Small hydropower load prediction method based on quadratic mode decomposition and improved sparrow algorithm

Abstract

The invention belongs to the field of power transmission network load prediction, and particularly provides a small hydropower load prediction method based on quadratic modal decomposition and an improved sparrow algorithm, which comprises the following steps: performing primary decomposition on load data by using variational modal decomposition to obtain a modal component and a residual sequence; carrying out secondary decomposition on the residual error by adopting improved self-adaptive noise complete set empirical mode decomposition; calculating permutation entropies of all mode components after decomposition, screening out low-frequency components and intermediate-frequency components according to a set threshold value, and reconstructing the low-frequency components and the intermediate-frequency components; building a two-way gating circulation unit network load prediction model, performing parameter optimization on the two-way gating circulation unit network load prediction model by using a sparrow algorithm based on firefly improvement, and predicting the screened low and intermediate frequency components through the optimized two-way gating circulation unit network load prediction model. And superposing the low-frequency predicted value and the intermediate-frequency predicted value to obtain a final predicted value. According to the prediction method, high-precision load prediction can be carried out aiming at the characteristics of the small hydropower station.

Description

Technical Field

[0001] The invention belongs to the field of transmission network load prediction, and in particular, relates to a small hydropower load prediction method based on quadratic mode decomposition and an improved sparrow algorithm. Background Art

[0002] As the scale of the power grid continues to expand, power inputs are becoming more and more diversified, and the proportion of renewable energy is increasing. For example, hydropower, wind power, photovoltaics and biomass power generation, etc., but their disadvantages are also obvious, all of which are that the power generation is uncontrollable, which increases the difficulty of load forecasting. For hydropower generation, which ranks second among renewable energy sources, it is widely used because of its low investment, short cycle, simple maintenance, and stable electricity prices. In some areas with abundant water resources, small hydropower generation is very abundant, which can not only meet the local load, but also be incorporated into the large power grid. However, due to its poor regulation ability, the output of small hydropower varies greatly under different rainfall conditions, and there are also flood seasons and dry seasons, which brings great difficulties to the accurate prediction of the regional network supply load curve.

[0003] At present, the methods of load forecasting at home and abroad are mainly divided into two categories: mathematical statistical model forecasting method and artificial intelligence forecasting method. Among them, the forecasting methods based on traditional mathematical statistical models include time series method, exponential smoothing method, multivariate linear regression, and gray model method. The advantages of this type of method are simple model construction and fast calculation speed; artificial intelligence methods include artificial neural network method and support vector machine. Both statistical model forecasting and artificial intelligence forecasting methods have corresponding research results in the field of load forecasting, among which load forecasting research for areas with many small hydropower stations is still relatively rare. There are a few related studies that use methods such as meteorological regression short-term load forecasting and two-stage reduction load forecasting. However, most of the research is not in-depth and detailed enough. At present, the small hydropower generation load basically relies on the experience of forecasting staff to formulate curves, but the accuracy is not very high. Therefore, there is an urgent need for a refined load forecasting method for areas with many small hydropower stations. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a small hydropower load forecasting method based on quadratic mode decomposition and improved sparrow algorithm, which can perform high-precision load forecasting according to the characteristics of small hydropower.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a small hydropower load forecasting method based on quadratic mode decomposition and improved sparrow algorithm, comprising the following steps:

[0006] S1. Data analysis and processing, including the following steps:

[0007] S1.1. Decompose the load data once using variational modal decomposition to obtain modal components and residual sequences;

[0008] S1.2, performing secondary decomposition of the residual by adopting improved adaptive noise complete set empirical mode decomposition;

[0009] S1.3, calculate the permutation entropy of all modal components after the secondary decomposition in S1.2, and select the low and medium frequency components according to the set threshold and reconstruct them;

[0010] S2. Build a bidirectional gated cyclic unit network load prediction model, use the sparrow algorithm improved by Firefly to optimize the parameters of the bidirectional gated cyclic unit network load prediction model, and use the optimized bidirectional gated cyclic unit network load prediction model to predict the low and medium frequency components screened out in step S1.3. The predicted values ​​of low frequency and medium frequency are superimposed to obtain the final predicted value.

