Novel electric power system equivalent inertia short-term prediction method adapting to seasonal change
Through the modal decomposition and feature fusion of ICEEMDAN and RCMFE algorithms, combined with the prediction model of BiLSTM neural network, the seasonal changes and complex data processing problems in the equivalent inertia prediction of the power system are solved, efficient and accurate short-term prediction is achieved, and the improvement of the Reliability and stability of power grid operation.
Patent Information
- Application Number
- CN202410779654.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-17
- Publication Date
- 2024-11-01
AI Technical Summary
The existing power system equivalent inertia prediction methods cannot effectively predict the future change of system inertia, and a single neural network model has limitations when processing complex data, making it difficult to effectively integrate seasonal characteristics.
Modal decomposition is performed using ICEEMDAN algorithm, feature fusion is combined with RCMFE algorithm, characteristic factors affecting the change of inertia are screened out, and short-term prediction models are constructed through BiLSTM neural network, taking into account the inherent laws of historical and future data, and Bayesian optimization of hyperparameters is used to optimize To improve prediction accuracy.
It realizes efficient and accurate short-term prediction of the inertia of the power system, improves the prediction accuracy and stability, helps the power grid dispatching department to formulate control measures, and enhances the understanding and judgment of the system status.
Smart Images

Figure CN118889366A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of power system dynamic analysis, and in particular to a novel power system equivalent inertia short-term prediction method adapting to seasonal changes. Background Art
[0002] With the rapid development of power systems and the large-scale access of renewable energy, the dynamic characteristics and operating environment of power grids have become increasingly complex. System inertia, as an important parameter to measure the dynamic response capability of the system, is crucial to ensuring the stability and reliability of the power grid. Existing inertia assessment work can only show the inertia level of the current operation or after the system is disturbed, and cannot predict its future changes. Predicting the equivalent inertia of the system helps to identify the time period when it is less than the critical level in advance, so as to make corresponding countermeasures in advance.
[0003] In recent years, neural network models have been widely used in power system prediction tasks due to their powerful nonlinear mapping capabilities and self-learning characteristics. However, neural network models with a single algorithm still have limitations when processing complex data, such as being prone to falling into local optimality and being sensitive to noise. In addition, the seasonal characteristics of power system data place higher demands on the design of prediction models, requiring the model to be able to effectively integrate and utilize these seasonal information. Therefore, a short-term prediction method for equivalent inertia of power systems is urgently needed. Summary of the invention
[0004] To solve the above problems, the present invention proposes a new short-term prediction method for equivalent inertia of power system that adapts to seasonal changes. This method introduces modal decomposition technology to decompose power system data into multiple intrinsic mode functions according to seasons, so as to better capture the seasonal characteristics and dynamic change characteristics of the data; at the same time, the advantages of multiple algorithms are comprehensively considered to balance the prediction accuracy and prediction time. This method makes full use of seasonal change information, ensures the high efficiency and accuracy of short-term prediction of power system inertia, and is conducive to the grid dispatching department to formulate corresponding control measures.
[0005] The technical solution adopted by the present invention is:
[0006] A novel short-term prediction method for equivalent inertia of power system adapting to seasonal changes includes the following steps:
[0007] Step 1: Divide the annual inertia of the power grid into four seasons and decompose it using the ICEEMDAN algorithm;
[0008] Step 2: Use the refined composite multi-scale fuzzy entropy RCMFE algorithm to aggregate the subsequences divided and decomposed according to the four seasons;
[0009] Step 3: Use the mRMR feature extraction algorithm to screen out the most relevant characteristic factors affecting the change of inertia in four seasons as the input data for system inertia prediction;
[0010] Step 4: Use the Min-Max normalization method to scale the input data required for system inertia prediction in step 3, maintain their relative relationship, and eliminate possible scale differences between different features;
[0011] Step 5: Construct a short-term prediction model for the power system inertia, predict the different components of inertia in different seasons, accumulate the final prediction results, and realize the power system inertia prediction.
