Runoff forecasting method based on EEMD-ann and meteorological factors and using multi-core parallel algorithm

Abstract

Disclosed in the present invention is a runoff forecasting method based on an EEMD-ANN and meteorological factors and using a multi-core parallel algorithm. The method comprises: S1, performing feature screening, involving: on the basis of historical data of previous inflows and meteorology, performing screening of lag terms of the previous inflows by means of a partial autocorrelation function, using an inverse distance weighting method to reduce the weight of meteorological data, using a cross-correlation function and an index of agreement to perform screening of lag terms of rainfall, then performing feature selection by means of a correlation coefficient γ, and finally, using a correlation coefficient τ to perform screening of input factors; S2, performing data pre-processing, involving: using an empirical mode decomposition technique to decompose the input factors, performing data normalization processing, and dividing data into a training set and a test set; and S3, performing model forecasting and assessment, involving: integrating the input factors and inputting same into a forecast model, then performing parameter optimization by means of a parallel algorithm, outputting a forecast result, and performing model assessment. The present invention effectively improves forecast accuracy, accelerates the efficiency of model parameter calibration, and reduces model training time.

Description

A runoff forecasting method based on EEMD-ANN and meteorological factors using a multi-core parallel algorithm Technical Field

[0001] The present invention belongs to the field of long-term runoff forecasting and relates to a runoff forecasting method based on EEMD-ANN and meteorological factors using a multi-core parallel algorithm. Technical Background

[0002] Due to the impacts of climate change, typhoons, and human activities, inflows have increased dramatically, and the accuracy of long-term runoff forecasts has decreased. This has directly led to difficulties in hydropower station scheduling, a significant increase in water abandonment, and a significant waste of clean energy. Therefore, hydropower stations require an effective long-term runoff forecasting system to accurately predict the large amounts of runoff generated by extreme weather events, helping them plan operations in advance and reduce water and hydropower resource losses. Due to the exponential growth in the complexity and randomness of runoff, the types and amount of hydrological data required for runoff forecasting have also increased accordingly. Faced with such a massive amount of information, traditional runoff forecasting methods, including mathematical statistics and causal analysis, can suffer from limitations such as low information utilization, reduced learning ability, and large prediction errors, making them difficult to meet the needs of practical hydrological data analysis.

[0003] Summary of the Invention

[0004] The purpose of the present invention is to overcome the above-mentioned shortcomings and improve a runoff forecasting method based on EEMD-ANN and meteorological factors using a multi-core parallel algorithm. On the basis of using the EEMD-ANN runoff model, a screening mechanism for input meteorological factors is added, and a multi-core parallel algorithm is used at the same time. It is not only suitable for long-term runoff forecast modeling, but also effectively improves the forecast accuracy, speeds up the efficiency of model parameter calibration, and saves model training time.

[0005] To achieve the purpose of the present invention, the technical solution adopted by the present invention is: a runoff forecasting method based on EEMD-ANN and meteorological factors using a multi-core parallel algorithm, which includes the following steps:

[0006] S1: Feature screening: Based on the historical data of previous inflow and meteorology, the partial autocorrelation function is used to screen the previous inflow of the lag term, the inverse distance weighting method is used to reduce the weight of the meteorological data, the cross-correlation function and consistency index are used to screen the rainfall lag term, and then the correlation coefficient γ is used for feature selection. Finally, the correlation coefficient τ is used to screen the input factors;

[0007] S2: Data preprocessing: Use empirical mode decomposition technology to decompose the input factors, perform data normalization, and divide the data into training and test groups;

[0008] S3: Model prediction and evaluation: Integrate the input factors and input them into the prediction model. Then optimize the parameters through parallel algorithms, output the prediction results, perform parallel evaluation over time, and evaluate the model through correlation coefficient, root mean square error, mean absolute error, mean relative error and certainty coefficient.

