Preferably, in the step S2.3, the normalization process is as follows:
[0018]
[0019] Wherein, x' is the normalized result of the inflow of a certain month or the atmospheric circulation or the sea temperature index, x is the original variable value, and min and max are the minimum and maximum values of the training period or the test period, respectively.
[0020] Preferably, the step S3 comprises the following process:
[0021] S3.1, calculate the correlation coefficient of each climate factor and the inflow, arrange each climate factor according to the correlation from large to small, and eliminate the climate factors without significant influence;
[0022] S3.2, calculate the correlation coefficient between each climate factor in S3.1 in turn, if a certain climate factor has significant correlation with the previous factor, it is eliminated, otherwise it is retained, and the climate factors with significant influence on the inflow and independent of each other are selected as the candidate prediction factors;
[0023] S3.3, taking the higher correlation as the selection condition, introducing the climate factors obtained in S3.2 into the GRU prediction model one by one, if the model simulation effect becomes worse after introducing a new factor, the factor is eliminated, otherwise it is retained; repeat the above process until all factors in S3.2 are introduced into the regression model, and the optimal prediction factor combination under different prediction time of the prediction object can be obtained.
[0024] Preferably, the step S4 comprises the following process:
[0025] S4.1, input the optimal prediction factor time series corresponding to the training expected prediction object into the GRU prediction model, output the prediction standardized value of the historical inflow of the training expected prediction month, restore it to the original space, and obtain the historical prediction value of the inflow of the month to be predicted in the corresponding year;
[0026] S4.2, input the optimal prediction factor time series corresponding to the test expected prediction object into the GRU prediction model trained in S4.1, output the historical prediction value of the inflow of the month to be predicted in the corresponding year in the test period, compare it with the historical runoff data, and evaluate the current model by using the correlation coefficient and KGE index.
[0027] Preferably, in the step S4.1, the GRU prediction model comprises three layers of input layer, hidden layer and output layer, the input layer is responsible for reading in the optimal prediction factor combination at each time, the hidden layer comprises 1 GRU layer, and the output layer is 1 fully connected layer, which is responsible for outputting the time series of the prediction object.
[0028] Preferably, the step S4.1 further comprises: using mean square error MSE as the loss function of the GRU prediction model, and using Adam to optimize the model parameters; at the same time, selecting Kling-Gupta Efficiency (KGE) index as the model performance evaluation function, and using Bayesian optimization method to automatically optimize the learning rate, number of neural units and other hyperparameters, and finally obtaining the optimal GRU prediction model.
[0029] Preferably, when the KGE index is selected as the model performance evaluation function, the performance of the model is comprehensively evaluated by considering the correlation r between the simulation value and the observed value, the proportion β of the mean value and the proportion γ of the standard deviation, and the specific formula is as follows:
[0030]
[0031] Preferably, the step S5 comprises the following process:
[0032] S5.1, for different prediction time, according to the prediction factor combination selected in step S3, the prediction factor time series corresponding to the target year and month is obtained;
[0033] S5.2, the prediction factor time series is input into the prediction model constructed in step S4, the standardized value of the target month runoff prediction is obtained, and the original runoff space is restored to obtain the predicted inflow.
[0034] The present application has the beneficial effects that: firstly, the present application is aimed at the inflow of different months and different prediction time of reservoirs in the upper reaches of the Yangtze River, and by fusing correlation analysis and stepwise regression method, a combination of prediction factors with physical correlation and mutual independence is selected as a characteristic value to input a machine learning model for training, which improves the prediction skill and effectively avoids overfitting, and also makes up for the problem of insufficient machine learning explanation; in the process of selecting factors by stepwise regression, the GRU model is introduced to evaluate the nonlinear influence of climate factors on inflow, which breaks through the limitation of traditional linear regression, and makes the factors more consistent with the actual hydrological climate process. Secondly, the present application only uses easily accessible climate factor data as prediction factors, which significantly reduces the data redundancy of the prediction model, and the model can automatically update the prediction factor combination, so that the prediction factor adjusts with the change of climate background, effectively solves the performance degradation problem of traditional fixed factor model under the background of climate change, and adapts to the prediction demand under different climate modalities. In addition, the present application uses the GRU model for prediction, and the training process automatically optimizes the hyperparameters through the Bayesian optimization algorithm, which can greatly reduce the iteration times, improve the model convergence speed and training efficiency. Finally, the present application improves the prediction skill and prediction period of the inflow of key reservoirs in the upper reaches of the Yangtze River, and provides an effective basis for water power dispatching and management BRIEF DESCRIPTION OF DRAWINGS
[0035] Figure 1 It is a flowchart of a long-term-ultra-long-term runoff prediction method based on dynamic prediction factors and artificial intelligence;
[0036] Figure 2 It is a prediction result evaluation diagram of the inflow of Wudongde Reservoir in July in the training period and test period with different prediction time (month). DETAILED DESCRIPTION
[0037] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0038] Example 1: As Figure 1 As shown, a long-term and ultra-long-term runoff prediction method based on dynamic predictor factors and artificial intelligence is presented. This method can automatically determine the training and test sets according to the reporting start date, and automatically update the predictor factor library and prediction model to adapt to climate change and thus improve prediction performance. The specific steps are as follows:
[0039] I. Inflow data from key reservoirs in the upper reaches of the Yangtze River were collected and processed, converting daily flow data into monthly average flows. Data on 88 monthly atmospheric circulation indices and 26 monthly sea surface temperature indices provided by the National Meteorological Administration Climate Center were also collected and processed. Indices with data completeness exceeding 70% were selected, and missing values were imputed using linear interpolation to preserve data characteristics to the greatest extent possible. Based on the completeness of historical data samples, 77 atmospheric circulation indices and 26 sea surface temperature indices were ultimately selected.