[0011] The present invention provides a small hydropower load forecasting method based on quadratic mode decomposition and improved sparrow algorithm, which has the following beneficial effects:

[0012] 1. This paper proposes a short-term prediction model for small hydropower load based on secondary decomposition and reconstruction and improved sparrow algorithm. The proposed model extracts low- and medium-frequency components and removes high-frequency signals through secondary decomposition and reconstruction, and predicts through the combined model of FASSA-BiGRU.

[0013] 2. The proposed secondary decomposition and reconstruction denoising strategy achieves a more refined decomposition of load data and effectively removes high-frequency noise in the data. Through comparative analysis, this strategy greatly improves the accuracy of subsequent prediction models by extracting effective information from the data.

[0014] 3. By improving the sparrow algorithm (FASSA), hyperparameter optimization is performed for different frequency components after reconstruction, and a BiGRU-based model is established to predict small hydropower load. By comparing with multiple groups of models, it is proved that FASSA-BiGRU can effectively complete the task of small hydropower load prediction, while ensuring accuracy and having extremely fast speed, demonstrating the ability of FASSA-BiGRU in small hydropower load prediction.

[0015] 4. Through comparative studies of multiple groups of prediction models, the MAPE, RMSE and MAE prediction errors of the model proposed in the present invention are all smaller than those of the comparison model, which proves that this method has superior prediction accuracy in small hydropower load prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0017] Figure 1 It is a secondary decomposition process based on VMD and ICEEMDAN;

[0018] Figure 2 It is a calculation flow chart based on permutation entropy;

[0019] Figure 3 This is the network training flow chart based on FASSA-BiGRU;

[0020] Figure 4 Compare the prediction results of different decomposition and reconstruction strategies;

[0021] Figure 5 is the fitness curve of FASSA, SSA and PSO;

[0022] Figure 6 Comparison of prediction results between FASSA-BiGRU and BiGRU. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0024] Combination Figure 1 to Figure 6 The specific embodiments of the present invention are described in further detail.

[0025] As a clean energy, hydropower is widely distributed in central China and southwest my country, and most small hydropower stations are run-of-river hydropower stations with small installed capacity and poor regulation ability. Run-of-river small hydropower stations are generally located in small watersheds, and their output depends entirely on the natural flow of the river and the rainfall in the region, showing strong randomness and uncertainty, which seriously affects local load forecasts and provincial grid power supply load forecasts. Therefore, accurate prediction of the output of run-of-river small hydropower is beneficial to the formulation of power generation plans and to the promotion of clean energy consumption. To this end, the present invention first decomposes the small hydropower load and extracts as many characteristic components as possible from the load. First, the variational mode decomposition (VMD) is used to decompose the load data once to obtain the mode components and residual sequence; at the same time, considering the influence of the residual sequence on the prediction result, the residual is decomposed twice by the improved adaptive noise complete set empirical mode decomposition (ICEEMDAN). Secondly, the permutation entropy of all components after decomposition is calculated, and the low and medium frequency components are screened out and reconstructed according to the set threshold. The improved sparrow algorithm is used to set network parameters for components of different frequencies, and a BiGRU network load prediction model is built. Finally, the predicted values ​​of each component are superimposed to obtain the final prediction result.

[0026] S1. Data analysis and processing:

[0027] VMD can better preserve the detailed features of the original data. At the same time, when the parameters are selected appropriately, it can avoid mode aliasing, which is something that other EMD algorithms cannot guarantee. ICEEMDAN has high computational efficiency and real-time performance, and is suitable for processing large-scale data. VMD itself has certain limitations: due to parameter settings, numerical errors and other reasons, the sequences before and after VMD decomposition are inconsistent, that is, the sum of all modes is not equal to the original sequence. The difference between the original sequence and the sum of all modal components is the VMD residual. These residuals may contain some important information that is not captured by VMD, thus affecting the prediction results. At the same time, VMD residuals are often a time series with extremely high complexity. If the important information contained in the residuals is to be used, it needs to be decomposed twice. This method uses ICEEMDAN to complete this task. The decomposition accuracy is improved through secondary decomposition, and the adaptability of the decomposition algorithm ensures the flexibility of decomposition, so that it can be applied to different types of signals and noise levels. Finally, the permutation entropy is introduced to extract and reconstruct the mid- and low-frequency components. While removing the high-frequency noise in the data, it also solves the problem of too many components after secondary decomposition.