[0009] Preferably, in step S1, the input data of the reanalysis meteorological dataset is processed and filtered using the inverse distance weighted method and the single correlation coefficient method;

[0010] The correlation between lagged runoff and total precipitation is demonstrated using partial autocorrelation and cross-correlation functions, and the optimal lag order can be determined using a 95% confidence interval. Furthermore, when the correlation coefficient changes slowly and does not fall within the 95% confidence interval, the optimal lag order is determined using trial and error and the consistency index (IA). The consistency index is an evaluation metric that adds one or more lag sequences to the input of the first lag order.

[0011] Among them, Y is the meteorological data of the research site; Z i is the meteorological value of each measurement point q; d i is the distance from each measurement point q to the research point; n is the number of measurement points; P is the weight index;

[0012] where x j and y j are the values ​​of runoff and meteorological factors in sequence j, respectively; x and y are the multi-year averages of runoff and meteorological factors, respectively; γ is the correlation coefficient between runoff and meteorological factors, τ is the partial correlation coefficient between runoff and meteorological factors; n is the number of samples;

[0013] in and Q i Represent the predicted flow and measured flow at time i respectively; represents the average value of the measured sequence; n is the length of the runoff series.

[0014] Preferably, in step S2, the EEMD-ANN prediction model uses EEMD to decompose the initial signal into several components, then uses ANN to predict each component, and finally adds the predicted values ​​of each component to obtain the final prediction result, thereby improving the prediction accuracy by eliminating "impurities" in the original signal;

[0015] In order to integrate the input data and speed up the learning efficiency of the model, the input signal is normalized and converted to the interval [0,1]. The deviation normalization method is used to normalize the input data, as shown in the following formula:

[0016] Among them, x scale is the reduced value; x i is the input value in the original sequence; x min and x max are the minimum and maximum values ​​in the original data, respectively.

[0017] Preferably, in step S3, after the input factors are integrated, they are input into the ANN model to obtain the output result. The ANN prediction model propagates the error E total And the learning rate η is used to adjust the weight of the model, gradually reducing the error and finally achieving the best prediction model;

[0018] The EEMD algorithm is used to decompose the original sequence, and the decomposition process is as follows:

[0019] The propagation error E of the ANN model total yes:

[0020] By E total and η to adjust the weights:

[0021] Among them E total is the error in the propagation process; D u is the actual value; u is the predicted value; x i and y j are the input value of the model and the output value of the hidden layer, f(x) and g(x) are the activation functions of the hidden layer and the output layer respectively; w ju and w ij is the weight in each propagation, through E total and η update.

[0022] Preferably, in step S3, while predicting by the EEMD-ANN model, a multi-core parallel algorithm is used to reduce the optimization time of the model and improve the prediction efficiency and accuracy of the model; the steps of the parallel algorithm are as follows:

[0023] (1) Parallel reading and analysis of main tasks;

[0024] (2) Decompose the main task into n subtasks according to actual needs and hardware environment;

[0025] (3) The processor is used to process and solve each subtask and obtain the output result of each subtask;

[0026] (4) Organize and merge the output results of each subtask to obtain the result of the main task.

[0027] Preferably, in step S3, after obtaining the forecast results, the accuracy of the model results is evaluated by correlation coefficient, root mean square error, mean absolute error, mean relative error and coefficient of certainty;

[0028] in Represents the mean value of the forecast series.

[0029] Beneficial effects of the present invention:

[0030] The present invention adopts a combination of EEMD-ANN and parallel algorithms, and takes into account the influence of specific meteorological factors. Compared with other forecasting models, EEMD-ANN can comprehensively decompose and analyze the characteristics of input runoff and has better forecasting performance. At the same time, adding specific meteorological factors to the forecasting system improves the learning ability and forecasting accuracy of the model. In addition, it is combined with parallel algorithms to accelerate the iterative process of the model, speed up the efficiency of model parameter calibration, save model training time, improve the efficiency of model construction, reduce time consumption in the model optimization process, improve the accuracy of the forecasting system, and ensure the high efficiency of the forecasting system. For long-term runoff forecasting, the construction of the forecasting model, forecast accuracy and model parameter verification efficiency are all very important. The long-term runoff forecasting framework proposed by the present invention effectively solves these problems and has certain reliability and practicality for the operation of the hydrological system. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 shows the statistics of abandoned water information (Hg: hydropower generation, An: abandoned water power, SC: Sichuan Province, YN: Yunnan Province);