[0012] In the step 1, according to the inherent time scale of the power grid inertia sequence, the power system data is decomposed into multiple intrinsic mode functions and residuals with frequencies from high to low according to the season, and each intrinsic mode function can represent the characteristic components of the power grid inertia at different time scales;
[0013] The decomposition steps of the ICEEMDAN algorithm are shown in formula (1):
[0014]
[0015] In the formula, x i (t) is the noise-added signal; x(t) is the original inertia sequence; α0 represents the initial value of the amplitude coefficient of the added noise;
[0016] E1(η i (t)) represents the first-order mode obtained by the EMD algorithm; η i (t) indicates adding the i-th white noise;
[0017] r k (t) represents the k-th order residual component; r k-1 (t) represents the k-1th order residual component;
[0018] α k-1 =β k-1 std(x(t)) / std(E k (w i (t))) is the standard deviation of white noise, which is used to remove noise; β k-1 is the scale factor; std is the standard deviation; w i (t) represents the white noise sequence added in the i-th iteration; E k (w i (t)) represents the result of the collective operation on all k added white noises;
[0019] E k (·) indicates the k-th mode obtained by the EMD algorithm; <·> indicates the mean value calculation; M(·) indicates the envelope calculation;
[0020] IMF k(t) represents the kth order modal component.
[0021] The ICEEMDAN algorithm adds white noise η i (t) and local average to achieve the continuity of frequency between adjacent scales, so as to adaptively decompose the non-stationary sequence x(t) into a finite number of intrinsic mode function components with different frequency characteristics and a residual component, so as to achieve the purpose of weakening the nonlinear characteristics of inertia and refining the analysis of inertia. Taking autumn as an example, the decomposition results of the original data of system inertia are as follows Figure 1 shown.
[0022] In step 2, a fine composite multi-scale fuzzy entropy RCMFE algorithm is used to calculate the RCMFE value of the system inertia of a certain season; the RCMFE algorithm is an improved version of the multi-scale fuzzy entropy (MFE) and the composite multi-scale fuzzy entropy (CMFE). The improvement is that it can analyze the complexity of the data more accurately at multiple scales. By calculating all coarse-grained sequences under the scale factor τ, then summing them, and then calculating the fuzzy entropy, the possibility of undefined entropy in CMFE is effectively reduced. The calculation formula for the system inertia RCMFE for a certain season is shown in formula (2):
[0023]
[0024] Where x={x1,x2,…,x N} is the inertia power time series; x1, x2,…, x N Indicates the value of each signal;
[0025] is a coarse-grained sequence; represents the coarse-grained sequence at different scales; 1≤k≤τ, 1≤j≤N / τ; k and j are index variables; τ is the scale factor; N represents the number of fuzzy entropies;
[0026] m is the embedding dimension; n is the gradient of the fuzzy function; r is the similarity tolerance;
[0027] represents the number of matching vector pairs calculated for the k-th coarse-grained sequence under the embedding dimension m+1 and the scaling factor τ;
[0028] represents the number of matching vector pairs calculated for the k-th coarse-grained sequence under the embedding dimension m and the scaling factor τ;
[0029] The subsequences divided and decomposed according to the four seasons are aggregated; the RCMFE values of each component obtained by decomposing the power system inertia are as follows: Figure 2 As shown, for different seasons, based on Figure 2The dividing line in the figure divides CIMF components with similar complexity into the same category and merges them into fewer IMF components. The aggregation result is as follows: Figure 3 shown.
[0030] In step 3, the correlation analysis of system inertia and climate and operation data is performed and features are further extracted. In addition to the historical inertia of each component, wind speed, light intensity I, average temperature t mean , precipitation PF, net load P net , Synchronous power output P syn , Asynchronous power output P non 、System frequency f sys Eight influencing factors were considered as input features;
[0031] Since the system inertia and influencing factors are mostly nonlinear, the mRMR feature extraction algorithm is used to select the system inertia power, traditional new energy and common new energy active output in spring, summer, autumn and winter for feature extraction, and screen out the most relevant characteristic factors affecting the inertia changes in the four seasons as the input data for system inertia prediction.