[0028] The secondary decomposition process based on VMD and ICEEMDAN is as follows Figure 1 As shown, the specific steps include:

[0029] S1.1. Use variational modal decomposition to decompose the load data once to obtain modal components and residual sequences.

[0030] In variational mode decomposition, the difference between the original sequence and the sum of all modal components is the VMD residual, and the formula is:

[0031]

[0032] Where R(t) is the VMD residual, X(t) is the original data, and u i (t) is the i-th mode, and K is the preset number of modes.

[0033] Variational mode decomposition includes the following steps:

[0034] S1.1.1. Construct variational problem: If the original signal x is decomposed into K components, the decomposition sequence is guaranteed to be a modal component with a finite bandwidth and a center frequency, and the sum of the estimated bandwidths of each mode is minimized. The constraint condition is that the sum of all modes is equal to the original signal. Then the constrained variational model of variational modal decomposition is as follows:

[0035]

[0036]

[0037] In the formula, u K={i1,i2,…,i k} and ω K ={f1,f2,…,f K} are the mode sets and their center frequencies, is the partial derivative with respect to t, δ(t) is the impulse function, st represents the constraint condition, Represents the square of the gradient norm.

[0038] S1.1.2. Solve the variational problem: Solve the above constrained optimization problem, transform the constrained variational problem into an unconstrained variational problem, take advantage of the quadratic penalty term and the Lagrangian multiplier method, and introduce the augmented Lagrangian function, which is expressed as:

[0039] L({i1,i2,…,i k}, {f1,f2,…,f k},λ(t))=; (4)

[0040]

[0041] In the formula, α is the penalty factor, λ(t) is the Lagrangian multiplier, and L represents the augmented Lagrangian function. is the partial derivative with respect to t, δ(t) is the impulse function, Represents the square of the gradient norm.

[0042] Use the alternating direction multiplier method to solve the above problem and update u K and K , where u K ={i1,i2,…,i k} and ω K ={f1,f2,…,f K}, the iteration formula is as follows:

[0043]

[0044] Among them, τ represents the noise tolerance, α is the penalty factor, λ(t) is the Lagrange multiplier, and f(t) represents the value of the variable at time t. and They are obtained by Fourier transform of f(t) and λ(t) respectively, ω represents the frequency variable, and its value range is [-∞,+∞].

[0045] Iterate until:

[0046]

[0047] Among them, ε is the given accuracy.

[0048] S1.2. The residual is decomposed twice by adopting the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN).

[0049] ICEEMDAN is used to perform secondary decomposition on the residual, and the original signal of the VMD residual is defined as y, w i is Gaussian white noise, E K (·) is the K-order modal component of the signal after empirical mode decomposition, N(·) is the local mean of the generated signal, and the ICEEMDAN algorithm includes the following steps:

[0050] S1.2.1. Add Gaussian white noise to the original signal y to obtain a new sequence y i , the first set of residuals R1 and the first modal component d1:

[0051] y i =y+β0E1(w i ); (5)

[0052] R1=N(y i ); (6)

[0053] d1=y-R1; (7)

[0054] S1.2.2. Continue to add Gaussian white noise to obtain the second set of residuals R2 and the second modal component d2:

[0055] R2=N(R1+β1E2(w i )); (8)

[0056] d2=R1-R2; (9)

[0057] S1.2.3, and so on, we can get the Kth residual R K and the Kth modal component d K :

[0058] R K =N(R K-1 +β K-1 E(w i )); (10)

[0059] d K =R K-1 -R K (11).