[0032] Figure 2 is a parallel algorithm flow chart;

[0033] Figure 3 (a) shows the PACF plot of (i) historical runoff and (ii) CCF plot of historical rainfall for a power station; (b) IA value of lag order; (c) single factor correlation coefficient plot of meteorological factor screening: (i) γ, (ii) τ;

[0034] Figure 4 (a) shows the decomposition results of the input factors (i) runoff; (ii) tcwv; (iii) tcc; (iv) tp; (v) d2m; (vi) tcw; (b) the parallel time comparison of the prediction models; (c) the prediction results of ANN and EEMD-ANN: (i) without adding refined meteorological factors; (ii) with adding refined meteorological factors;

[0035] FIG5 is a framework diagram of a forecast model of the present invention;

[0036] FIG6 is an enlarged schematic diagram of region i in FIG4c;

[0037] FIG7 is an enlarged schematic diagram of area ii in FIG4c. DETAILED DESCRIPTION

[0038] The present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

[0039] Example 1: As shown in FIG5 , a runoff forecasting method based on EEMD-ANN and meteorological factors using a multi-core parallel algorithm includes the following steps:

[0040] S1: Feature screening: Based on the historical data of previous inflow and meteorology, the partial autocorrelation function is used to screen the previous inflow of the lag term, the inverse distance weighting method is used to reduce the weight of the meteorological data, the cross-correlation function and consistency index are used to screen the rainfall lag term, and then the correlation coefficient γ is used for feature selection. Finally, the correlation coefficient τ is used to screen the input factors;

[0041] Because input variables, including previous inflow and meteorological factors, are disordered and irregular, rationally selecting input variables can reduce the dimensionality of the data and accelerate the learning efficiency of the prediction model. The inverse distance weighting method and the simple correlation coefficient method are used to process and filter the input data of the reanalysis meteorological dataset.

[0042] The correlation between lagged runoff and total precipitation (tp) can be demonstrated using the partial autocorrelation function (PACF) and cross-correlation function (CCF), and the optimal lag order can be determined using a 95% confidence interval. Furthermore, when the correlation coefficient changes slowly and consistently and does not fall within the 95% confidence interval, the optimal lag order is determined using trial and error and the consistency index (IA). The consistency index is an evaluation metric that adds one or more lag sequences to the first lag order.

[0043] Among them, Y is the meteorological data of the research site; Z i is the meteorological value of each measurement point q; d i is the distance from each measurement point q to the research point; n is the number of measurement points; P is the weight index;

[0044] where x j and y jare the values ​​of runoff and meteorological factors in sequence j, respectively; x and y are the multi-year averages of runoff and meteorological factors, respectively; γ is the correlation coefficient between runoff and meteorological factors, τ is the partial correlation coefficient between runoff and meteorological factors; n is the number of samples;

[0045] in and Q i Represent the predicted flow and measured flow at time i respectively; represents the average value of the measured sequence; n is the length of the runoff series.

[0046] S2: Data preprocessing: Use empirical mode decomposition technology to decompose the input factors, perform data normalization, and divide the data into training and test groups;

[0047] The principle of the EEMD-ANN prediction model is to use EEMD to decompose the initial signal into several components, then use ANN (artificial neural network) to predict each component, and finally add up the predicted values ​​of each component to obtain the final prediction result. This can improve the prediction accuracy by eliminating the "impurities" in the original signal.