[0032] In step 3, the feature screening calculation process of the mRMR feature extraction algorithm is shown in formula (3):
[0033]
[0034] Where: X and Y represent two characteristic variables respectively; p(x) and p(y) represent the marginal probability functions of the two characteristic variables respectively; p(x, y) represents the joint probability distribution of the two characteristic variables; x and y represent the specific values of the two characteristic variables respectively;
[0035] maxD(S,P) represents the maximum correlation between features and categories; D is used to measure the correlation between features and categories; S represents the feature set; P represents the target category; D(S,P) represents a measure of the correlation between features and categories; f i represents the i-th feature in the set; I(f i ,P) represents the mutual information between feature i and the target category. Mutual information is used to measure the correlation between variables;
[0036] minR(S) represents the minimum redundancy between features; I(f i ,f j ) represents the mutual information between feature i and feature j, which is used to measure the correlation between variables; f j represents the jth feature in the set;
[0037] maxJ(D,R) represents the evaluation function of the feature selection algorithm; J represents the evaluation function itself; R represents the redundancy between features; J(D,R) represents a specific evaluation function;
[0038] Maximizeφ means maximizing the difference between single variables and correlation and redundancy; I(f i ,c) represents mutual information, which is used to measure the degree of mutual dependence between two variables; c represents the target variable; S t-1 represents the feature set at step t-1; t represents a specific time point in the time series;
[0039] From the collection SS t-1 To select the t-th feature in , we need to determine the added t-th feature by maximizing the difference between univariate correlation and redundancy.
[0040] In step 4, the Min-Max normalization method is used to scale the original data, and its expression is shown in formula (4):
[0041]
[0042] In the formula: x' is the normalized data; x is the original data; x max 、x min are the maximum and minimum values in the original data.
[0043] In step 5, the system inertia sequence has obvious positive and negative regularity. When predicting the system equivalent inertia, the influence of historical inertia and future inertia on the prediction accuracy should be considered at the same time;
[0044] A short-term prediction model for power system inertia is constructed based on the BiLSTM neural network. The overall framework of the prediction model is as follows: Figure 4 As shown in the figure. The BiLSTM neural network consists of two LSTM neural networks, one forward and one reverse, and can perform bidirectional time series feature extraction on the input data. The BiLSTM neural network considers the inherent rules of the forward and backward data at the same time, and makes predictions from the two directions of history and the future. It uses two independent hidden layers to process the data from the forward and reverse directions respectively, and then uses the outputs of the forward and reverse hidden layers as the inputs of the output layer, thereby neutralizing the phase error.
[0045] In step 5, the hyperparameter optimization process of the BiLSTM neural network based on Bayesian optimization includes the following steps:
[0046] S5.1: Determine the possible value range of the hyperparameters of the BiLSTM neural network;
[0047] S5.2: Use historical data to build a suitable probability model, namely the Gaussian process, to predict the output of new data points based on existing data points. In the process of hyperparameter optimization, the Gaussian process is used to predict the impact of different hyperparameter combinations on model performance;
[0048] S5.3: Select a set of hyperparameters from the hyperparameter interval to train the BiLSTM neural network model;
[0049] S5.4: Train the BiLSTM neural network and evaluate its performance on the validation set.
[0050] S5.5: If a better parameter is found or the maximum number of iterations is reached, stop, otherwise use a new proxy model, which is also referred to as a Gaussian process. In each round of iteration, based on the existing evaluation results, the Gaussian process will update its predictions, which can more accurately guide the next round of hyperparameter selection and continue to find a better hyperparameter combination;
[0051] S5.6: Extract the best performing hyperparameter combination from the iterative process.
[0052] The BiLSTM hyperparameter optimization process based on Bayesian optimization is as follows Figure 5 described.
[0053] In order to verify the prediction performance of the model proposed in this invention, the mean absolute error E MAE , Root mean square error E RMSE , mean absolute percentage error E MAPE As an evaluation index, to verify the accuracy of the model in predicting the system inertia. MAE Reflects the actual error level of the model; E RMSE Indicates the impact of extreme outliers in the prediction results; E MAPE It represents the average value of the absolute percentage error and the inertia size, which can reflect the overall prediction effect of the prediction model. The evaluation indicators are shown in formula (5).
[0054]
[0055] in: y i are the predicted value and the true value respectively. The smaller the MAE, RMSE, and MAPE, the better the model fitting effect.
[0056] The present invention provides a novel short-term prediction method for equivalent inertia of a power system that adapts to seasonal changes, and the technical effects are as follows:
[0057] 1) The present invention utilizes the ICEEMDAN algorithm for modal decomposition and the RCMFE algorithm for feature fusion to accurately capture the seasonal variation of the power system inertia, thereby improving the accuracy of the prediction.
[0058] 2) The present invention utilizes the mRMR feature extraction algorithm to effectively screen out the characteristic factors most relevant to the seasonal inertia changes, thereby improving the quality of input features while reducing the prediction burden of the model.
[0059] 3) The present invention uses BiLSTM neural network to consider the inherent laws of forward and backward data at the same time, and makes predictions from two directions, namely, history and future, and uses two independent hidden layers to process data from forward and backward directions respectively. This structure can perform bidirectional time series feature extraction on input data, further improving the globality and integrity of time series feature extraction.
