Parallel architecture runoff prediction method fusing feature decomposition and error correction strategy

Abstract

The invention belongs to the technical field of hydrological prediction, and discloses a parallel architecture runoff prediction method fusing characteristic decomposition and an error correction strategy, and the method comprises the steps: collecting a runoff time sequence of a target region, and carrying out the decomposition of the runoff time sequence through employing a VMD optimized through an SSA algorithm; screening a subsequence most related to runoff change as a runoff prediction factor; inputting the runoff prediction factors into a parallel CNN-GRU model to extract spatial-temporal characteristics, and optimizing parameters of the parallel CNN-GRU model by using an SSA algorithm; inputting a runoff prediction factor into the optimized parallel CNN-GRU model for preliminary prediction, and predicting an error value based on the optimized parallel CNN-GRU model; and superposing the prediction error value to the initial prediction value to correct the prediction deviation, and obtaining a final runoff prediction value. According to the method, by combining the parallel CNN-GRU architecture, the SSA optimization algorithm and the error correction mechanism, the precision of regional runoff prediction is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of hydrological prediction, and specifically relates to a parallel architecture runoff prediction method integrating feature decomposition and error correction strategy. Background Art

[0002] In water-scarce areas, water shortage and uneven temporal and spatial distribution have always been one of the main problems restricting regional sustainable development. Accurate and reliable runoff prediction and forecasting technology is an effective means to promote efficient use of water resources and regional sustainable development. Among them, medium- and long-term runoff forecasts can provide an important scientific basis for water resource allocation, hydropower operation and management, and are crucial to solving water problems caused by social and economic development. However, as an element of a complex hydrodynamic evolution system, runoff shows a strong correlation with multiple factors such as human activities, vegetation type, soil type, evaporation intensity and precipitation. Runoff sequences are highly nonlinear, unstable and complex.

[0003] In recent years, with the rapid development of computer technology and the continuous improvement of machine learning theory, data-driven methods represented by machine learning models have been well applied in runoff prediction. Machine learning models have a strong ability to capture nonlinear relationships and can effectively predict the changing trend of runoff. However, in current practical applications, the performance of machine models faces some challenges, such as the impact of feature preprocessing of input data and hyperparameter selection on model performance, the difficulty of a single model to fully identify the internal changes of the data sequence, and the uncertainty of the prediction result error. Summary of the invention

[0004] In view of the above defects or improvement needs of the prior art, the present invention aims to propose a runoff prediction method based on a parallel architecture hybrid model with characteristic mode decomposition and error correction strategy, so as to overcome the influence of the uncertainty of the initial prediction error of the single feature input and the machine learning model on the runoff prediction accuracy in the prior art, and improve the accuracy of the runoff prediction model.

[0005] To achieve the above object, the present invention provides a parallel architecture runoff prediction method integrating feature decomposition and error correction strategy, the method comprising the following steps: S100: Collect the runoff time series of the target area and decompose the runoff time series using VMD (variational mode decomposition) optimized by SSA (sparrow search) algorithm; S200: selecting the subsequence most relevant to runoff change from the decomposed subsequences as the runoff prediction factor; S201: Analyze the decomposed subsequences using autocorrelation function and partial correlation function to determine time-related features; S202: using a hierarchical prediction factor selection algorithm to select the subsequence most relevant to runoff change as the runoff prediction factor; S300: Input the runoff prediction factor into the parallel CNN-GRU (Convolutional Neural Network-Gated Recurrent Unit) model to extract spatiotemporal features, and use the SSA algorithm to optimize the parameters of the parallel CNN-GRU model; S400: inputting the runoff prediction factor into the optimized parallel CNN-GRU model for preliminary prediction, and predicting the error value based on the optimized parallel CNN-GRU model; S500: superimposing the prediction error value onto the preliminary prediction value to correct the prediction deviation and obtain the final runoff prediction value.

[0006] Furthermore, the decomposition of the runoff time series by using the VMD optimized by the SSA algorithm in S100 specifically includes: S101: Decompose the runoff time series using VMD to obtain several subsequences and construct a constrained variational problem; S102: Introducing a quadratic penalty factor and the Lagrange multiplication operator Convert the constrained variational problem to an unconstrained problem; S103: Calculate the corresponding modal components based on the extreme value solution and the center frequency ; S104: Taking the minimum average envelope entropy as the objective function, the VMD parameters are iteratively optimized using the SSA algorithm; S105: After the iteration is completed, the optimal decomposition subsequence is output.