[0048] EEMD is an improvement on the empirical mode decomposition (EMD) algorithm, which adaptively transforms the nonstationary and nonlinear initial signal sequence f(t) into multiple stable components, including several intrinsic mode functions (IMFs) and a residual (Re). By adding white noise and continuously averaging the decomposition results, the mode mixing phenomenon in EMD is effectively avoided, enhancing the signal clarity and practicality. Ultimately, the corresponding decomposed signals (IMFs and Re) are integrated to obtain a stable initial signal.

[0049] In order to integrate the input data and speed up the learning efficiency of the model, the input signal can be normalized and converted to the interval [0,1]. The deviation normalization method can be used to normalize the input data, as shown in the following formula:

[0050] Among them, x scale is the reduced value; x i is the input value in the original sequence; x min and x max are the minimum and maximum values ​​in the original data, respectively.

[0051] S3: Model prediction and evaluation: Integrate the input factors and input them into the prediction model. Then optimize the parameters through parallel algorithms, output the prediction results, perform parallel evaluation over time, and evaluate the model through correlation coefficient, root mean square error, mean absolute error, mean relative error and certainty coefficient.

[0052] Since the invention considers the addition of meteorological factors, the diverse types of input factors lead to the diversity of decomposition quantities. Therefore, it is necessary to integrate the diverse decomposition components into the input factors of the model according to certain rules.

[0053] After integrating the input factors, they are input into the ANN model to obtain the output results. total And the learning rate η is used to adjust the weight of the model, gradually reducing the error and finally achieving the optimal prediction model.

[0054] The ensemble empirical mode decomposition (EEMD) algorithm is used to decompose the original sequence. The decomposition principle is as follows:

[0055] The propagation error E of the ANN model total yes:

[0056] By E total and η to adjust the weights:

[0057] Among them E total is the error in the propagation process; D u is the actual value; u is the predicted value; x i and y j are the input value of the model and the output value of the hidden layer, f(x) and g(x) are the activation functions of the hidden layer and the output layer respectively; w ju and w ij is the weight in each propagation, through E total and η update.

[0058] While using the EEMD-ANN model for prediction, a multi-core parallel algorithm is employed to reduce model optimization time and improve the model's prediction efficiency and accuracy. A parallel algorithm is a method that utilizes multiple processors to solve problems. With the continuous advancement of computer science and technology, modern computers are equipped with multi-core systems. Therefore, parallel algorithms can be easily used on modern computers to effectively solve large-scale data processing problems. The steps for using a parallel algorithm are shown in Figure 2. The main steps of a parallel algorithm are as follows:

[0059] (1) Parallel reading and analysis of main tasks;

[0060] (2) Decompose the main task into n subtasks according to actual needs and hardware environment;

[0061] (3) The processor is used to process and solve each subtask and obtain the output result of each subtask;

[0062] (4) Organize and merge the output results of each subtask to obtain the result of the main task.

[0063] In step S3, after obtaining the forecast results, the accuracy of the model results is evaluated by correlation coefficient, root mean square error, mean absolute error, mean relative error and coefficient of certainty;

[0064] in Represents the mean value of the forecast series.

[0065] Example 2: Sichuan and Yunnan provinces are the regions with the richest hydropower resources in China, and are also important power supply bases for China's "West-to-East Power Transmission" project, but they are also facing serious water abandonment problems. Figure 1 shows that the abandoned water energy in Sichuan and Yunnan provinces is equivalent to the hydropower generation in Germany. From 2018 to 2020, the excess energy accounted for a large proportion of hydropower generation. This increasingly serious wastewater problem should be solved in a timely manner. Therefore, hydropower stations need an effective long-term runoff forecasting system to accurately forecast the large amount of runoff generated by extreme weather events, help hydropower stations plan operations in advance, and reduce the loss of water resources and hydropower resources. The method of the present invention is used below to establish a runoff forecast model for a power station in a certain river basin in southwest my country, and the forecast results are analyzed.