[0060] 4) The combined prediction model constructed by the present invention significantly improves the prediction accuracy while reducing the later prediction tasks through a series of processes such as modal decomposition, complexity evaluation, sequence aggregation and hyperparameter optimization of the original data of power system inertia in each season, which helps power system operators understand and judge the system status. BRIEF DESCRIPTION OF THE DRAWINGS
[0061] Figure 1 This is the decomposition result of the original data of the autumn system inertia.
[0062] Figure 2 RCMFE value of each component obtained by decomposing the inertia of the power system.
[0063] Figure 3 It is the aggregated result of each component obtained by decomposing the inertia of the autumn system.
[0064] Figure 4 Schematic diagram of the overall framework of the prediction model.
[0065] Figure 5 This is a flowchart for BiLSTM hyperparameter optimization based on Bayesian optimization.
[0066] Figure 6 This is a comparison chart of power system inertia prediction and evaluation indicators.
[0067] Figure 7 This is a comparison chart of the power system inertia prediction effect.
[0068] Figure 8 This is a comparison chart of the actual and predicted values of the power system inertia. DETAILED DESCRIPTION
[0069] A seasonal equivalent inertia prediction method for power systems based on a multi-algorithm hybrid neural network model with modal decomposition and feature fusion is proposed. The method first uses the improved adaptive noise complete ensemble empirical mode decomposition (ICEEMDAN) method to decompose the inertia of the four seasons, and reconstructs the decomposed components of the four seasons into a new sequence based on the refined composite multi-scale fuzzy entropy (RCMFE) value in entropy theory. The minimum redundancy maximum relevance (mRMR) method is used to screen out a subset of features with high correlation with equivalent inertia, and the data preprocessing is combined with seasonal feature analysis to form a feature set of input variables of the prediction model. Secondly, the bidirectional long short-term memory network (BiLSTM) model based on the Bayesian optimization algorithm is used to predict different components of inertia in different seasons, and the final prediction results are accumulated to realize the inertia prediction of the new power system.
[0070] A novel short-term prediction method for equivalent inertia of power system adapting to seasonal changes includes the following steps:
[0071] Step 1: According to the inherent time scale of the power grid inertia series, the power system data is decomposed into multiple intrinsic mode functions and residuals with high to low frequencies according to season. Each intrinsic mode function can represent the characteristic component of the power grid inertia at different time scales.
[0072] The decomposition steps of the ICEEMDAN algorithm are shown in formula (1):
[0073]
[0074] In the formula, x i (t) is the noise-added signal; x(t) is the original inertia sequence; α0 represents the initial value of the amplitude coefficient of the added noise;
[0075] E1(η i (t)) represents the first-order mode obtained by the EMD algorithm; η i (t) indicates adding the i-th white noise;
[0076] r k (t) represents the k-th order residual component; r k-1 (t) represents the k-1th order residual component;
[0077] α k-1 =β k-1 std(x(t)) / std(E k (w i (t))) is the standard deviation of white noise, which is used to remove noise; β k-1 is the scale factor; std is the standard deviation; w i (t) represents the white noise sequence added in the i-th iteration; E k (w i(t)) represents the result of the collective operation on all k added white noises;
[0078] E k (·) indicates the k-th mode obtained by the EMD algorithm; <·> indicates the mean value calculation; M(·) indicates the envelope calculation;
[0079] IMF k (t) represents the kth order modal component.
[0080] The ICEEMDAN algorithm adds white noise η i (t) and local average to achieve the continuity of frequency between adjacent scales, so as to adaptively decompose the non-stationary sequence x(t) into a finite number of intrinsic mode function components with different frequency characteristics and a residual component, so as to achieve the purpose of weakening the nonlinear characteristics of inertia and refining the analysis of inertia. Taking autumn as an example, the decomposition results of the original data of system inertia are as follows Figure 1 Shown
[0081] Step 2: Use the refined composite multiscale fuzzy entropy RCMFE algorithm to calculate the RCMFE value of the system inertia in a certain season; the RCMFE algorithm is an improved version of the multiscale fuzzy entropy (MFE) and the composite multiscale fuzzy entropy (CMFE). The improvement is that it can analyze the complexity of data more accurately at multiple scales. By calculating all coarse-grained sequences under the scale factor τ, then summing them up and calculating the fuzzy entropy, the possibility of undefined entropy in CMFE is effectively reduced. The calculation formula for the system inertia RCMFE for a certain season is shown in formula (2).