[0007] Furthermore, the constrained variational problem described in S101 is constructed according to the following formula:

[0008] Where: K is the number of modes to be decomposed; and are the functions of different modal components and the center frequency; is the Dirac function; is the convolution operation; for The Hilbert transform result of ; is the original time series; Represents the imaginary part of a complex number; Express Find the partial derivative.

[0009] Furthermore, the secondary penalty factor is introduced as described in S102 and the Lagrange multiplication operator The formula for converting a constrained variational problem into an unconstrained problem is as follows:

[0010] Where: L represents the Lagrangian function.

[0011] Further, in step S103, the corresponding modal component is obtained according to the extreme value solution and the center frequency According to the following formula:

[0012]

[0013] Where: , , and They are , , and The Fourier transform of Indicates the signal k The modal components pass through n +Time function after 1 iteration; The current remaining amount Wiener filtering; Indicates other modes ( ); is the center of gravity of the power spectrum of the current mode function.

[0014] Furthermore, the calculation formula of the average envelope entropy in S104 is as follows:

[0015] In the formula, and is the best parameter combination, k is the number of modal decompositions, For each intrinsic mode function, the envelope entropy is calculated as follows:

[0016]

[0017] Where N is the number of samples, for The standardized form of is the envelope signal of the source signal after Hibert transform.

[0018] Furthermore, the step of selecting the subsequence most relevant to the runoff change from the decomposed subsequences as the runoff prediction factor in S200 specifically includes: S201: Analyze the decomposed subsequences using autocorrelation function and partial correlation function to determine time-related features; S202: Using a hierarchical prediction factor selection algorithm to select subsequences most relevant to runoff changes as runoff prediction factors.

[0019] Furthermore, the step of using the hierarchical prediction factor selection algorithm to select the subsequence most relevant to the runoff change as the runoff prediction factor in S202 specifically includes: (1) Construct a hierarchical evaluation model; (2) Construct a judgment matrix; (3) Determine the weight of each subsequence based on the eigenvalues ​​and eigenvectors of the judgment matrix; (4) Sort each subsequence by weight from large to small, and use the subsequence with the highest weight as the runoff prediction factor.

[0020] Furthermore, the step S300 of inputting the runoff prediction factor into the parallel CNN-GRU model to extract spatiotemporal features and optimizing the parameters of the parallel CNN-GRU model using the SSA algorithm specifically includes: S301: Input the subsequences used as prediction factors into the CNN and GRU models respectively, extract the local features between different subsequences through CNN, and extract the temporal features of each subsequence through GRU; S302: The parallel CNN-GRU model is trained using the root mean square error as the loss function, and the SSA algorithm is used to optimize the model parameters.

[0021] Furthermore, in S400, the runoff prediction factor is input into the optimized parallel CNN-GRU model for preliminary prediction, and the prediction error value based on the optimized parallel CNN-GRU model specifically includes: S401: Input the sub-series as prediction factors into the optimized parallel CNN-GRU model and output a preliminary prediction value; S402: Calculate the error between the preliminary predicted value and the actual observed value, and arrange them in chronological order to form an error sequence; S403: Input the error sequence into the optimized parallel CNN-GRU model to obtain a prediction error value.

[0022] Compared with the prior art, the present invention can achieve the following beneficial effects: (1) Aiming at the multilateral and instability of monthly runoff series, the present invention significantly improves the accuracy of regional runoff prediction by combining the parallel CNN-GRU architecture, SSA optimization algorithm and error correction mechanism. The parallel strategy extracts different features of the original runoff series through different base models, which can expand the complex features of the runoff series and predict and aggregate different features in different ways, and finally obtain more accurate prediction results. Compared with the two single models, the prediction performance of the parallel CNN-GRU (PCG) model is better. Compared with other optimization algorithms, SSA enhances the local optimization computing power, avoids being trapped in the local optimization results, and can provide higher quality solutions. In addition, the SSA algorithm can shorten the prediction time of the model, expand its performance advantage, and make it more competitive. VMD with parameter optimization function can effectively preprocess hydrological data, reduce the complexity of the original runoff series, effectively enhance the generalization ability of the model, and reduce the pressure of the subsequent main prediction model.