[0060] S1.3, if Figure 2 As shown, the permutation entropy of all modal components after secondary decomposition by S1.2 is calculated, and the low and medium frequency components are screened out and reconstructed according to the set threshold;

[0061] The calculation of permutation entropy of modal components includes the following steps:

[0062] S1.3.1. Set an embedding dimension m and a delay time τ, then reconstruct the original sequence and define each subsequence as:

[0063] X(i)=u(i),u(i+τ),…,u(i+(m-1)τ);

[0064] The acquired time series small hydropower historical load data is reconstructed in phase space to obtain the reconstructed moment Y:

[0065]

[0066] Where m is the embedding dimension; τ is the delay time; K = N-(m-1)t, t = 1, 2, 3, ..., N;

[0067] S1.3.2. Each row in the matrix Y is a reconstruction component, and the column index of the element is j1, j2, …, j m It means that by comparing the numerical values ​​of the elements, each reconstructed component is rearranged in ascending order. If the numerical values ​​of two elements are the same, the column indexes of the two elements are compared and arranged in ascending order. After reconstruction, a set of symbol sequences S(l) can be obtained:

[0068] S(l)=(j1,j2,j3,…,j m ); (13)

[0069] Where l = 1, 2, 3, ..., K, and K ≤ m!, and m column indices have m! different permutations;

[0070] S1.3.3. The number of times each symbol sequence appears is divided by m! to obtain the probability P of each sequence appearing. K , permutation entropy H pe The calculation formula is:

[0071]

[0072] S1.3.4. The permutation entropy value of the time series is calculated by formula (13), and the maximum value is ln(m!). The permutation entropy value of the time series is normalized to obtain:

[0073]

[0074] The permutation entropy values ​​of each group of time series are normalized and then used for permutation entropy sorting and screening.

[0075] S2. Build a bidirectional gated recurrent unit (BiGRU) network load prediction model, use the firefly-based improved sparrow algorithm (FA-SSA) to optimize the parameters of the BiGRU network load prediction model, and use the optimized BiGRU network load prediction model to predict the low and medium frequency components screened out in step S1.3. The predicted values ​​of the low and medium frequencies are superimposed to obtain the final predicted value.

[0076] FA-SSA is used to optimize the parameters of the BiGRU network load forecasting model. Figure 3 As shown, the following steps are included:

[0077] S2.1. Parameter initialization: Determine the parameters of FA-SSA and the network structure of BiGRU respectively, initialize the sparrow population through Cubic mapping, and optimize the number of hidden layer neurons and learning rate of the BiGRU network load forecasting model.

[0078] The Cubic mapping formula is:

[0079]

[0080] S2.2, select the objective function: take the mean percentage error (MAPE) of the initial BiGRU network load forecasting model as the fitness function of FA-SSA, calculate the fitness value of each sparrow, and determine the number of three types of searchers, joiners and scouts;

[0081] The first s with the largest fitness values ​​are selected as searchers, and the rest are followers and scouts. The proportion of searchers is 20%, the proportion of followers is 70%, and the rest are scouts.

[0082] S2.3, Optimization process: While the searcher, joiner and scout are updating their positions, the firefly perturbation strategy is used to update the position of the mutant sparrow, recalculate the fitness value of each individual and update the optimal position again to obtain the best position of each individual. When the set maximum number of iterations is reached, the optimal hyperparameter value is obtained.

[0083] S2.4, BiGRU network load forecasting model training: The optimal hyperparameter solution obtained by the FASSA algorithm is substituted into the BiGRU network load forecasting model, and the training test is re-performed to determine the final BiGRU network load forecasting model.

[0084] In this embodiment, the number of iterations and the number of sparrow populations are both set to 100, and the warning value is 0.8. After steps S2.1 to S2.4, the important parameters of the BiGRU network, namely the number of hidden layer neurons and the learning rate, are optimized according to the FASSA algorithm. Finally, the optimal parameters of the BiGRU obtained by FASSA optimization are {[23,28], 0.0024}.

[0085] Example:

[0086] In order to analyze the impact of different decomposition and reconstruction strategies on prediction accuracy, the secondary decomposition and reconstruction proposed in the present invention are compared with direct prediction and one-time decomposition and reconstruction using the FASSA-BiGRU algorithm. The direct prediction does not involve the decomposition and reconstruction process; the one-time decomposition and reconstruction uses the VMD and ICEEMDAN algorithms respectively, and uses the same reconstruction method. At the same time, in order to verify the effectiveness of removing high-frequency components in improving prediction accuracy, a comparison is set up: all components that retain high-frequency components and all components that do not retain high-frequency components are reconstructed and input into the FASSA-BiLSTM prediction respectively. The prediction results are compared as follows: Figure 4 The error comparison is shown in Table 1.

