Overhead transmission line current-carrying capacity self-correction prediction method based on multi-combination mode decomposition and integrated prediction model

Abstract

An overhead transmission line current-carrying capacity self-correction prediction method based on multi-combination mode decomposition and an integrated prediction model comprises the following steps: firstly, selecting current-carrying capacity influence factors by using a correlation analysis method, and constructing an initial data set according to the selected current-carrying capacity influence factors; then, preprocessing the initial data set by using a weighted moving average method and a modal decomposition algorithm; and finally, substituting the preprocessed initial data set into the integrated prediction model for prediction, and performing weight distribution on the result of the integrated prediction model by using the constructed weight analysis system, thereby obtaining a final prediction result. The method can effectively excavate the potential relationship between the current-carrying capacity of the overhead transmission line and the meteorological environment. The current-carrying capacity prediction precision of the overhead transmission line is simultaneously improved from three aspects of a prediction source, a prediction process and a prediction result.

Description

Technical Field

[0001] The present invention relates to the technical field of overhead transmission line current-carrying capacity prediction, and particularly relates to a self-correcting prediction method for overhead transmission line current-carrying capacity based on multi-combination modal decomposition and an integrated prediction model. Background Technique

[0002] At present, the current-carrying capacity of overhead transmission lines in China is mainly characterized by the current-carrying capacity of overhead transmission lines. There are many influencing factors for the current-carrying capacity of overhead transmission lines, mainly involving wind speed, wind direction, solar radiation intensity, environmental temperature, etc. In addition, the above meteorological factors are all important influencing factors affecting the current-carrying capacity of overhead transmission lines. However, due to the strong volatility and randomness of the above meteorological factors themselves, it is difficult to accurately predict the current-carrying capacity of overhead transmission lines. Therefore, improving the existing line current-carrying capacity prediction method is the key to accurately predicting the current-carrying capacity of overhead transmission lines.

[0003] According to existing research, the existing prediction methods for the current-carrying capacity of overhead transmission lines mainly have the following problems: 1) The models used in the existing prediction methods for the current-carrying capacity of overhead transmission lines are all single prediction models, and the model complexity is relatively low, making it difficult to effectively cope with the strong volatility and randomness of meteorological factors. 2) The existing prediction methods for the current-carrying capacity of overhead transmission lines all consider how to improve the prediction accuracy from the perspective of the prediction process, and cannot comprehensively improve the existing methods from three aspects: the prediction source, the prediction process, and the prediction result, so as to comprehensively improve the prediction accuracy of the line current-carrying capacity.

[0004] At present, the research progress on the prediction of the current-carrying capacity of overhead transmission lines is as follows:

[0005] (1) Chinese Patent: "Interval Prediction Method and Equipment for Overhead Transmission Line Current-Carrying Capacity Based on LSSVM Model" (Publication No.: CN117096860A), mainly uses the processed sample data set and the LSSVM model to obtain the current-carrying capacity of overhead transmission lines in the future time period. However, there are two problems with the prediction scheme mentioned in this invention patent. ①: The hyperparameters required by the model need to be selected manually, which results in weak generalization and low robustness of the model, and it is difficult to process a large amount of sample non-linear data. ②: The influencing factors of the conductor current-carrying capacity involve meteorological data with large randomness such as wind speed and wind direction angle, and directly predicting the data is difficult to ensure the prediction accuracy.

[0006] In summary, this prediction scheme only considers improving the prediction result from the prediction process, and there are defects in the improvement scheme, mainly including: the model used is a single prediction model, the prediction accuracy is low, and the generalization ability is insufficient. In addition, this scheme also ignores improving the prediction result from the two aspects of the prediction source and the prediction result.

[0007] (2) Chinese Patent: "A Method for Predicting Dynamic Ampacity of Cables Based on BiLSTM" (Publication No.: CN115730527A). First, a dataset of cable temperature and ampacity sequences is constructed. After preprocessing the dataset through MEEMD, the BiLSTM model is used to predict the cable ampacity. However, there are two problems with the prediction method mentioned in this invention patent. ①: The hyperparameters of a single BiLSTM model need to be selected manually, which will lead to low generalization ability and poor robustness of the model. ②: Although this prediction scheme considers using MEEMD to preprocess the dataset at the prediction source, there are still problems such as the need to manually select hyperparameters such as the decomposition layer number in the MEEMD model.

[0008] In summary, although this prediction scheme attempts to improve the prediction accuracy from both the prediction source and the prediction process, the proposed scheme has problems such as the need to manually select hyperparameters, weak generalization ability of the model, and low prediction accuracy, and still does not solve the problem of accurate data prediction well. Therefore, this prediction scheme considers improving the prediction accuracy from the prediction source and the prediction process, but does not consider improving the prediction accuracy from the prediction result, and there are certain defects in the proposed prediction scheme.

[0009] (3) Chinese Patent: "A Method for Predicting Dynamic Ampacity of Overhead Transmission Lines Based on RNN Model" (Publication No.: CN117592591A). The ampacity data of the overhead transmission line for a future time period is predicted based on the sample dataset and the RNN model. However, the following problems exist with the prediction method mentioned in this invention patent. ①: The influencing factors of the conductor ampacity include long-term time series data such as wind speed. Due to the characteristics of parameter sharing and multiple consecutive multiplications in a single RNN prediction model, problems such as gradient explosion are likely to occur. ②: The hyperparameters required by the model need to be extracted manually, which results in relatively weak generalization ability of the model.