[0023] (2) The present invention adopts an error correction strategy to effectively reduce the prediction error of the runoff extreme point and improve the model's peak prediction performance, providing an important scientific basis for the rational allocation of regional water resources and helping to promote the sustainable development of the region.

[0024] (3) The multi-method hybrid model of decomposition, optimization algorithm, parallel mode and error correction strategy proposed in this paper improves the accuracy of monthly runoff forecast and significantly enhances the prediction effect of extreme point data. It is an effective and useful strategy. This provides a new perspective for the medium- and long-term prediction and accuracy improvement of hydrological runoff series. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of a multi-strategy monthly runoff prediction hybrid model (SVMD-SPCG-EC) according to a specific embodiment of the present invention; Figure 2 It is a monthly runoff sequence diagram of Beidao hydrological station, Zhangjiashan hydrological station and Huaxian hydrological station in a specific embodiment of the present invention; Figure 3 It is a component diagram after normalization and decomposition of the initial runoff sequence of Beidao hydrological station, Zhangjiashan hydrological station and Huaxian hydrological station in a specific embodiment of the present invention; Figure 4 The figures are the autocorrelation function (ACF) and partial correlation function (PACF) of the runoff series of the three hydrological stations of Zhangjiashan, Huaxian and Beidao according to the specific embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0027] In order to improve the runoff prediction accuracy, the present invention provides a parallel architecture runoff prediction method integrating feature decomposition and error correction strategy, the method comprising the following steps: 1. Collect the runoff time series of the target area and decompose the runoff time series using VMD optimized by SSA algorithm.

[0028] The historical long series of monthly runoff data in the target area were collected and normalized. The sparrow search algorithm (SSA) was used to optimize the parameters of the variational mode decomposition (VMD) to determine the optimal parameters, including the penalty factor and the number of modes. By minimizing the average envelope entropy, the parameter combination that makes the VMD decomposition effect the best was found. The VMD parameters optimized by SSA were applied to the original runoff sequence to decompose it into multiple intrinsic mode functions (IMFs) and residual components. This step aims to reduce the nonlinearity and non-stationarity of the sequence and prepare for subsequent model input.

[0029] 1.1 VMD is used to decompose the original runoff sequence, and the highly complex and non-stationary original runoff sequence is decomposed into a series of low-complexity and relatively stable subsequences. The decomposition process can be regarded as an iterative solution process of the constrained variation problem. The resulting constrained variation problem is:

[0030] Where: K is the number of modes to be decomposed; and are the functions of different modal components and the center frequency; is the Dirac function; is the convolution operation; for The Hilbert transform result of ; is the original time series; Represents the imaginary part of a complex number; Express Find the partial derivative.

[0031] 1.2 Introducing the quadratic penalty factor and the Lagrange multiplication operator Convert a constrained variational problem to an unconstrained problem:

[0032] Where: L represents the Lagrangian function.

[0033] 1.3 To adjust different modal components and center frequencies, the corresponding modal components are obtained according to the corresponding extreme value solutions. and the center frequency :

[0034]

[0035] Where: , , and They are , , and The Fourier transform of Indicates the signal k The modal components pass through n +Time function after 1 iteration; The current remaining amount Wiener filtering; Indicates other modes ( ); is the center of gravity of the power spectrum of the current mode function.

[0036] 1.4 Taking the minimum average envelope entropy as the objective function, SSA is used to optimize the dual parameters of VMD to achieve adaptive decomposition of monthly runoff series.

[0037] The calculation formula of average envelope entropy is as follows:

[0038] In the formula, and is the best parameter combination, k is the number of modal decompositions; For the envelope entropy of each IMFs, the calculation formula is as follows:

[0039]

[0040] Where N is the number of samples, for The standardized form of is the envelope signal of the source signal after Hibert transform.

[0041] The specific process of SSA optimization includes: (1) Obtain the initial parameters of SSA, including population size, maximum number of iterations, etc.

[0042] (2) Calculate the initial population fitness.

[0043] (3) Calculate the positions of the discoverer, participants, and scouts.