[0040] Second, determine the month to be predicted and the prediction period, set up the training and test sets, and standardize the data. The specific steps are as follows:
[0041] 2-1) Determine the reporting start date based on the month to be predicted and the forecast lead time. For example, if the inflow of water into Wudongde Reservoir is predicted 7 months in advance in July 2024, the reporting start date is January 2024.
[0042] 2-2) Determine the training and testing periods based on the reporting start date. Training period data is used for predictive factor selection and predictive model training, while testing period data is used for model validation. For example, if reporting begins in January 2024, the periods 1960-2010 and 2011-2023 will be designated as the training and testing periods, respectively. If reporting begins in January 2025, the periods 1961-2011 and 2012-2024 will be designated as the training and testing periods, respectively.
[0043] 2-3) Standardize the inflow data of key reservoirs and the atmospheric circulation and sea surface temperature index data during the training and testing periods, specifically as follows:
[0044]
[0045] Where x' represents the standardized result of monthly inflow, atmospheric circulation, or sea surface temperature index, and x represents the original variable value. min and max are the minimum and maximum values during the training or testing period, respectively.
[0046] Third, the inflow of key reservoirs in the month to be predicted is used as the prediction object. Based on the prediction lead time, the atmospheric circulation index and sea surface temperature index for the 24 months prior to the reporting date are used as candidate predictor factors (a total of 24 × (77 + 26) = 2472 climate factors). The optimal combination of predictor factors is selected by combining correlation analysis and stepwise regression methods. The specific steps are as follows:
[0047] 3-1) Calculate the correlation coefficient between each climate factor and the inflow, arrange each climate factor in descending order of correlation, and remove climate factors with no significant impact.
[0048] 3-2) Calculate the correlation coefficients between the climate factors in 3-1) in turn. If a climate factor is significantly correlated with the previous factors, it is removed; otherwise, it is retained. Select climate factors that have a significant impact on the inflow and are independent of each other as candidate prediction factors.
[0049] 3-3) Using factors with high correlation as the selection criterion, the climate factors obtained in 3-2) are introduced one by one into the GRU prediction model. If the model simulation effect deteriorates after introducing a new factor, the factor is removed; otherwise, it is retained. This process is repeated until all factors in 3-2) are introduced into the regression model, thus obtaining the optimal combination of predictive factors for different prediction lead times. In this embodiment of the invention, correlation and root mean square error (RMSE) are used to evaluate the performance of the stepwise regression model. If the correlation coefficient increases and the RMSE decreases after introducing a new factor, that factor is retained.
[0050] IV. In this embodiment of the invention, the GRU prediction model is selected to construct the inbound flow prediction model. The specific steps are as follows:
[0051] 4-1) Input the time series of the optimal predictor factors corresponding to the expected prediction object into the GRU prediction model, output the standardized predicted value of the historical inflow of the expected prediction month, restore it to the original space, and obtain the historical predicted value of the inflow of the expected prediction month for the corresponding year. In this embodiment of the invention, the GRU prediction model consists of three layers: an input layer, a hidden layer, and an output layer. The input layer is responsible for reading the optimal combination of predictor factors at each time point, the hidden layer contains one GRU layer, and the output layer is a fully connected layer responsible for outputting the time series of the prediction object. In this embodiment of the invention, the mean squared error (MSE) is used as the loss function of the GRU model, and the Adam algorithm is used to optimize the model parameters. At the same time, the Kling-Gupta Efficiency (KGE) index is selected as the model performance evaluation function, and the Bayesian optimization method is used to automatically optimize hyperparameters such as the learning rate and the number of neurons, finally obtaining the optimal GRU prediction model.