[0010] In summary, this prediction scheme only considers improving the prediction result of the conductor ampacity from the prediction process, does not consider improving the prediction accuracy from both the prediction source and the prediction result, and there are certain defects in the proposed prediction scheme.

[0011] It is not difficult to find from existing research that there is currently no prediction scheme that can comprehensively consider from the prediction source, the prediction process, and the prediction result to improve the prediction accuracy of the ampacity of overhead transmission lines in all aspects. Moreover, existing prediction schemes mostly use a single prediction model to predict the ampacity of overhead transmission lines, and the model complexity is relatively low, making it difficult to achieve accurate prediction of the ampacity of overhead transmission lines. Summary of the Invention

[0012] To solve the above technical problems, the present invention provides a self-correcting prediction method for the ampacity of overhead transmission lines based on multi-combination modal decomposition and integrated prediction model, which can effectively explore the potential relationship between the ampacity of overhead transmission lines and the meteorological environment. More importantly, the present invention can improve the prediction accuracy of the ampacity of overhead transmission lines simultaneously from three aspects: the prediction source, the prediction process, and the prediction result, providing a sufficient theoretical basis for the safety and capacity increase of overhead transmission lines to a great extent.

[0013] The technical solution adopted by the present invention is as follows:

[0014] A self-correcting prediction method for the ampacity of overhead transmission lines based on multi-combination modal decomposition and integrated prediction model, comprising the following steps:

[0015] Step 1: First, select the influencing factors of the ampacity by using the correlation analysis method, and construct an initial data set according to the selected influencing factors of the ampacity.

[0016] Step 2: Then, preprocess the initial data set by using the weighted moving average method and the modal decomposition algorithm.

[0017] Step 3: Finally, input the preprocessed initial data set into the integrated prediction model for prediction, and use the constructed weight analysis system to allocate weights to the results of the integrated prediction model, so as to obtain the final prediction result.

[0018] The said Step 1 includes the following steps:

[0019] Step 1.1: First, obtain all the meteorological data of the influencing factors of the ampacity of the overhead transmission line and the conductor ampacity data, so as to construct an initial data set containing meteorological data and conductor ampacity data.

[0020] Step 1.2: Respectively use the Person correlation coefficient method and the Spearman correlation coefficient method to conduct a correlation analysis on the meteorological influencing factors of the overhead transmission line ampacity data.

[0021] Step 1.3: Determine the screening threshold of the influencing factors of the ampacity of the overhead transmission line according to the actual engineering requirements, and eliminate the weakly correlated influencing factors according to the screening threshold.

[0022] Step 1.4: Take the intersection of the results obtained by the Person correlation coefficient method and the Spearman correlation coefficient method to finally determine the key influencing factor set of the transmission line ampacity.

[0023] In the said Step 1.2, the formula of the Person correlation coefficient method is:

[0024]

[0025] In the formula: and represents the sample mean; r represents the correlation between variables. The closer the absolute value of r is to 1, the stronger the correlation between the two variables; X i represents the true value of the sample of the i-th meteorological factor; Y i represents the true value of the sample of the i-th overhead transmission line current-carrying capacity; n represents the total number of samples.

[0026] The formula for the Spearman correlation coefficient method is:

[0027]

[0028] In the formula: d i is the difference in the rank values of the i-th data pair. Specifically, d i represents the difference between the current-carrying capacity data and the meteorological data after aligning their ranks; P represents the correlation relationship between samples. When the absolute value of P is closer to 1, it indicates that there is a stronger monotonic correlation relationship between the two variables.

[0029] In step 1.3, determine the screening threshold of the above meteorological influence factors, and eliminate the meteorological influence factors with weak correlations according to the screening threshold. Among them, the screening threshold of the meteorological influence factors is as follows:

[0030]

[0031] In the formula: ε i represents the correlation value of the i-th meteorological influence factor; m represents the number of types of meteorological influence factors.

[0032] In step 1.4, take the intersection of the results obtained by the Person correlation coefficient method and the Spearman correlation coefficient method; specifically as follows: Assume that the set of meteorological influence factors calculated by the Person correlation coefficient method is A, and the set of meteorological influence factors calculated by the Spearman correlation coefficient method is B. Use the following formula to take the intersection of the two sets to obtain the final set of meteorological influence factors C.

[0033] C = A ∩ B.

[0035] Step two includes the following steps:

[0036] Step 2.1: Use the Z-score method to perform data preprocessing on the data of the influence factor set, so as to eliminate the abnormal data in the data of the influence factor set;

[0037] Step 2.2: Divide the data of the influence factor set into a training set and a validation set according to a certain ratio;

[0038] Step 2.3: Use the ensemble empirical mode decomposition to perform a preliminary decomposition of the training set data to obtain multiple primary modal components; Step 2.4: Use the variational mode decomposition algorithm to perform a secondary decomposition of the high-frequency modal components that still exist after the primary decomposition to obtain the residual modal components;

[0039] Step 2.5: The multiple primary modal components and the remaining modal components obtained in step 2.3 and step 2.4 are arranged in order and are ready to be sent to the subsequent prediction model for prediction.