[0044] The calculation formula for the finder's position is:

[0045] Where: For the The position of the i-th sparrow in the j-th dimension during the iteration; is the maximum number of iterations; is a random number with a value range of (0,1], The warning value range is [0,1]; The safety value range is [0.5,1]; Q follows a normal distribution and is a random number in the interval [0,1]; is a 1×d matrix with all elements set to 1.

[0046] The participant position is calculated as:

[0047] Where: , A is a 1×d matrix that randomly generates 1 and -1; For the The position of the i-th sparrow in the j-th dimension during the iteration; For the population The worst position at the iteration; For the population The optimal position at the iteration For the population The optimal position at the iteration; n is the total number of sparrows.

[0048] The calculation formula for the scout's position is:

[0049] Where: is the step size control parameter; It is the direction of movement of the sparrow; is a very small number; For the population The worst position at the iteration; , , They are the individual, best and worst fitness of sparrows respectively.

[0050] (4) Check whether the boundary is exceeded.

[0051] (5) Update position and population fitness.

[0052] (6) Determine whether the maximum number of iterations has been reached. If so, output the optimal decomposition result. Otherwise, return to step (3) to continue iterating.

[0053] 2. Select the subsequence most relevant to runoff change from the decomposed subsequences as the runoff prediction factor.

[0054] The decomposed subsequences were analyzed using autocorrelation function (ACF) and partial correlation function (PACF) to determine the time-related characteristics. A hierarchical predictor selection algorithm was used to identify and select the subsequences most relevant to runoff changes as runoff predictors, which can build a more accurate prediction model.

[0055] 2.1 The decomposed subsequences are analyzed using the autocorrelation function (ACF) and partial correlation function (PACF) to determine the time-related characteristics.

[0056] 2.1.1 Calculate ACF. Calculate ACF for each subsequence to determine the correlation between the sequence and its lagged version, draw an ACF graph, and observe the correlation values ​​at different lag orders.

[0057] 2.1.2 Calculate PACF. Calculate PACF for each subsequence to determine the direct correlation between the series and its lagged version after controlling for other lag effects, draw a PACF graph, and observe the correlation values ​​at different lag orders.

[0058] 2.1.3 Identify significant correlation peaks through ACF and PACF plots, which indicate important time-dependent structures in the time series. For ACF plots, look for significant positive or negative correlation lag orders; for PACF plots, look for significant peaks, which usually correspond to the parameters of the model.

[0059] 2.2 A hierarchical prediction factor selection algorithm is used to identify and select the prediction factors most relevant to runoff change, construct a judgment matrix and determine the weight of each subsequence, and determine the required subsequence as the prediction factor based on the weight.

[0060] 2.2.1 Constructing a hierarchical evaluation model First, it is necessary to determine the target layer, criterion layer, and solution layer of the entire decision-making event.

[0061] 2.2.2 Constructing the judgment matrix. In the criterion layer, each criterion is compared pairwise to construct a pairwise comparison matrix.

[0062] 2.2.3 Perform pairwise comparisons. Score each pair of criteria using a 1-9 scale, reflecting the importance of one criterion relative to the other.

[0063] 2.2.4 Calculate weights. Determine the relative weights of each criterion by calculating the eigenvalues ​​and eigenvectors of the judgment matrix.

[0064] 2.2.5 Consistency test. Check the consistency of the judgment matrix to ensure that the decision maker's judgment is reasonable and consistent. The consistency ratio (CR) should be less than 0.1 to ensure the consistency of the judgment.

[0065] 3. Input the runoff prediction factor into the parallel CNN-GRU model to extract spatiotemporal features, and use the SSA algorithm to optimize the parameters of the parallel CNN-GRU model.

[0066] The sub-series as predictors are input into the parallel CNN-GRU model to extract spatiotemporal features, where CNN is responsible for extracting spatial features and GRU is responsible for extracting temporal features, ensuring that the model can process spatial and temporal information simultaneously and enhance prediction capabilities. The SSA algorithm is used to optimize the parameters of the parallel CNN-GRU model to reduce the uncertainty of parameter selection and improve the prediction accuracy of the model.

[0067] 3.1 The sub-series used as predictors are input into the CNN and GRU models respectively, where CNN is used to extract local features between different decomposition sequences, and GRU is used to extract the temporal features of each decomposition structure.