[0052] In this embodiment of the invention, KGE is used as the model evaluation index. KGE is a comprehensive evaluation index that comprehensively evaluates the model performance by considering the correlation (r) between simulated and observed values, the proportion of the mean (β), and the proportion of the standard deviation (γ).
[0053]
[0054] 4-2) Input the time series of the optimal predictor (from the training period) corresponding to the target to be predicted into the GRU prediction model trained in 4-1), output the historical predicted value of the inflow of the month to be predicted in the corresponding year of the test period, compare it with the historical runoff data, and evaluate the current model using the correlation coefficient and KGE index.
[0055] V. Real-time prediction is performed based on the aforementioned predictor set and GRU prediction model. The specific steps are as follows:
[0056] 5-1) For different forecast timeframes, based on the combination of forecast factors selected in (III), the forecast factor time series corresponding to the target year and month are obtained;
[0057] 5-2) Input the time series of the above predictors into the prediction model constructed in (IV) to obtain the standardized value of the target month's runoff prediction, restore it to the original runoff space, and obtain the predicted inflow.
[0058] Example 2: Taking the forecast of the inflow into the Wudongde Reservoir in July as an example (reporting from January 2024), the specific process is as follows:
[0059] 1) Collect and organize the inflow data of Wudongde Reservoir from 1960 to July 2023, and standardize it according to formula (1);
[0060] 2) Collect and organize atmospheric circulation and sea surface temperature index data from January to December 1951 to 2023, perform elimination and interpolation processing according to step one above, and standardize each index data according to formula (1);
[0061] 3) The period from 1960 to 2010 was selected as the training period, and the period from 2011 to 2023 as the testing period. Standardized inflow rates for July from 1960 to 2010 were used as the prediction target. Climate factor indices 1-24 months before the annual reporting start date (e.g., January) were used as candidate prediction factors. The correlation coefficients between each climate factor and the inflow rate were calculated. Factors were ranked from highest to lowest correlation, and the correlation coefficients between factors were calculated. Factors significantly correlated with the preceding factor were removed. The GRU prediction model was selected as the regression model to capture the nonlinear dependence between the prediction factors and the prediction target. The retained factors were then introduced into the GRU model in descending order of correlation. KGE was used as the evaluation index to remove factors that worsened the model fit, while retaining factors that improved the fit. The optimal combination of prediction factors was finally obtained, as shown in Table 1.
[0062] Table 1. Predictor factors for runoff from Wudongde Hydropower Station in July 2024 at different prediction timeframes.
[0063]
[0064] 4) GRU machine learning was selected as the prediction model. The optimal predictor time series for July of each year from 1960 to 2010 and the July inflow of the Wudongde Reservoir were input into the GRU prediction model, and the model was trained using the method described in step four. The predictor series for July of each year from 2011 to 2023 were then input into the trained GRU prediction model. The output results were compared with the actual inflow data, and the correlation coefficient and KGE index were calculated to evaluate the model's predictive performance. The model prediction results during the testing and training periods are shown below. Figure 2 As shown, the model can predict the interannual variation of runoff relatively well. The predictive performance does not change significantly with increasing forecast period, and does not show a marked decreasing trend. Except for the 15-16 month forecast period, the correlation coefficients are all above 0.75, and the KGE index is above 0.63.
[0065] 5) Select the predictor data corresponding to July 2024, input its time series into the GRU prediction model constructed above, and the standardized value of the reservoir inflow prediction in July 2024 can be output. Combine it with the climatological data from 1960 to 2023 to restore it to the original runoff space, and the predicted reservoir inflow data for July 2024 can be obtained.
[0066] The method described in this invention can select reliable combinations of predictive factors for different forecast lead times and target months, and train multiple prediction models accordingly. This improves the accuracy of long-term and ultra-long-term runoff forecasts in the upper reaches of the Yangtze River while reducing computational resource consumption.
[0067] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.
Claims
1. A long-term and ultra-long-term runoff prediction method based on dynamic predictor factors and artificial intelligence, characterized in that: It includes the following steps: S1. Data collection and preprocessing; S2. Determine the month to be predicted and the prediction timeframe, set up the training set and test set, and standardize the data. S3. The inflow of key reservoirs in the month to be predicted is taken as the prediction object. Based on the prediction time, the atmospheric circulation index and sea surface temperature index of the 24 months before the reporting date are taken as candidate prediction factors. The optimal combination of prediction factors is selected by combining correlation analysis and stepwise regression methods. S4. Construction and training of the GRU runoff prediction model in the upper reaches of the Yangtze River; S5. Real-time forecasting of inflow into the upper reaches of the Yangtze River based on dynamic predictor factors and GRU prediction models.