[0066] Historical runoff data from January 1979 to December 2019 and 25 meteorological data points for the power station were selected as candidate input data. Table 1 shows the 25 candidate meteorological factors for the power station. Within the power station's control basin, there are 26 meteorological data monitoring points. The inverse distance weighting (IDW) method can be used to reduce the weight of the meteorological data at each monitoring point. This data was downloaded from the ERA5 website. The meteorological data covers the same time period as the historical runoff data.

[0067] Table 1

[0068] S1: Feature screening: When screening input factors, the optimal historical runoff lag order PACF (Figure 3a) and the corresponding 95% confidence interval are used to determine the optimal historical runoff. The figure shows that before the 12th order, the PACF is basically outside the 95% confidence interval, but after the 12th order, the PACF is basically reduced to within the 95% confidence interval. The PACF is truncated at p = 12. Therefore, the previous inflow Q with 12 lag orders is selected. t-1 :Q t-12(1st to 12th order lagged runoff Q) serves as the inflow input of the forecast model proposed in the present invention.

[0069] The total precipitation (tp) factor input is filtered by the CCF and IA evaluation index respectively through tp of different lag orders. As shown in Figure 3a, the correlation between runoff and tp first decreases slightly from lag number 1 to 12, and then increases slightly, but never falls within the 95% confidence interval. Obviously, the optimal lag order of tp cannot be obtained by CCF alone. In order to obtain the accurate order, the trial and error method and IA index are used to further determine the optimal lag order: the tp lag orders from 1 to 12 are used as input factors of the test model, and then the IA index is used to evaluate the correlation of each lag input factor, and the most relevant input is selected. The experimental results of the IA index are shown in Figure 3b. Among the 12 input factors, the IA value of input factor 6 is the largest; therefore, the correlation between tp and runoff in input factor 6 is the largest. Finally, it was determined that the lag order of 6 (R t-1 、R t-2 、R t-3 、R t-4 、R t-5 and R t-6 ) is used as the total precipitation (tp) input of the prediction model proposed in the present invention.

[0070] With the exception of total precipitation (tp), the lag order of the other 24 candidate meteorological factors appears to have little effect on the runoff formation process. Therefore, the correlation coefficients γ and τ can be directly used (with a confidence level α of 0.05) to identify the truly relevant meteorological factors as model inputs. The results in Figure 3c show that many candidate meteorological factors have a good single correlation with the historical runoff of the power station, and the independent coupling between meteorological factors is very low. Finally, the present invention selects meteorological factors with τ > 8.

[0071] In summary, the five filtered meteorological factors selected in the forecast model proposed in the present invention are total precipitation (tp), total atmospheric water vapor content (tcwv), total columnar water content (tcw), total cloud cover (tcc) and 2-meter dew point temperature (d2m), which are consistent with the input factors of the long-term runoff forecast framework proposed in the present invention.

[0072] S2: Data preprocessing: When decomposing the input factors, the input factors in the present invention are composed of six signals; two of them are the previous inflow with a 12th-order lag and tp with a 6th-order lag, while tcwv, tcw, tcc and d2m with a 1st-order lag constitute the remaining four. Then, the input factors are decomposed into a finite number of subsequences using the EEMD algorithm. The decomposition result is shown in Figure 4a. Since each initial signal contains different data types and attributes, the decomposition results are also different. These include the previous inflow, tp, tcc and tcwv (containing 7 IMF components), Re, and d2m and tcw (containing 6 IMF and Re components).

[0073] S3: Model Forecast and Evaluation: When integrating input factors, Figure 4a shows that the IMF6 components of all input signals contain more than one signal cycle, while the IMF7 components of the antecedent inflow, TP, TCC, and TCWV do not have a complete cycle. Therefore, based on this characteristic, the diverse components can be integrated as inputs to the ANN model, which is consistent with the modeling approach of the proposed long-term runoff forecast framework. Detailed integration results are shown in Table 2.

[0074] Table 2

[0075] In this invention, EEMD decomposes the previous inflow signal into eight components. These components only affect the input factors of the ANN model and do not participate in the underlying prediction process of the ANN. Therefore, when the number of nodes in the input layer (N1), hidden layer (N2), and output layer (N3) is known, a parallel algorithm is used to optimize other parameters such as the learning rate (η) and the number of training epochs to determine the model results. In addition, the experimental results of the fuzzy neural network (FNN) model were compared with the ANN as a reference control group.