[0068] 3.1.1 Input the subsequences as predictors into the CNN model. Since CNN is good at capturing spatial features (which can be understood as local patterns or trends in time series), it is suitable for extracting local features. CNN processes the input subsequences through convolutional layers to identify and extract local patterns or trends in the sequence, which may be periodic or seasonal features in runoff changes.

[0069] 3.1.2 Input the same subsequence into the GRU model. GRU is a variant of recurrent neural network (RNN) that is particularly suitable for processing and predicting sequence data and can capture long-term dependencies. GRU processes sequence data through its gating mechanism (update gate and reset gate) to extract temporal features in the time series, such as trends, cycles, and anomalies.

[0070] 3.2 The subsequences used as predictors are randomly divided into training sets and test sets. The training set is input into the parallel CNN-GRU model. The model is trained with the root mean square error (RMSE) as the loss function, and the SSA algorithm is used to optimize the model parameters.

[0071] The specific process of using the SSA algorithm to optimize model parameters is as follows: (1) Read data and obtain SSA initial parameters (population size, maximum number of iterations, etc.); (2) Calculate the initial fitness value based on the parameters; (3) Update the locations of discoverers, participants, and scouts; (4) Calculate the fitness of the updated sparrow’s new position; (5) Modify the current optimal solution and its fitness; (6) Repeat (3)-(5) until the iteration stop condition is reached and output the optimal solution.

[0072] 3.3 Bring the SSA optimization parameters into the parallel CNN-GRN to create a prediction model.

[0073] 4. Input the runoff prediction factors into the optimized parallel CNN-GRU model for preliminary prediction, and predict the error value based on the optimized parallel CNN-GRU model.

[0074] The optimized parallel CNN-GRU prediction model is used to make preliminary predictions on the test sample data, and the error between the preliminary prediction results of the test sample data and the actual observed value (test sample data) is calculated to form an error sequence. The error sequence is input into the parallel CNN-GRU model optimized by SSA to predict the error value.

[0075] 4.1 Input the test set into the optimized parallel CNN-GRN model and output the preliminary prediction value.

[0076] 4.2 Calculate the error between the initial predicted value and the actual observed value to form an error sequence.

[0077] 4.2.1 For each corresponding pair of actual observations and preliminary predictions, calculate the difference between them. The error can be obtained by simple subtraction: Error = actual observation − preliminary prediction.

[0078] 4.2.2 Arrange all calculated error values ​​in chronological order to form an error sequence. This sequence will be used to analyze the prediction deviation of the model and provide a basis for further model optimization.

[0079] 4.2.3 The parallel CNN-GRN model optimized by the SSA algorithm is used to predict and analyze the error sequence to obtain the predicted error value.

[0080] 5. Superimpose the prediction error value on the preliminary prediction value to correct the prediction deviation and obtain the final runoff prediction value.

[0081] The preliminary prediction value is combined with the predicted error value, and the final runoff prediction value is obtained through error correction strategy adjustment to ensure that the prediction result is closer to the actual observed value and improve the accuracy of the prediction. The prediction performance of the model is evaluated using evaluation indicators such as root mean square error (RMSE), mean absolute error (MAE), determination coefficient (R) and Nash coefficient (NSC).

[0082] 5.1 Combine the preliminary prediction value with the prediction error value and adjust the error correction strategy to obtain the final runoff prediction value.

[0083] 5.1.1 For each time point, the forecast error value is added to the preliminary forecast value to correct the forecast deviation. If the forecast error value is negative, its absolute value is subtracted from the preliminary forecast value. Its expression is as follows:

[0084]

[0085] Where: represents the actual observed value, is the initial forecast value, is the error sequence, is the prediction error value, Represents the final runoff prediction value.

[0086] Make sure all calculations are performed on the same numerical scale to avoid errors caused by inconsistent scales.

[0087] 5.1.2 As needed, the corrected forecast values ​​shall be smoothed to reduce the impact of random fluctuations and outliers. The smoothing process may be performed by moving average, exponential smoothing, etc.

[0088] 5.2 The prediction performance of the model was evaluated using evaluation indicators such as root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R) and Nash coefficient (NSC).