[0040] In step 2.1, the impact factor set data is expressed as follows:

[0041] X={X 1 ,X 2 ,…,X n}

[0042] Where, X 1 Represents the first meteorological factor data; X 2 represents the second meteorological factor data; and X n Represents the nth meteorological factor data;

[0043] The calculation formula of the Z-score method is as follows:

[0044]

[0045] Where: X n Represents the nth meteorological factor data; μ n represents the mean value of the nth meteorological factor data; J n Represents the standard deviation of the nth meteorological factor data.

[0046] Set the Z-score threshold. There is usually no unified standard for setting this threshold. You can consider the larger extreme value after calculating all meteorological factor data as the Z-score threshold. Compare all meteorological factor data calculated by the Z-score method with the Z-score threshold, and remove some abnormal data that exceeds the Z-score threshold.

[0047] In step 2.2, the influencing factor set data is divided into a training set and a validation set according to a certain ratio. The training set is a data subset used to train the machine learning model. The data in the training set is mainly used to adjust the parameters of the model; while the validation set is a data subset used to verify the learning effect of the machine learning model, and the data in the validation set is mainly used to evaluate the predictive performance of the model. At this point, it should be noted that the number of data in the training set should far exceed the number of data in the validation set. In most cases, the ratio of the training set to the training set can be selected as 7:3 or 8:2.

[0048] In step 2.3, the essence of the ensemble empirical mode decomposition is an empirical mode decomposition algorithm with multiple additions of Gaussian white noise:

[0049] Y n (t) = X n (t) + U n (t)

[0050] Where: X n (t) is the set of data of various influencing factors, U n (t) is the Gaussian white noise signal, and Y n (t) is the set of data of various influencing factors after adding Gaussian white noise.

[0051] The specific steps are as follows: First, use the Empirical Mode Decomposition (EMD) algorithm to decompose Y n (t), and decompose Y n (t) into the sum of a series of modal components and a residual component. The calculation formula is:

[0052]

[0053] Where: are the respective modal components obtained by decomposing the original signal function after adding white noise; A represents the number of iterations of the EMD algorithm; Q represents the number of algorithm integrations; res(t) represents the residual component; M represents the number of iterations to stop the algorithm;

[0054] Repeat the above steps, add white noise signals with different amplitudes each time of decomposition, and then perform a mean operation on the set of component signals to obtain multiple initial modal components:

[0055]

[0056] After obtaining the ensemble empirical mode decomposition, For the convenience of writing, IMF_n is used instead in the following text;

[0057] In step 2.4, after the ensemble empirical mode decomposition, some of the components may still have some components with rich information that are not decomposed thoroughly. Such components are usually the high-frequency IMF1 components. Therefore, perform a secondary decomposition on the sub-signal components IMF1 of the modal components that still contain rich information, and the secondary decomposition method selects the variational mode decomposition algorithm in the completely non-recursive method; use the variational mode decomposition algorithm to continue decomposing the sub-signal component IMF1 into sub-sub-signal components with different frequencies;

[0058] Among them, the set of sub-sub-signal components obtained after the variational mode decomposition of IMF1 is shown in the following formula:

[0059] IMF1K = {IMF 11 , IMF 12 , IMF 13 ,....., IMF 1K}

[0060] Where: IMF1K represents the set of sub-sub signal components; IMF 11 represents the first component in the set of sub-sub signal components; and so on until the Kth; IMF 1K represents the Kth component in the set of sub-sub signal components.

[0061] In step 2.5, the multiple modal component sets are shown in the following formula:

[0062] IMF = {IMF 11 , IMF 12 , IMF 13 , IMF 14 ,......, IMF 1K , IMF 2 , IMF 3 ,......IMFn}

[0063] Where, IMF n represents the n components of the ensemble empirical mode decomposition; IMF 1K represents the Kth component obtained by the second decomposition of the first high-frequency component of the ensemble empirical mode decomposition.

[0064] Step three includes the following steps:

[0065] Step 3.1: Select an artificial intelligence optimization algorithm to extract the optimal hyperparameters of the bidirectional long short-term memory network prediction model and the gated recurrent unit prediction model;

[0066] Step 3.2: Input the multiple modal components obtained in step 2.5 into two prediction models, namely the bidirectional long short-term memory network and the gated recurrent unit, which have been optimized with hyperparameters, for prediction;

[0067] Step 3.3: Perform reduction processing on the predicted modal component results obtained in step 3.2 according to the rule of linear summation. The reduction processing results are divided into two major categories: the prediction results of the bidirectional long short-term memory network and the prediction results of the gated recurrent unit; Step 3.4: Establish a comprehensive time-varying weight system according to the grey correlation coefficient method:

[0068] First, calculate the time-varying weights of the two prediction result matrices in step 3.3 by the grey correlation coefficient method;

[0069] Then, multiply the two prediction result matrices by the time-varying weights and add them to obtain the first prediction result of the overhead transmission line current-carrying capacity;

[0070] Step 3.5: Calculate the difference between the first predicted value obtained in Step 3.4 and the true value, and put this difference into the bidirectional long short-term memory network in Step 3.2 for secondary prediction to obtain the second difference prediction result;

[0071] Step 3.6: Consider using an error judgment threshold to screen the secondary difference prediction, and use the second difference prediction result to compensate and correct the first prediction result to obtain the final prediction result;

[0072] Step 3.7: For the second prediction result, use an evaluation index system to judge the quality of the prediction result.