[0076] With the support of the parallel algorithm, the optimal model parameters of the forecast model were quickly obtained, as shown in Table 3. Figure 4b also visually demonstrates the optimization time of the ANN model when the parallel algorithm is run on processors with 2, 4, 8, 16, and 32 cores. The optimization time of the ANN-Q (EEMD-ANN-Q) model decreases from 10,438 seconds on a single core to 383 seconds on 32 cores. Similarly, the ANN-QE (EEMD-ANN-QE) model exhibits the same optimization process as the ANN-Q (EEMD-ANN-Q) model. Furthermore, due to the inclusion of meteorological factors in the input data, the optimization time of the ANN-QE (EEMD-ANN-QE) model is consistently longer than that of the ANN-Q (EEMD-ANN-Q) model with the same number of cores. Therefore, the acceleration effect of the parallel algorithm is perfectly optimized, which is consistent with the computational approach of the proposed long-term runoff forecast framework.

[0077] Comparison of forecast results: Finally, the forecast indicators of each model are shown in Table 4. According to the five accuracy evaluation indicators, the forecast performance of each model is ranked as follows:

[0078] EEMD-AN-QE>EEMD-AN-Q>ANN-QE>FYN-Q>ANN-Q. Due to the combination of fuzzy theory and neural network, the prediction performance of FNN model is better than that of ordinary ANN model.

[0079] Table 3

[0080] Table 4

[0081] Model Prediction Results: The correlation of the model prediction results is clearly shown in Figure 4c (or Figures 6 and 7) (black dots represent the inflow predicted by the EEMD-ANN, and gray dots represent the inflow predicted by the ANN). From Figure 4c, we can clearly see that the black dots are densely clustered around y = x, while the gray dots are relatively scattered on both sides of y = x. This shows that under the same input factors, the prediction results of the EEMD-ANN model are more consistent with the preceding inflow than the prediction results of the conventional ANN, and therefore have higher prediction accuracy. From Figure 4c, we can clearly see that the black dots are densely clustered around y = x, while the gray dots are relatively scattered on both sides of y = x. This shows that under the same input factors, the prediction results of the EEMD-ANN model are more consistent with the preceding inflow than the prediction results of the conventional ANN, and therefore have higher prediction accuracy. In addition, from the perspective of input factors, when the selected meteorological factors are added in Figure 4c(ii), under the same conditions (EEMD-ANN-QE is 0.91; EEMD-ANN-Q is 0.85; ANN-QE is 0.83; ANN-Q is 0.81), the fitting slope of the model's predicted inflow is greater than that without adding the selected meteorological factors (as shown in Figure 4c(i)). This indicates that the forecast performance of the forecast model has been greatly improved after adding the selected meteorological factors.

[0082] In summary, it can be determined that the optimal prediction model is the EEMD-ANN-QE proposed in this invention. The proposed method significantly shortens the prediction time of power plants and greatly improves the prediction accuracy of power plants. This has high practical value for the prediction problems of large power plants in my country.

Claims

1. A runoff forecasting method based on EEMD-ANN and meteorological factors using a multi-core parallel algorithm, characterized by: It includes the following steps: S1: Feature screening: Based on the historical data of previous inflow and meteorology, the partial autocorrelation function is used to screen the previous inflow of the lag term, the inverse distance weighting method is used to reduce the weight of meteorological data, the cross-correlation function and consistency index are used to screen the rainfall lag term, and then the correlation coefficient γ is used for feature selection, and finally the correlation coefficient τ is used to screen the input factors; S2: Data preprocessing: Use the empirical mode decomposition technique to decompose the input factors, perform data normalization, and divide the data into training and test groups; S3: Model prediction and evaluation: Integrate the input factors and input them into the prediction model. Then optimize the parameters through parallel algorithms, output the prediction results, perform parallel evaluation over time, and evaluate the model through correlation coefficient, root mean square error, mean absolute error, mean relative error and certainty coefficient.