[0089] 5.2.1 Calculate evaluation indicators such as root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R), and Nash coefficient (NSC). The above evaluation indicators are expressed as follows:

[0090]

[0091]

[0092]

[0093] in: and are the actual observed values ​​and their average values, and are the predicted values ​​and their average values.

[0094] 5.2.2 Evaluate the prediction performance of the model based on the evaluation indicators.

[0095] Example This embodiment provides a parallel architecture runoff prediction method that integrates feature decomposition and error correction strategy. Figure 1 The multi-strategy monthly runoff prediction hybrid model (SVMD-SPCG-EC) shown in the figure is used to predict runoff. The method includes the following steps: 1. Collect the historical long series of monthly runoff data in the target area and normalize them.

[0096] In this embodiment, the monthly runoff data comes from the Beidao hydrological station, Zhangjiashan hydrological station and Huaxian hydrological station. Figure 2 The initial runoff sequence information of Zhangjiashan Station, Huaxian Station and Beidao Station is shown. The monthly runoff series of the three hydrological stations are 660 months from 1956 to 2010.

[0097] 2. VMD optimized by SSA algorithm is used to decompose the runoff time series.

[0098] The components of the normalized and decomposed initial runoff series at the Beidao, Zhangjiashan and Huaxian hydrological stations are shown in Figure 2. Figure 3 shown.

[0099] 3. Select the subsequence most relevant to runoff change from the decomposed subsequences as the runoff prediction factor.

[0100] 3.1 The decomposed subsequences are analyzed using the autocorrelation function (ACF) and partial correlation function (PACF) to determine the time-related characteristics.

[0101] The ACF and PACF diagrams of runoff series at the three hydrological stations of Zhangjiashan, Huaxian and Beidao are shown in Figure 4 As shown, the ACF and PACF plots are used to identify significant correlation peaks, which indicate important time-dependent structures in the time series. For the ACF plot, look for significant positive or negative correlation lag orders; for the PACF plot, look for significant peaks, which usually correspond to the parameters of the model.

[0102] 3.2 A hierarchical predictor selection algorithm is used to identify and select the subsequences most relevant to runoff changes as runoff predictors.

[0103] 4. Input the runoff prediction factor into the parallel CNN-GRU model to extract spatiotemporal features, and use the SSA algorithm to optimize the parameters of the parallel CNN-GRU model.

[0104] The subsequences used as predictors are input into the parallel CNN-GRU model to extract spatiotemporal features, where CNN is responsible for extracting spatial features and GRU is responsible for extracting temporal features. The subsequences used as predictors are randomly divided into training sets and test sets, and the ratio of the length of the training and test time series in the runoff data of each hydrological station is 8:2. The training set is input into the parallel CNN-GRU model, and the model is trained with the root mean square error (RMSE) as the loss function. The SSA algorithm is used to optimize the parameters of the parallel CNN-GRU model.

[0105] 5. Input the runoff prediction factors into the optimized parallel CNN-GRU model for preliminary prediction, and predict the error value based on the optimized parallel CNN-GRU model.

[0106] The decomposed runoff sequence is input into the parallel CNN-GRU model optimized by SSA to obtain preliminary prediction results. CNN (Convolutional Neural Network), GRU (Gated Recurrent Unit), CNN and GRU parallel PCG (Parallel CNN-GRU model, Parallel CNN-GRU), PSO-PCG (Parallel CNN-GRU model optimized by particle swarm optimization algorithm PSO), SSA-PCG (SPCG) (Parallel CNN-GRU model optimized by sparrow optimization algorithm SSA), EMD-PCG (Parallel CNN-GRU model after processing data by empirical mode decomposition EMD method), CEEMD-PCG (Parallel CNN-GRU model after processing data by complete empirical mode decomposition CEEMD method), SVMD-PCG (Parallel CNN-GRU model after processing data by variational mode decomposition VMD method), EMD-SPCG (Parallel CNN-GRU model after processing data by empirical mode decomposition EMD method and then optimized by sparrow optimization algorithm SSA) are used as the prediction results. model), CEEMD-SPCG (a parallel CNN-GRU model that uses the complete empirical mode decomposition (CEEMD) method to process data and then uses the sparrow optimization algorithm (SSA) to optimize the model), SVMD-SPCG (a parallel CNN-GRU model that uses the variational mode decomposition (VMD) method to process data and then uses the sparrow optimization algorithm (SSA) to optimize the model), and SVMD-SPCG-EC (a multi-strategy combination model that adds the error correction module EC to the SVMD-SPCG model) are applied to the monthly runoff prediction of the three hydrological stations of Zhangjiashan, Huaxian and Beidao, and a comparison is made to determine the superiority of the model.