[0082]
[0083] Where x={x1,x2,…,x N} is the inertia power time series; x1, x2,…, x N Represent each signal value respectively;
[0084] is a coarse-grained sequence; Respectively represent the coarse-grained sequence at different scales; 1≤k≤τ, 1≤j≤N / τ; k and j are index variables; τ is the scale factor; N represents the number of fuzzy entropy;
[0085] m is the embedding dimension; n is the gradient of the fuzzy function; r is the similarity tolerance;
[0086] represents the number of matching vector pairs calculated for the k-th coarse-grained sequence under the embedding dimension m+1 and the scaling factor τ;
[0087] represents the number of matching vector pairs calculated for the k-th coarse-grained sequence under the embedding dimension m and the scaling factor τ;
[0088] The subsequences divided and decomposed according to the four seasons are aggregated; the RCMFE values of each component obtained by decomposing the power system inertia are as follows: Figure 2 As shown, it can be seen that the RCMFE value of each CIMF component in each season is different, which means that the data complexity of each component is different.
[0089] For different seasons, based on Figure 2 The dividing line in the figure divides CIMF components with similar complexity into the same category and merges them into fewer IMF components. The aggregation result is as follows: Figure 3 As shown. Figure 3 It can be seen that the autumn inertia CIMF components are combined to form four IMF components, which are more stable than the original inertia data and have more obvious regularity.
[0090] Step 3: Analyze the correlation between system inertia and climate and operation data and further extract features. In addition to the historical inertia of each component, we also select wind speed, light intensity I, average temperature t mean , precipitation PF, net load P net , Synchronous power output P syn , Asynchronous power output P non 、System frequency f sys Eight influencing factors were considered as input features;
[0091] Since the system inertia and influencing factors are mostly nonlinear, the mRMR feature extraction algorithm is used to select the system inertia power, traditional new energy and common new energy active output in spring, summer, autumn and winter for feature extraction, and screen out the most relevant characteristic factors affecting the inertia changes in the four seasons as the input data for system inertia prediction.
[0092] In step 3, the feature screening calculation process of the mRMR feature extraction algorithm is shown in formula (3):
[0093]
[0094] Where: X and Y represent two characteristic variables respectively; p(x) and p(y) represent the marginal probability functions of the two characteristic variables respectively; p(x, y) represents the joint probability distribution of the two characteristic variables; x and y represent the specific values of the two characteristic variables respectively;
[0095] maxD(S,P) represents the maximum correlation between features and categories; D is used to measure the correlation between features and categories; S represents the feature set; P represents the target category; D(S,P) represents a measure of the correlation between features and categories; fi represents the i-th feature in the set; I(f i ,P) represents the mutual information between feature i and the target category. Mutual information is used to measure the correlation between variables;
[0096] minR(S) represents the minimum redundancy between features; I(f i ,f j ) represents the mutual information between feature i and feature j, which is used to measure the correlation between variables; f j represents the jth feature in the set;
[0097] maxJ(D,R) represents the evaluation function of the feature selection algorithm; J represents the evaluation function itself; R represents the redundancy between features; J(D,R) represents a specific evaluation function;
[0098] I(f i ,c) represents mutual information, which is used to measure the degree of mutual dependence between two variables; c represents the target variable; S t-1 represents the feature set at step t-1; t represents a specific time point in the time series;
[0099] In the selection set SS t-1 The t-th feature is selected by maximizing the difference between the univariate relevance of the feature and its redundancy.
[0100] Step 4: Use the Min-Max normalization method to scale the original data. The expression is shown in formula (4):
[0101]
[0102] In the formula: x' is the normalized data; x is the original data; x max 、x min are the maximum and minimum values in the original data.
[0103] Step 5: The system inertia sequence has obvious positive and negative regularity. When predicting the system equivalent inertia, the influence of historical inertia and future inertia on the prediction accuracy should be considered at the same time;
[0104] A short-term prediction model for power system inertia is constructed based on the BiLSTM neural network. The overall framework of the prediction model is as follows: Figure 4As shown in the figure. The BiLSTM neural network consists of two LSTM neural networks, one forward and one backward, and can perform bidirectional time series feature extraction on the input data. The BiLSTM neural network considers the inherent rules of the forward and backward data at the same time, and makes predictions from the two directions of history and the future. It uses two independent hidden layers to process the data from the forward and backward directions respectively, and then uses the outputs of the forward and backward hidden layers as the inputs of the output layer, thereby neutralizing the phase error.