[0073] In the above-mentioned Step 3.1, since both the bidirectional long short-term memory network prediction model and the gated recurrent unit prediction model contain three unknown hyperparameters: learning rate U, number of hidden layers T, and time step S. Among them, the learning rate U determines the step size of the model in each gradient update and is an important hyperparameter affecting the training speed and final accuracy; the number of hidden layers determines the depth of the model; if the time step S is too short, the model may lose long-term dependence relationships, and if it is too long, the computational complexity may increase. At the same time, because the raccoon algorithm has advantages such as strong global search ability, high efficiency, and wide applicability, the raccoon algorithm is used here to extract the optimal hyperparameter values of the three unknown parameters: learning rate U, number of hidden layers T, and time step S in the bidirectional long short-term memory network prediction model and the gated recurrent unit prediction model.

[0074] In the above-mentioned Step 3.2, the modal component matrices are respectively input into two prediction models: the bidirectional long short-term memory network and the gated recurrent unit after hyperparameter optimization for prediction;

[0075] First, the modal component matrix is as follows:

[0076]

[0077] In the formula: n refers to the number of all modal components; m is less than k and refers to the m-th classification of the primary mode; IMF 1k refers to the k-th modal component of the primary decomposition.

[0078] The decomposed modal component matrix X is respectively input into the optimized bidirectional long short-term memory network and the gated recurrent unit. The modal component matrix X is input into the bidirectional long short-term memory network after hyperparameter optimization. The bidirectional long short-term memory network is mainly composed of cell memory units and gating mechanisms (input gate, forget gate, output gate) to ensure the long-term dependence of data.

[0079] (1) Among them, the calculation formula of the forget gate is:

[0080] f t = σ(W f x t + U f h t-1 + b f )

[0081] In the formula: f t refers to the output of the forget gate, and the output range is [0, 1], indicating the memory degree of the previous state C t-1 . σ refers to the activation function, and the output range is [0, 1]; W f refers to the input weight matrix of the forget gate; U f refers to the hidden state weight matrix of the forget gate; b f is the bias term of the forget gate; x t refers to the input feature vector at the current time t; h t-1 refers to the hidden state at the previous time t - 1.

[0082] (2) In addition, the calculation formula of the input gate is:

[0083] i t = σ(W i x t + U i h t-1 + b i )

[0084]

[0085] Among them, i t refers to the output of the input gate, indicating the degree to which the current information is written into the cell state C t ; refers to the candidate memory state, which is calculated through the current input and the previous hidden state; tanh() refers to the hyperbolic sine function, which compresses the value to [-1, 1]; W i and W C refer to the input weight matrices of the input gate and the candidate state; U i and U C refer to the hidden state weight matrices of the candidate state; b i and b C refer to the bias terms of the candidate state.

[0086] (3) Again, the calculation formula of the cell memory unit is as follows:

[0087]

[0088] In the formula: Ct refers to the cell state at the current time t; ⊙ refers to element-wise multiplication; C t-1 refers to the cell state at the previous time; f t refers to the degree of forgetting that the forget gate controls the memory; refers to the proportion that the input gate determines the new information to be written into the cell state.

[0089] (4) Calculation formula of the output gate:

[0090] o t = σ(W o x t + U o h t-1 + b o )

[0091] h t = o t ⊙ tanh(C t )

[0092] In the formula: where o t refers to the output of the output gate, and the output range is between [0, 1], which controls the proportion of the cell state C t output to the hidden state h t ; h t refers to the hidden state at the current time t, which is used as the output and the input of the next layer; W o refers to the input weight matrix of the output gate; U o refers to the hidden state weight matrix of the output gate; b o refers to the bias term of the output gate.

[0093] (5) Calculation formula of the final predicted value:

[0094]

[0095] In the formula: is the predicted value at time t; W out is the weight matrix of the fully connected layer; b out is the bias term of the fully connected layer.

[0096] The gated recurrent unit is a simplified version of the recurrent neural network, which controls the flow and memory of information through the gating mechanism, avoiding the problem of gradient disappearance or explosion in long sequences. The prediction process of the gated recurrent unit model is as follows:

[0097] 1) Calculation formula of the update gate:

[0098] z t = σ(W z x t + U z ht-1 +b z )

[0099] In the formula: z t refers to the output of the update gate, and the output range is between [0, 1], indicating the proportion of retaining the information of the previous hidden state; σ refers to the activation function, and the output range is [0, 1]; W z refers to the input weight matrix of the forget gate; U z refers to the hidden state weight matrix of the forget gate; b z bias term of the forget gate; x t refers to the input feature vector at the current time t; h t-1 refers to the hidden state at the previous time t - 1.