[0107] 5.1 CNN, GRU and PCG: For both CNN and GRU, the normalized original runoff sequence is used as data input, and the data window size is determined based on the ACF and PACF time series correlation results. The CNN model uses a two-layer one-dimensional convolution structure with a convolution kernel size of 1×3; the GRU model uses two 256-dimensional hidden layers. The PCG model uses the same CNN and GRU model architectures, and achieves the synthesis of parallel results and outputs the results through the fully connected layer.

[0108] 5.2 PSO-PCG and SSA-PCG: The PSO algorithm and SSA algorithm are used to optimize the hyperparameters of PCG respectively, and the parameters obtained by the two optimization algorithms are applied to the PCG model for prediction. The population size of the PSO optimization algorithm is 20, c1=c2=0.5, the inertia coefficient is 0.8, and the maximum number of iterations is 100; the population size of the SSA optimization algorithm is 20, and the maximum number of iterations is 100. The optimization range of the hyperparameters is set to [0.0001, 0.01].

[0109] 5.3 EMD-PCG, CEEMD-PCG and SVMD-PCG: The normalized time series data is decomposed by three different signal methods: EMD, CEEMD and VMD. EMD and CEEMD adaptively decompose the normalized initial data into IMFs; VMD obtains the optimal decomposition by the SSA optimization algorithm with the goal of minimizing entropy. Finally, the decomposition results are input into the PSG model for prediction.

[0110] 5.4 EMD-SPCG, CEEMD-SPCG and SVMD-SPCG: The normalized initial runoff sequence decomposed by EMD, CEEMD and SVMD is used as input, and the SSA algorithm is used to optimize the PCG model parameters under multiple inputs. Finally, the decomposition results are input into the optimal model to obtain the prediction results.

[0111] 5.5 SVMD-SPCG-EC: The normalized initial runoff sequence decomposed by SVMD is input into the SPCG model to obtain the initial prediction value and prediction error. The error sequence is then re-input into SVMD-SPCG to obtain the predicted deviation result.

[0112] 6. Superimpose the prediction error value on the preliminary prediction value to correct the prediction deviation and obtain the final runoff prediction value.

[0113] 6.1 Combine the preliminary prediction value with the prediction error value and adjust the error correction strategy to obtain the final runoff prediction value.

[0114] 6.2 Evaluate the predictive performance of the model.

[0115] According to the evaluation indicators, such as root mean square error (RMSE), mean absolute error (MAE), determination coefficient (R) and Nash coefficient (NSC), the proposed multi-strategy combination model (SVMD-SPCG-EC) is used to predict the monthly runoff series of Zhangjiashan hydrological station, Huaxian hydrological station and Beidao hydrological station, respectively. This example uses four evaluation indicators, RMSE, MAE, R and NSC, to evaluate the prediction accuracy of the proposed model and 12 comparison models. The results show that SVMD-SPCG-EC shows the best prediction ability and fitting effect after adding the error correction module, demonstrating the model's generalization ability for time data of different dimensions. The proposed multi-method combination model has advantages in runoff prediction.

[0116] The above-mentioned embodiments only express the implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A parallel architecture runoff prediction method integrating feature decomposition and error correction strategy, characterized in that: The method comprises the following steps: S100: Collect the runoff time series of the target area and decompose the runoff time series using VMD optimized by SSA algorithm; S200: selecting the subsequence most relevant to runoff change from the decomposed subsequences as the runoff prediction factor; S201: Analyze the decomposed subsequences using autocorrelation function and partial correlation function to determine time-related features; S202: using a hierarchical prediction factor selection algorithm to select the subsequence most relevant to runoff change as the runoff prediction factor; S300: Input the runoff prediction factor into the parallel CNN-GRU model to extract spatiotemporal features, and use the SSA algorithm to optimize the parameters of the parallel CNN-GRU model; S400: inputting the runoff prediction factor into the optimized parallel CNN-GRU model for preliminary prediction, and predicting the error value based on the optimized parallel CNN-GRU model; S500: superimposing the prediction error value onto the preliminary prediction value to correct the prediction deviation and obtain the final runoff prediction value.