[0105] In step 5, the hyperparameter optimization process of the BiLSTM neural network based on Bayesian optimization includes the following steps:
[0106] S5.1: Determine the possible value range of the hyperparameters of the BiLSTM model;
[0107] S5.2: Use historical data to build appropriate probabilistic models to predict the performance of different hyperparameter combinations;
[0108] S5.3: Select a set of hyperparameters from the hyperparameter interval for model training;
[0109] S5.4: Train the BiLSTM model and evaluate its performance on the validation set;
[0110] S5.5: If a better parameter is found or the maximum number of iterations is reached, stop, otherwise use a new proxy model to continue looking for a better hyperparameter combination;
[0111] S5.6: Extract the best performing hyperparameter combination from the iterative process.
[0112] The BiLSTM hyperparameter optimization process based on Bayesian optimization is as follows Figure 5 described.
[0113] In order to verify the prediction performance of the model proposed in this invention, the mean absolute error E MAE , Root mean square error E RMSE , mean absolute percentage error E MAPE As an evaluation index, to verify the accuracy of the model in predicting the system inertia. MAE Reflects the actual error level of the model; E RMSE Indicates the impact of extreme outliers in the prediction results; E MAPE It represents the average value of the absolute percentage error and the inertia size, which can reflect the overall prediction effect of the prediction model. The evaluation indicators are shown in formula (5).
[0114]
[0115] in: y iare the predicted value and the true value respectively. The smaller the MAE, RMSE, and MAPE, the better the model fitting effect.
[0116] Step 6: The accuracy of the short-term prediction model of the power system established above is verified through typical cases.
[0117] A hybrid neural network model based on modal decomposition and feature fusion multi-algorithm is used to perform short-term prediction of system inertia for a regional power grid, based on the measured data from March 31, 2017 to February 28, 2018 in the region, with a time resolution of 15 minutes. The input features such as meteorology and system operation obtained by mRMR screening are shown in Table 1.
[0118] Taking autumn as an example, the decomposition results of the original inertia data are as follows: Figure 3 The RCMFE calculation results of each component in each season are shown in Figure 4 The aggregation results are shown as Figure 5 shown.
[0119] Three models, ICEEMDAN-mRMR-Bayes-BiLSTM (IMBB), mRMR-Bayes-BiLSTM (MBB), and ICEEMDAN-RCMFE-mRMR-SSA-BiLSTM (IRMSB), were selected for comparative analysis with the (ICEEMDAN-RCMFE-mRMR-Bayes-BiLSTM, IRMBB) model proposed in this invention.
[0120] Table 1 Input characteristics of inertia IMF components in each season
[0121]
[0122] Then, the statistical model predicts the time required, and uses the three indicators of MAPE, MAE, and RMSE to evaluate the model. The results are as follows Figure 6 Shown
[0123] To further verify the superiority of the model proposed in this invention, five additional comparison models are set up: ICEEMDAN-RCMFE-mRMR-BiLSTM (IRMB); ICEEMDAN-RCMFE-mRMR-Bayes-SVM (IRMBS); ICEEMDAN-RCMFE-mRMR-Bayes-XGBoost (IRMBX); ICEEMDAN-RCMFE-mRMR-Bayes-LSTM (IRMBL); and ICEEMDAN-RCMFE-mRMR-Bayes-GRU (IRMBG). The prediction result evaluation indicators are shown in Table 2.
[0124] Table 2 Prediction performance indicators of six models in different seasons
[0125]
[0126]
[0127] The results show that the IRMBB method outperforms other combination models in all evaluation indicators in the four seasons. Compared with the benchmark models in the field of deep learning such as LSTM and GRU, BiLSTM has a good ability to capture the bidirectional dependency of time series data, so its evaluation indicators are slightly better than LSTM and GRU. MAPE The values are between 52% and 83% of IRMBL and IRMBG. LSTM is slightly better than GRU. SVM and XGBoost, typical models in the field of machine learning, are inferior to LSTM and GRU in prediction effect, and are much lower than BiLSTM. Even in the summer when IRMBX has the best prediction effect, its E MAE 、E RMSE 、E MAPE The three indicators are 1.64, 1.62, and 1.71 times that of IRMBB respectively; and the evaluation indicators of IRMBS in the four seasons are more than 3 times that of IRMBB. However, if BiLSTM is not optimized by Bayes, even if the data is processed by ICEEMDAN-RCMFE method, it is impossible to get a good prediction effect. For example, the evaluation indicators of IRMB are more than 5 times that of IRMBB.