[0100] 2) Calculation formula of the reset gate:

[0101] r t = σ(W r x t + U r h t-1 + b r )

[0102] In the formula: r t refers to the output of the reset gate, and the output range is [0, 1], indicating the proportion of discarding the information of the previous hidden state; W r represents the input weight matrix of the reset gate; U r refers to the hidden state weight matrix of the reset gate; b r refers to the bias term of the reset gate.

[0103] 3) Calculation formula of the candidate hidden state:

[0104]

[0105] In the formula: refers to the candidate hidden state; tanh refers to the hyperbolic sine function, which is used to compress the value to [-1, 1]; W h refers to the input weight matrix; U h refers to the hidden state weight matrix; b h refers to the bias term; r t ⊙ U h h t-1 refers to the previous hidden state h t-1 through the calculation controlled by the reset gate.

[0106] 4) Formula for updating the hidden state:

[0107]

[0108] In the formula: ht refers to the hidden state at the current time step t; (1 - z t ) refers to the control of the candidate hidden state 's impact on the current hidden state; refers to the candidate hidden state; ⊙ refers to element-wise multiplication; z t ⊙h t-1 refers to the update gate controlling the impact of the previous hidden state on the current hidden state.

[0109] 5) The calculation formula for the final predicted value:

[0110]

[0111] In the formula: refers to the predicted value at time t; W out refers to the weight matrix of the fully connected layer; b out refers to the bias term of the fully connected layer.

[0112] In step 3.3, the predicted modal components obtained are superimposed according to the reduction principle,

[0113] For each refined modal component IMF in the matrix ij , use a bidirectional long short-term memory network or a gated recurrent unit to predict its future k-step values respectively, obtaining a predicted value matrix:

[0114]

[0115] In the formula: n is the index range of the initial modal components; m is the index range of the refined modal components; refers to the predicted value at t + k of the j-th refined component of the i-th initial modal component.

[0116] The formula for modal component reduction is as follows:

[0117]

[0118] In the formula: refers to the predicted value at t + k of the j-th refined component of the i-th initial modal component; refers to the residual component at the corresponding time.

[0119] The obtained results are divided into two categories: the single-point value prediction result P of the optimized bidirectional long short-term memory network LSTM , and the single-point value prediction result P of the optimized gated recurrent unit GRU .

[0120] In step 3.4, the formula for the time-varying weight system is as follows:

[0121] Di = ζ 1 × P LSTM + ζ 2 × P GRU

[0122] Where: D i represents the first comprehensive prediction result of the current-carrying capacity of the overhead transmission line, and ζ 1 is the single prediction value weight of the bidirectional long short-term memory network prediction model; ζ 2 is the single prediction value weight of the gated recurrent unit prediction model; P LSTM is the single prediction value of the bidirectional long short-term memory network prediction model; P GRU is the single prediction value of the gated recurrent unit prediction model;

[0123] The calculation formula for the single prediction value weight of the bidirectional long short-term memory network prediction model is as follows:

[0124]

[0125] In the formula: ξ 1 is the comprehensive error value of the bidirectional long short-term memory network prediction model, and ξ 2 is the comprehensive error value of the gated recurrent unit prediction model;

[0126] ξ 1 , ξ 2 The calculation formula is as follows:

[0127]

[0128] In the formula: and respectively represent the difference between the actual value and the predicted value of the sample; represents the average of the differences between the actual values and the predicted values of all sample points.

[0129] In step 3.4, when selecting the time-varying weights of the two prediction models using the grey correlation coefficient method, the calculation formula is as follows:

[0130]

[0131] In the formula: w it is the time-varying weight of the i-th single prediction model at this moment; r i,t-1 is the grey correlation degree value of the prediction model at the previous moment, and the formula is as follows:

[0132]

[0133] In the formula: ξ is the monotonic coefficient; m and M are the absolute values of the minimum error and the maximum error of the single prediction model respectively; Δ i(t - 1) is the absolute value of the error of the i-th sub-prediction model at the (t - 1)-th moment.

[0134] In step 3.5, calculate the first prediction error value and define the error E i The formula is as follows:

[0135] E i = D i - V i

[0136] In the formula: D i represents the first comprehensive prediction result of the overhead transmission line current-carrying capacity; V i represents the true value of the overhead transmission line current-carrying capacity.

[0137] Put E i into any one of the prediction models in the prediction model for secondary prediction, and the set of secondary prediction values obtained is as follows:

[0138] S i = {S 1 , S 2 , S 3 ,..., S n}

[0139] In the formula: S i is the set of secondary prediction values, i = 1, 2, 3......n.

[0140] In step 3.6, consider using an error judgment criterion to screen the secondary prediction values. Among them, the calculation formula of the error judgment criterion is as follows:

[0141] |S i - E i | ≤ |E i |

[0142] If the secondary prediction value satisfies the above formula, then add the secondary prediction value and the first prediction value to obtain the final total prediction value. The calculation formula of the final total prediction value is as follows:

[0143] R i = S i + D i

[0144] In the formula: R i is the final total prediction value.