2. The runoff forecasting method based on EEMD-ANN and meteorological factors using a multi-core parallel algorithm according to claim 1, characterized in that: In the step S1, the input data of the reanalysis meteorological data set is processed and filtered using the inverse distance weighted method and the single correlation coefficient method; The correlation between lagged runoff and total precipitation is demonstrated by partial autocorrelation function and cross-correlation function, and the optimal lag order can be determined by 95% confidence interval; in addition, when the correlation coefficient changes slowly and does not belong to the 95% confidence interval, the optimal lag order needs to be determined by trial and error method and consistency index (IA), which is an evaluation index of continuously adding one or more lag sequences based on the input of the first lag order; Among them, Y is the meteorological data of the research point; Z i is the meteorological value of each measuring point q; d i is the distance from each measurement point q to the research point; n is the number of measurement points; P is the weight index; where x j and j are the values ​​of runoff and meteorological factors in sequence j, respectively; x and y are the multi-year averages of runoff and meteorological factors, respectively; γ is the correlation coefficient between runoff and meteorological factors, τ is the partial correlation coefficient between runoff and meteorological factors; n is the number of samples; in and Q i Represent the predicted flow and measured flow at time i respectively; represents the average value of the measured sequence; n is the length of the runoff series.

3. The runoff forecasting method based on EEMD-ANN and meteorological factors using a multi-core parallel algorithm according to claim 1, characterized in that: In step S2, the EEMD-ANN prediction model uses EEMD to decompose the initial signal into several components, then uses ANN to predict each component, and finally adds the predicted values ​​of each component to obtain the final prediction result, thereby improving the prediction accuracy by eliminating "impurities" in the original signal; In order to integrate the input data and speed up the learning efficiency of the model, the input signal is normalized and converted to the interval [0,1]. The deviation normalization method is used to normalize the input data, as shown in the following formula: Among them, x scale is the reduced value; x i is the input value in the original sequence; x min and x max are the minimum and maximum values ​​in the original data respectively.

4. The runoff forecasting method based on EEMD-ANN and meteorological factors using a multi-core parallel algorithm according to claim 1, characterized in that: In step S3, after the input factors are integrated, they are input into the ANN model to obtain the output result. The ANN prediction model propagates the error E total And the learning rate η is used to adjust the weight of the model, gradually reduce the error, and finally reach the best prediction model; The EEMD algorithm is used to decompose the original sequence, and the decomposition process is as follows: The propagation error E of the ANN model total yes: By E total and η to adjust the weights: Where E total is the error in the propagation process; D u is the actual value; u is the predicted value; x i and j are the input value of the model and the output value of the hidden layer, respectively. f(x) and g(x) are the activation functions of the hidden layer and the output layer, respectively. ju and w ij is the weight in each propagation, through E total and η update.

5. The runoff forecasting method based on EEMD-ANN and meteorological factors using a multi-core parallel algorithm according to claim 1, characterized in that: In step S3, while predicting through the EEMD-ANN model, a multi-core parallel algorithm is used to reduce the optimization time of the model and improve the prediction efficiency and accuracy of the model; the steps of the parallel algorithm are as follows: (1) Parallel reading and analysis of main tasks; (2) Decompose the main task into n subtasks according to actual needs and hardware environment; (3) The processor is used to process and solve each subtask and obtain the output result of each subtask; (4) Sort and merge the output results of each subtask to obtain the result of the main task.

6. The runoff forecasting method based on EEMD-ANN and meteorological factors using a multi-core parallel algorithm according to claim 1, characterized in that: In step S3, after obtaining the forecast results, the accuracy of the model results is evaluated by correlation coefficient, root mean square error, mean absolute error, mean relative error and certainty coefficient; in Represents the mean value of the forecast series.

Images