[0128] Figure 7 The violin error graphs of the six methods in four seasons show that the IRMB method without Bayes hyperparameter optimization has the largest overall error, followed by IRMBS and IRMBX. The overall errors of IRMBG and IRMBL are not much different. Due to the large scale of the training set used in this paper, IRMBL is slightly better than IRMBG. The overall error change of the IRMBB proposed in this paper is the smallest among the six groups of models, which shows its excellent prediction performance. In summary, the superiority of the IRMBB method proposed in this paper compared with the other five methods has been verified.
[0129] Figure 8 The prediction results of autumn inertia by six methods are shown in Figure 2. The results show that compared with the other five methods, the model proposed in this invention better fits the actual inertia curve.
Claims
1. A new short-term prediction method for equivalent inertia of power system that adapts to seasonal changes, characterized by The following steps are involved: Step 1: Divide the annual inertia of the power grid into four seasons and decompose it using the ICEEMDAN algorithm; Step 2: Use the refined composite multi-scale fuzzy entropy RCMFE algorithm to aggregate the subsequences divided and decomposed according to the four seasons; Step 3: Use the mRMR feature extraction algorithm to screen out the most relevant characteristic factors affecting the change of inertia in four seasons as the input data for system inertia prediction; Step 4: Use the Min-Max normalization method to scale the input data required for system inertia prediction in step 3; Step 5: Construct a short-term prediction model for the power system inertia, predict the different components of inertia in different seasons, accumulate the final prediction results, and realize the power system inertia prediction.
2. According to claim 1, a novel short-term prediction method for equivalent inertia of a power system adapting to seasonal changes is characterized by: In the step 1, according to the inherent time scale of the power grid inertia sequence, the power system data is decomposed into multiple intrinsic mode functions and residuals with frequencies from high to low according to the season, and each intrinsic mode function can represent the characteristic components of the power grid inertia at different time scales; The decomposition steps of the ICEEMDAN algorithm are shown in formula (1): In the formula, x i (t) is the noise-added signal; x(t) is the original inertia sequence; α0 represents the initial value of the amplitude coefficient of the added noise; E1(η i (t)) represents the first-order mode obtained by the EMD algorithm; η i (t) indicates adding the i-th white noise; r k (t) represents the k-th order residual component; r k-1 (t) represents the k-1th order residual component; α k-1 =β k-1 std(x(t)) / std(E k (w i (t))) is the standard deviation of white noise, which is used to remove noise; β k-1 is the scale factor; std is the standard deviation; w i (t) represents the white noise sequence added in the i-th iteration; E k (w i (t)) represents the result of the collective operation on all k added white noises; E k (·) indicates the kth mode obtained by EMD algorithm; <·> indicates mean value calculation; M(·) indicates envelope calculation; IMF k (t) represents the kth order modal component.
3. According to claim 2, a new short-term prediction method for equivalent inertia of a power system adapting to seasonal changes is characterized by: The ICEEMDAN algorithm adds white noise η i (t) and local average to achieve the continuity of frequency between adjacent scales, so as to adaptively decompose the non-stationary sequence x(t) into a finite number of intrinsic mode function components with different frequency characteristics and a residual component.
4. According to claim 1, a novel short-term prediction method for equivalent inertia of a power system adapting to seasonal changes is characterized by: In step 2, a fine composite multi-scale fuzzy entropy RCMFE algorithm is used to calculate the RCMFE value of the system inertia in a certain season. The calculation formula of the system inertia RCMFE in a certain season is shown in formula (2): Where x={x1,x2,…,x N } is the inertia power time series; x1, x2,…, x N Indicates the value of each signal; is a coarse-grained sequence; Represents coarse-grained sequences at different scales; 1≤k≤τ,1≤j≤N / τ; k and j are index variables; τ is the scale factor; N represents the number of fuzzy entropy; m is the embedding dimension; n is the gradient of the fuzzy function; r is the similarity tolerance; represents the number of matching vector pairs calculated for the k-th coarse-grained sequence under the embedding dimension m+1 and the scaling factor τ; represents the number of matching vector pairs calculated for the k-th coarse-grained sequence with embedding dimension m and scaling factor τ.