[0145] In step 3.7, regarding how to evaluate the prediction accuracy of the final total prediction value, the present invention considers using a comprehensive evaluation index system to judge the accuracy of the prediction result. Specifically, the composition of the comprehensive evaluation index system can be expressed by the following formula:

[0146]

[0147] Where: E rmse is the root mean square error, which is the square root of the mean square error. It describes the average difference degree between the predicted value and the true value, and is sensitive to outliers; E MAE is the mean absolute error, which describes the average absolute difference degree between the predicted value and the true value. Although its sensitivity to outliers is not as high as that of Ermse, since it uses absolute values, positive and negative errors will not cancel each other out, and it can accurately reflect the actual situation of the error; R 2 is the goodness of fit, which describes the goodness of fit of the prediction model.

[0148] Generally speaking, among the above three evaluation indexes, the smaller E rmse and E MAE are, the higher the prediction accuracy of the prediction model is, while the larger R 2 is, the higher the prediction accuracy of the prediction model is.

[0149] The self-correcting prediction method for the ampacity of overhead transmission lines based on multi-combination modal decomposition and integrated prediction model of the present invention has the following technical effects:

[0150] 1) Compared with the traditional single-modal decomposition method, the multi-combination modal decomposition method used in the present invention does not rely on a single decomposition algorithm, realizes the complementary advantages between multiple decomposition algorithms, makes full use of their strengths and avoids their weaknesses, and can obtain multi-layer modal components with richer features, which helps to improve the prediction accuracy. The multi-combination modal decomposition method reduces the prediction difficulty from the prediction source and improves the prediction accuracy of the model.

[0151] 2) Compared with the traditional single prediction model, the integrated prediction model used in the present invention combines the advantages of various prediction models, redistributes the prediction results, effectively avoids the influence of "bad points" with large prediction deviations on the prediction results, and at the same time reduces the situation where the prediction error of a single prediction model becomes larger as the prediction step increases. The integrated prediction model has the advantages of high prediction accuracy, good model generalization ability, strong robustness, etc., and effectively improves the prediction accuracy during the prediction process.

[0152] 3) The self-correcting prediction method for the ampacity error of the overhead transmission line used in the present invention sends the first prediction error into the prediction model again for secondary prediction. Combining threshold judgment, the second prediction value is added to the first prediction value, which effectively eliminates the influence of the prediction error on the prediction result to a certain extent. The self-correcting prediction effectively improves the final prediction accuracy from the prediction result level and ensures the accuracy of the prediction result. Brief Description of the Drawings

[0153] Figure 1 is the overall flowchart of the prediction method of the present invention.

[0154] Figure 2 This is the first decomposition diagram of the prediction method of the present invention.

[0155] Figure 3 It is a secondary decomposition diagram of the prediction method of the present invention.

[0156] Figure 4 It is a comparison chart of the iterative optimization effects of the raccoon algorithm used in the prediction method of the present invention and other algorithms.

[0157] Figure 5 It is a comparison chart of the prediction results of the prediction method of the present invention and other methods.

[0158] Figure 6 It is a comparison chart of the evaluation index system results of the prediction method of the present invention and other methods. DETAILED DESCRIPTION

[0159] The self-correction prediction method for the current carrying capacity of overhead transmission lines based on multi-combination modal decomposition and integrated artificial intelligence model first overcomes the problems of low prediction accuracy, weak robustness, and insufficient generalization ability of a single prediction model by integrating different data preprocessing methods with different prediction models, thereby improving the overall prediction accuracy from two aspects: the prediction source and the prediction process. Subsequently, a self-correction method is used to perform a secondary prediction on the prediction error of the conductor current carrying capacity, and the secondary prediction result is combined with the first prediction result to eliminate the error of the first prediction result to a certain extent, thereby further improving the prediction accuracy of the model from the prediction result. The present invention can effectively explore the potential relationship between the current carrying capacity of overhead transmission lines and the meteorological environment, and simultaneously improve the prediction accuracy of the current carrying capacity of overhead transmission lines from three aspects: the prediction source, the prediction process, and the prediction result, thereby effectively ensuring the safe capacity increase of overhead transmission lines.

[0160] The following steps are involved:

[0161] Step S1: Use the meteorological monitoring equipment of the micro-meteorological station to obtain the historical meteorological data of the factors affecting the current carrying capacity of the transmission line, and use the current carrying capacity data of the conductor obtained by the SCADA system of the power grid company to construct a complete initial data set of meteorological data and conductor current carrying capacity. Subsequently, use the data software MATLAB to eliminate the abnormal outlier values ​​that may exist in the data set.

[0162] In step S1, the micro-meteorological station obtains meteorological data including the ambient temperature around the conductor, the wind speed around the conductor, the solar radiation intensity around the conductor, the wind direction angle around the conductor, the altitude of the conductor, the humidity around the conductor, etc. As shown in Table 1.

[0163] Table 1 The influence of meteorological data on the current carrying capacity of conductors

[0164]

[0165] Traditionally, according to the actual engineering requirements of the wire current-carrying capacity in a certain area, the influencing factors of the wire current-carrying capacity can be selected as wind speed, sunshine intensity, temperature, wind direction angle, wire parameters, etc., as shown in Table 2.