2. The parallel architecture runoff prediction method integrating feature decomposition and error correction strategy according to claim 1 is characterized in that: The decomposition of the runoff time series using VMD optimized by the SSA algorithm described in S100 specifically includes: S101: Decompose the runoff time series using VMD to obtain several subsequences and construct a constrained variational problem; S102: Introducing a quadratic penalty factor and the Lagrange multiplication operator Convert the constrained variational problem into an unconstrained problem; S103: Calculate the corresponding modal components based on the extreme value solution and the center frequency ; S104: Taking the minimum average envelope entropy as the objective function, the VMD parameters are iteratively optimized using the SSA algorithm; S105: After the iteration is completed, the optimal decomposition subsequence is output.

3. The parallel architecture runoff prediction method integrating feature decomposition and error correction strategy according to claim 2 is characterized in that: The constrained variational problem described in S101 is constructed according to the following formula: Where: K is the number of modes to be decomposed; and are the functions of different modal components and the center frequency; is the Dirac function; is the convolution operation; for The Hilbert transform result of ; is the original time series; Represents the imaginary part of a complex number; Express Find the partial derivative.

4. The parallel architecture runoff prediction method integrating feature decomposition and error correction strategy according to claim 3 is characterized in that: S102 introduces a secondary penalty factor and the Lagrange multiplication operator The formula for converting a constrained variational problem into an unconstrained problem is as follows: Where: L represents the Lagrangian function.

5. The parallel architecture runoff prediction method integrating feature decomposition and error correction strategy according to claim 4 is characterized in that: S103: Obtaining the corresponding modal component according to the extreme value solution and the center frequency According to the following formula: Where: , , and They are , , and The Fourier transform of The signal k The modal components pass through n +Time function after 1 iteration; The current remaining amount Wiener filtering; Indicates other modes ( ); is the center of gravity of the power spectrum of the current mode function.

6. The parallel architecture runoff prediction method integrating feature decomposition and error correction strategy according to claim 5 is characterized in that: The calculation formula of the average envelope entropy in S104 is as follows: In the formula, and is the best parameter combination, k is the number of modal decompositions, For each intrinsic mode function, the envelope entropy is calculated as follows: Where N is the number of samples, for The standardized form of is the envelope signal of the source signal after Hibert transform.

7. The parallel architecture runoff prediction method integrating feature decomposition and error correction strategy according to claim 1 is characterized in that: S202 adopts the hierarchical prediction factor selection algorithm to select the subsequence most relevant to the runoff change as the runoff prediction factor, which specifically includes: (1) Construct a hierarchical evaluation model; (2) Construct a judgment matrix; (3) Determine the weight of each subsequence based on the eigenvalues ​​and eigenvectors of the judgment matrix; (4) Sort each subsequence by weight from large to small, and use the subsequence with the highest weight as the runoff prediction factor.

8. The parallel architecture runoff prediction method integrating feature decomposition and error correction strategy according to claim 1 is characterized in that: S300 inputs the runoff prediction factor into the parallel CNN-GRU model to extract spatiotemporal features, and uses the SSA algorithm to optimize the parameters of the parallel CNN-GRU model, specifically including: S301: Input the subsequences used as prediction factors into the CNN and GRU models respectively, extract the local features between different subsequences through CNN, and extract the temporal features of each subsequence through GRU; S302: The parallel CNN-GRU model is trained using the root mean square error as the loss function, and the SSA algorithm is used to optimize the model parameters.

9. The parallel architecture runoff prediction method integrating feature decomposition and error correction strategy according to claim 1 is characterized in that: In S400, the runoff prediction factor is input into the optimized parallel CNN-GRU model for preliminary prediction, and the prediction error value based on the optimized parallel CNN-GRU model specifically includes: S401: Input the sub-series as prediction factors into the optimized parallel CNN-GRU model and output a preliminary prediction value; S402: Calculate the error between the preliminary predicted value and the actual observed value, and arrange them in chronological order to form an error sequence; S403: Input the error sequence into the optimized parallel CNN-GRU model to obtain a prediction error value.

Images