[0087] Table 1 Comparison of model errors

[0088]

[0089] From Table 1 and Figure 4 It can be seen that the error during direct prediction is the largest, with a MAPE error of 50.7%; the error of primary decomposition and reconstruction is second, among which the MAPE error when VMD is used as the decomposition algorithm is 37.66%, and the MAPE error when ICEEMDAN is used as the decomposition algorithm is 34.42%; the accuracy of secondary decomposition and reconstruction is the best, among which the MAPE error when the high-frequency components are not eliminated is 19.31%, and the comprehensive MAPE error when the high-frequency components are eliminated is 11.21%.

[0090] At the same time, in order to verify the superiority of FASSA parameter setting and the improvement of prediction accuracy, the sparrow algorithm SSA and particle swarm algorithm (PSO) are selected for fitness comparison, such as Figure 5 As shown; and then verified by comparing the prediction results of FASSA-BiGRU and BiGRU, as shown Figure 6 And as shown in Table 2.

[0091] Figure 5 It can be seen that among the three optimization algorithms, the convergence speed and convergence accuracy of the improved FASSA algorithm are improved compared with the SSA algorithm and other optimization algorithms, and convergence is achieved after 20 iterations.

[0092] Table 2 Error comparison

[0093]

[0094] From Table 2 and Figure 6It can be seen that the FASSA algorithm can not only improve the prediction efficiency, but also has a good effect on improving the prediction accuracy. The RMSE, MAPE and MAE of FASSA-BiGRU are 0.36, 11.21% and 0.30 respectively, while the RMSE, MAPE and MAE of BiGRU are 0.40, 15.42% and 0.35 respectively. It can be seen that the optimal hyperparameters are important for the load forecasting model.

Claims

1. A small hydropower load forecasting method based on quadratic mode decomposition and improved sparrow algorithm, characterized in that: The following steps are involved: S1. Data analysis and processing, including the following steps: S1.

1. Decompose the load data once using variational modal decomposition to obtain modal components and residual sequences; S1.2, performing secondary decomposition of the residual by adopting improved adaptive noise complete set empirical mode decomposition; S1.3, calculate the permutation entropy of all modal components after the secondary decomposition in step S1.2, and select the low and medium frequency components according to the set threshold and reconstruct them; S2. Build a bidirectional gated cyclic unit network load prediction model, use the sparrow algorithm improved by Firefly to optimize the parameters of the bidirectional gated cyclic unit network load prediction model, and use the optimized bidirectional gated cyclic unit network load prediction model to predict the low and medium frequency components screened out in step S1.

3. The predicted values ​​of low frequency and medium frequency are superimposed to obtain the final predicted value.

2. The method for predicting small hydropower load based on quadratic mode decomposition and improved sparrow algorithm according to claim 1 is characterized in that: In step S1.1, in the variational mode decomposition, the difference between the original sequence and the sum of all modal components is the VMD residual, and the formula is as follows: Where R(t) is the VMD residual, X(t) is the original data, and u i (t) is the i-th mode, and K is the preset number of modes.

3. The method for small hydropower load forecasting based on quadratic mode decomposition and improved sparrow algorithm according to claim 1 is characterized in that: In step S1.1, the variational mode decomposition includes the following steps: S1.1.

1. Construct variational problem: If the original signal x is decomposed into K components, the constrained variational model of variational mode decomposition is as follows: In the formula, let u K ={i1,i2,…,i k } and ω K ={f1,f2,…,f K } are the mode sets and their center frequencies, represents the partial derivative with respect to t, δ(t) is the impulse function, st represents the constraint condition, Represents the square of the gradient norm; S1.1.

2. Solve the variational problem: L({i1,i2,…,i K },{f1,f2,…,f K },λ(t))=; (4) Where α is the penalty factor, λ(t) is the Lagrangian multiplier, and L represents the augmented Lagrangian function.