[0166] Table 2 Influencing factors of wire current-carrying capacity selected according to actual engineering requirements

[0167]

[0168]

[0169] Step S2: Use two correlation analysis methods, namely the Person correlation coefficient method and the Spearman correlation coefficient method, to conduct a correlation analysis on the influencing factors of the overhead transmission line current-carrying capacity. Then, determine the screening threshold of the influencing factors according to the actual engineering requirements. Finally, eliminate the weakly correlated influencing factors according to the screening threshold, and take the intersection of the results obtained by the above two correlation coefficient methods to determine the final set of influencing factors for the line current-carrying capacity. When using the two correlation analysis methods, it is necessary to select the threshold of the measured influencing factors in combination with the actual engineering. Without other special regulations, the influencing factor threshold can be selected as the average value of all influencing factors. Specifically as follows:

[0170] First, determine the set of influencing factors for the wire current-carrying capacity through repeated screening of the results of the two correlation coefficient methods, and then combine and determine the screening threshold according to the actual engineering requirements. Eliminate the weakly correlated influencing factors according to the screening threshold to finally determine the number of influencing factors for the wire current-carrying capacity;

[0171] Among them: The formula for the Person correlation coefficient method is:

[0172]

[0173] In the formula: and represent the sample mean; r represents the correlation between variables. The closer the absolute value of r is to 1, the stronger the correlation between the two variables.

[0174] The formula for the Spearman correlation coefficient method is:

[0175]

[0176] In the formula, d i is the difference in the rank values of the i-th data pair; n is the total number of observed sample values; the closer the absolute value of P is to 1, the stronger the monotonic correlation relationship between the two variables.

[0177] The set of influencing factors is screened out by the above two methods, and their intersection is taken. Suppose the set of influencing factors screened out by the Pearson correlation coefficient method is:

[0178] Pearson = {ambient temperature, wind speed, solar radiation intensity, altitude}

[0179] The influencing factor set screened by the Spearman correlation coefficient method is:

[0180] Spearman set = {ambient temperature, wind speed, solar radiation intensity, altitude}

[0181] The intersection of the two is: {ambient temperature, wind speed, solar radiation intensity, altitude}. Finally, the above-screened influencing factor set is reselected based on actual engineering needs. For example, some factors (such as wind speed) may be further excluded because their impact on current carrying capacity is not significant under certain conditions. The final set of influencing factors will be used to establish the prediction model. Using the above two correlation coefficient methods, two sets of influencing factors of conductor current carrying capacity can be obtained, and the intersection of the two sets of influencing factors is selected as the total set of influencing factors of conductor current carrying capacity to be predicted next.

[0182] Step S3: Use the Z-score method to perform data preprocessing on the influencing factor set data, so as to eliminate abnormal data in the influencing factor set data; then divide the influencing factor set data into a training set and a validation set according to a certain ratio.

[0183] The impact factor set data is expressed as follows:

[0184] X={X 1 ,X 2 ,…,X n}

[0185] In the formula, X 1 Represents the first meteorological factor data; X 2 represents the second meteorological factor data; and X n Represents the nth meteorological factor data;

[0186] The calculation formula of the Z-score method is as follows:

[0187]

[0188] Where: X n Represents the nth meteorological factor data; μ n represents the mean value of the nth meteorological factor data; J n Represents the standard deviation of the nth meteorological factor data.

[0189] Set the Z-score threshold. There is usually no unified standard for setting this threshold. You can consider the larger extreme value after calculating all meteorological factor data as the Z-score threshold. Compare all meteorological factor data calculated by the Z-score method with the Z-score threshold, and remove some abnormal data that exceeds the Z-score threshold.

[0190] After removing abnormal data from all meteorological factor data, it is necessary to divide the influencing factor set data into training set and validation set according to a certain ratio. Among them, the training set is a data subset used to train the machine learning model. The data in the training set is mainly used to adjust the parameters of the model. The validation set is a data subset used to verify the learning effect of the machine learning model. The data in the validation set is mainly used to evaluate the prediction performance of the model. At this time, it is important to note that the number of data in the training set should be far greater than the number of data in the validation set. In most cases, the ratio of training set to validation set can be 7:3 or 8:2.

[0191] Step S4: Use ensemble empirical mode decomposition and variational mode decomposition, a total of two mode decomposition algorithms, to perform decomposition preprocessing on the meteorological data in the new data set in step S3 in order. Specifically, firstly, ensemble empirical mode decomposition is used to perform an initial decomposition of the data in the new data set to obtain multiple mode components. Then, the variational mode decomposition algorithm is used to perform a secondary decomposition of the high-frequency mode components that still exist after the initial decomposition to obtain the remaining mode components.

[0192] The essence of ensemble empirical mode decomposition is a multiple empirical mode decomposition with Gaussian white noise added:

[0193] Y n (t) = X n (t)+U n (t)

[0194] Where: X n (t) is the original time signal sequence, U n (t) is Gaussian white noise signal, Y n (t) is the time signal series after adding Gaussian white noise.