2. The long-term and ultra-long-term runoff prediction method based on dynamic predictive factors and artificial intelligence according to claim 1, characterized in that: Step S1 includes the following process: S1.1 Collect and organize daily inflow data of key reservoirs in the upper reaches of the Yangtze River during historical periods, and process them into monthly runoff data; S1.2 Collect and organize the monthly atmospheric circulation index and monthly sea surface temperature index provided by the Climate Center of the National Meteorological Administration as candidate predictive factors.
3. The long-term and ultra-long-term runoff prediction method based on dynamic predictive factors and artificial intelligence according to claim 1, characterized in that: Step S2 includes the following process: S2.1 Determine the reporting start date based on the month to be forecasted and the forecast lead time; S2.2 Determine the training period and testing period based on the reporting start date; The training period data is used for predictor selection and predictor model training, while the testing period data is used for model validation. S2.3 Standardize the inflow data, atmospheric circulation data, and sea surface temperature index data of key reservoirs during the training and testing periods.
4. The long-term and ultra-long-term runoff prediction method based on dynamic predictive factors and artificial intelligence according to claim 3, characterized in that: In step S2.3, the standardization process is as follows: Where x' is the standardized result of the monthly inflow, atmospheric circulation, or sea surface temperature index, x is the original variable value, and min and max are the minimum and maximum values during the training or testing period, respectively.
5. The long-term and ultra-long-term runoff prediction method based on dynamic predictive factors and artificial intelligence according to claim 1, characterized in that: Step S3 includes the following process: S3.1 Calculate the correlation coefficient between each climate factor and the inflow, arrange each climate factor in descending order of correlation, and remove climate factors with no significant impact. S3.2 Calculate the correlation coefficient between each climate factor in S3.1 in sequence. If a climate factor is significantly correlated with the previous factors, it is removed; otherwise, it is retained. Select climate factors that have a significant impact on the inflow and are independent of each other as candidate prediction factors. S3.
3. Select the factors with high correlation as the selection criteria. Introduce the climate factors obtained in S3.2 into the GRU prediction model one by one. If the model simulation effect deteriorates after introducing a new factor, the factor will be removed; otherwise, it will be retained. Repeat the above process until all factors in S3.2 are introduced into the regression model, and the optimal combination of predictive factors for different prediction timeframes of the predicted object can be obtained.
6. The long-term and ultra-long-term runoff prediction method based on dynamic predictor factors and artificial intelligence according to claim 1, characterized in that: Step S4 includes the following process: S4.1 Input the time series of the optimal predictor corresponding to the training expected prediction object into the GRU prediction model, output the predicted standardized value of the historical inbound flow of the training expected prediction month, restore it to the original space, and obtain the historical predicted value of the inbound flow of the month to be predicted in the corresponding year. S4.2 Input the time series of the optimal predictor corresponding to the test target into the GRU prediction model trained in S4.1, output the historical predicted value of the inflow of the month to be predicted in the corresponding year of the test period, compare it with the historical runoff data, and evaluate the current model using the correlation coefficient and KGE index.
7. The long-term and ultra-long-term runoff prediction method based on dynamic predictor factors and artificial intelligence according to claim 6, characterized in that: In step S4.1, the GRU prediction model consists of three layers: an input layer, a hidden layer, and an output layer. The input layer is responsible for reading in the optimal combination of prediction factors at each time step. The hidden layer contains one GRU layer, and the output layer is a fully connected layer responsible for outputting the time series of the prediction object.
8. The long-term and ultra-long-term runoff prediction method based on dynamic predictor factors and artificial intelligence according to claim 7, characterized in that: Step S4.1 further includes: using the mean squared error (MSE) as the loss function of the GRU prediction model, and using Adam to optimize the model parameters; at the same time, selecting the Kling-Gupta Efficiency (KGE) index as the model performance evaluation function, and using Bayesian optimization methods to automatically optimize hyperparameters such as learning rate and number of neurons, and finally obtaining the optimal GRU prediction model.
9. The long-term and ultra-long-term runoff prediction method based on dynamic predictor factors and artificial intelligence according to claim 8, characterized in that: When selecting the KGE index as the model performance evaluation function, the model performance is comprehensively evaluated by considering the correlation r between simulated and observed values, the proportion of the mean β, and the proportion of the standard deviation γ. The specific formula is as follows:
10. The long-term and ultra-long-term runoff prediction method based on dynamic predictor factors and artificial intelligence according to claim 1, characterized in that: Step S5 includes the following process: S5.
1. For different forecast lead times, based on the combination of forecast factors selected in step S3, obtain the forecast factor time series corresponding to the target year and month. S5.2 Input the time series of the above-mentioned predictor factors into the prediction model constructed in step S4 to obtain the standardized value of the target month's runoff prediction, restore it to the original runoff space, and obtain the predicted inflow.