5. According to claim 4, a novel short-term prediction method for equivalent inertia of a power system adapting to seasonal changes is characterized by: The subsequences divided and decomposed according to the four seasons are aggregated; CIMF components with similar complexity in different seasons are classified into the same category and merged into fewer IMF components.
6. According to claim 1, a novel short-term prediction method for equivalent inertia of a power system adapting to seasonal changes is characterized by: In step 3, the correlation analysis of system inertia and climate and operation data is performed and features are further extracted. In addition to the historical inertia of each component, wind speed, light intensity I, average temperature t mean , precipitation PF, net load P net , Synchronous power output P syn , Asynchronous power output P non 、System frequency f sys Eight influencing factors were considered as input features; Since the system inertia and influencing factors are mostly nonlinear, the mRMR feature extraction algorithm is used to select the system inertia power, traditional new energy and common new energy active output in spring, summer, autumn and winter for feature extraction, and screen out the most relevant characteristic factors affecting the inertia changes in the four seasons as the input data for system inertia prediction.
7. A novel short-term prediction method for equivalent inertia of a power system adapted to seasonal changes according to claim 6, characterized in that: The feature screening calculation process of the mRMR feature extraction algorithm is shown in formula (3): Where: X and Y represent two feature variables respectively; p(x) and p(y) represent the marginal probability functions of the two feature variables respectively; p(x, y) represents the joint probability distribution of the two feature variables; x and y represent the specific values of the two feature variables respectively; maxD(S, P) represents the maximum correlation between the feature and the category; D is used to measure the correlation between the feature and the category; S represents the feature set; P represents the target category; D(S,P) represents a measure of the correlation between features and categories; f i represents the i-th feature in the set; I(f i ,P) represents the mutual information between feature i and the target category. Mutual information is used to measure the correlation between variables; minR(S) represents the minimum redundancy between features; I(f i ,f j ) represents the mutual information between feature i and feature j, which is used to measure the correlation between variables; f j represents the jth feature in the set; maxJ(D,R) represents the evaluation function of the feature selection algorithm; J represents the evaluation function itself; R represents the redundancy between features; J(D,R) represents a specific evaluation function; Maximizeφ means maximizing the difference between single variables and correlation and redundancy; I(f i ,c) represents mutual information, which is used to measure the degree of mutual dependence between two variables; c represents the target variable; S t-1 represents the feature set at step t-1; t represents a specific time point in the time series.
8. According to claim 1, a novel short-term prediction method for equivalent inertia of a power system adapting to seasonal changes is characterized by: In step 4, the Min-Max normalization method is used to scale the original data, and its expression is shown in formula (4): In the formula: x' is the normalized data; x is the original data; x max 、x min are the maximum and minimum values in the original data.
9. According to the novel short-term prediction method of equivalent inertia of power system adapting to seasonal changes as claimed in claim 1, it is characterized by: In step 5, the system inertia sequence has obvious positive and negative regularity. When predicting the system equivalent inertia, the influence of historical inertia and future inertia on the prediction accuracy should be considered at the same time; A short-term prediction model of power system inertia is constructed based on the BiLSTM neural network. The BiLSTM neural network consists of two LSTM neural networks, forward and reverse, which can perform bidirectional time series feature extraction on the input data. The BiLSTM neural network considers the inherent laws of the forward and backward data at the same time, and makes predictions from the two directions of history and future. It uses two independent hidden layers to process the data from the forward and reverse directions respectively, and then uses the outputs of the forward and reverse hidden layers as the input of the output layer to neutralize the phase error.
10. A novel short-term prediction method for equivalent inertia of a power system adapted to seasonal changes according to claim 9, characterized in that: In step 5, the hyperparameter optimization process of the BiLSTM neural network based on Bayesian optimization includes the following steps: S5.1: Determine the value range of the hyperparameters of the BiLSTM neural network; S5.2: Use historical data to build a suitable probability model and predict the impact of different hyperparameter combinations on model performance; S5.3: Select a set of hyperparameters from the hyperparameter interval to train the BiLSTM neural network model; S5.4: Train the BiLSTM neural network and evaluate its performance on the validation set. S5.5: If a better parameter is found or the maximum number of iterations is reached, stop. Otherwise, use a new proxy model to continue looking for a better hyperparameter combination. S5.6: Extract the best performing hyperparameter combination from the iterative process.