4. The method for predicting small hydropower load based on quadratic mode decomposition and improved sparrow algorithm according to claim 1 is characterized in that: In step S1.2, the residual is decomposed twice using the improved adaptive noise complete set empirical mode decomposition, and the original signal of the VMD residual is defined as y, w i is Gaussian white noise, E K (·) is the K-order modal component of the signal after empirical mode decomposition, N(·) is the local mean of the generated signal, and the improved adaptive noise complete set empirical mode decomposition algorithm includes the following steps: S1.2.

1. Add Gaussian white noise to the original signal y to obtain a new sequence y i , the first set of residuals R1 and the first modal component d1: y i =y+β0E1(w i ); (5) R1=N(y i ); (6) d1=y-R1; (7) S1.2.

2. Continue to add Gaussian white noise to obtain the second set of residuals R2 and the second modal component d2: R2=N(R1+β1E2(w i )); (8) d2=R1-R2; (9) S1.2.3, and so on, we can get the Kth residual R K and the Kth modal component d K : R K N(R K-1 +β K-1 E K (w i ))(10) d K =R K-1 -R K (11)。 5. The method for small hydropower load forecasting based on quadratic mode decomposition and improved sparrow algorithm according to claim 1 is characterized in that: In step S1.3, the calculation of the permutation entropy of the modal component includes the following steps: S1.3.

1. Reconstruct the phase space of the acquired time series small hydropower historical load data to obtain the reconstructed moment Y: Where m is the embedding dimension; τ is the delay time; K = N-(m-1)t, t = 1, 2, 3, ..., N; S1.3.

2. Each row in the matrix Y is a reconstruction component, and the column index of the element is j1, j2, …, j m It means that by comparing the numerical values ​​of the elements, each reconstructed component is rearranged in ascending order. If the numerical values ​​of two elements are the same, the column indexes of the two elements are compared and arranged in ascending order. After reconstruction, a set of symbol sequences S(l) can be obtained: S(l)=(j1,j2,j3,…,j m ); (13) Where l = 1, 2, 3, ..., K, and K ≤ m!, and m column indices have m! different permutations; S1.3.

3. The number of times each symbol sequence appears is divided by m! to obtain the probability P of each sequence appearing. K , permutation entropy H pe The calculation formula is: S1.3.

4. The permutation entropy value of the time series is calculated by formula (13), and the maximum value is ln(m!). The permutation entropy value of the time series is normalized to obtain: The permutation entropy values ​​of each group of time series are normalized and then used for permutation entropy sorting and screening.

6. The method for small hydropower load forecasting based on quadratic mode decomposition and improved sparrow algorithm according to claim 1 is characterized in that: In step S2, the parameters of the bidirectional gated cyclic unit network load forecasting model are optimized using the sparrow algorithm improved based on Firefly, including the following steps: S2.1, parameter initialization: determine the parameters of the firefly-based improved sparrow algorithm and the bidirectional gated recurrent unit network load prediction model respectively, initialize the sparrow population through Cubic mapping, and optimize the number of hidden layer neurons and learning rate of the bidirectional gated recurrent unit network load prediction model. The Cubic mapping formula is: S2.2, select the objective function: take the average percentage error of the initial bidirectional gated cyclic unit network load forecasting model as the fitness function of the sparrow algorithm based on the firefly improvement, calculate the fitness value of each sparrow, and determine the number of three types of searchers, joiners and scouts; S2.3, Optimization process: While the searcher, joiner and scout are updating their positions, the firefly perturbation strategy is used to update the position of the mutant sparrow, recalculate the fitness value of each individual and update the optimal position again to obtain the best position of each individual. When the set maximum number of iterations is reached, the optimal hyperparameter value is obtained; S2.4, Bidirectional gated recurrent unit network load prediction model training: Based on the Firefly-improved Sparrow algorithm, the optimal hyperparameter solution is obtained and substituted into the bidirectional gated recurrent unit network load prediction model, and retraining and testing are performed to determine the final bidirectional gated recurrent unit network load prediction model.

7. The method for predicting small hydropower load based on quadratic mode decomposition and improved sparrow algorithm according to claim 6 is characterized in that: In step S2.2, the first s with the largest fitness values ​​are selected as searchers, and the rest are followers and scouts. The proportion of searchers is 20%, the proportion of followers is 70%, and the rest are scouts.

Images