[0195] Use EMD algorithm to calculate Y n (t) is decomposed into n (t) is decomposed into the sum of a series of modal components and a residual component. The specific calculation formula is as follows:

[0196]

[0197] Where: are the modal components obtained by decomposing the original signal function after adding white noise. A represents the number of iterations of the EMD algorithm, and Q represents the number of algorithm integrations.

[0198] Repeat the above steps, but each decomposition requires adding white noise signals with different amplitudes, and then perform a mean operation on the set of component signals to obtain the final sub-signal components:

[0199]

[0200] After obtaining the ensemble empirical mode decomposition For the sake of convenience in writing, IMF_n is used instead in the following. After the ensemble empirical mode decomposition, some of the components may still have some components with rich information that are not decomposed thoroughly. Such components are usually the high-frequency IMF1 component. Therefore, for the sub-signal components IMF1 of the modal components that still contain rich information, a second decomposition is performed, and the variational mode decomposition algorithm in the completely non-recursive method is selected as the second decomposition method.

[0201] The purpose of using the variational mode decomposition algorithm is to further decompose the component IMF1 with rich information into sub-signal components of different frequencies. The core of this decomposition method is to construct and solve a variational type problem. This decomposition method uses iterative search for the optimal solution to determine the center frequency and bandwidth of each decomposed component. Since the relevant principles of the variational mode decomposition algorithm are already very mature, they will not be introduced in detail here. Among them, the set of sub-signal components obtained after the variational mode decomposition of IMF1 is shown in the following formula:

[0202] IMF1_K = {IMF 11 , IMF 12 , IMF 13 ,......, IMF 1K}

[0203] Step S5: Arrange the multiple modal components and the remaining modal components obtained in step S4, mainly by converting them into a matrix expression for subsequent prediction calculations.

[0204] The set of multiple modal components is shown in the following formula:

[0205] IMF = {IMF 11 , IMF 12 , IMF 13 , IMF 14 ,......, IMF 1K , IMF 2 , IMF 3 ,...... IMF_n}

[0206] In the formula, IMF nrepresents n components of the ensemble empirical mode decomposition, and IMF 1K represents the Kth component obtained by the second decomposition of the first high-frequency component of the ensemble empirical mode decomposition.

[0207] The modal component matrix is as follows:

[0208]

[0209] In the formula: n refers to the number of all modal components; m is less than k and refers to the mth classification of the primary mode; IMF 1k refers to the kth modal component of the primary decomposition.

[0210] Step S6: Select an artificial intelligence optimization algorithm to extract the optimal hyperparameters of the bidirectional long short-term memory network and the gated recurrent unit. Specifically, the hyperparameters to be optimized mainly include the learning rate and the number of hidden layers, aiming to ensure that the bidirectional long short-term memory network and the gated recurrent unit are in the best prediction performance state.

[0211] The bidirectional long short-term memory network is mainly composed of cell memory units and gating mechanisms (input gate, forget gate, output gate) to ensure the long-term dependence of data.

[0212] (1) Among them, the calculation formula of the forget gate is:

[0213] f t =σ(W f x t +U f h t-1 +b f )

[0214] In the formula: f t refers to the output of the forget gate, and the output range is [0,1], indicating the memory degree of the previous state C t-1 ; σ refers to the activation function, and the output range is [0,1]; W f refers to the input weight matrix of the forget gate; U f refers to the hidden state weight matrix of the forget gate; b f the bias term of the forget gate; x t refers to the input feature vector at the current time t; h t-1 refers to the hidden state at the previous time t-1.

[0215] (2) In addition, the calculation formula of the input gate:

[0216] i t =σ(W i x t +U i h t-1 +bi )

[0217]

[0218] where i t represents the output of the input gate, indicating the degree to which the current information is written into the cell state C t ; represents the candidate memory state, calculated from the current input and the previous hidden state; tanh() represents the hyperbolic tangent function, which compresses the value to [-1, 1]; W i and W C represent the input weight matrices of the input gate and the candidate state; U i and U C represent the hidden state weight matrices of the candidate state; b i and b C represent the bias terms of the candidate state.

[0219] (3) Again, the calculation formula of the cell memory unit is as follows:

[0220]

[0221] In the formula: C t represents the cell state at the current time t; ⊙ represents element-wise multiplication; C t-1 represents the cell state at the previous time; f t represents the degree of forgetting of the memory controlled by the forget gate; represents the proportion of the new information written into the cell state determined by the input gate.

[0222] (4) The calculation formula of the output gate:

[0223] o t = σ(W o x t + U o h t-1 + b o )

[0224] h t = o t ⊙ tanh(C t )

[0225] In the formula: o t represents the output of the output gate, and the output range is between [0, 1], controlling the proportion of the cell state C t output to the hidden state h t ; h t represents the hidden state at the current time t, serving as the output and the input of the next layer; W o represents the input weight matrix of the output gate; Uo refers to the hidden state weight matrix of the output gate; b o refers to the bias term of the output gate.

[0226] (5) Calculation formula for the final predicted value:

[0227]

[0228] In the formula: is the predicted value at time t; W out is the weight matrix of the fully connected layer; b out is the bias term of the fully connected layer.

[0229] The gated recurrent unit is a simplified version of the recurrent neural network. It controls the flow and memory of information through a gating mechanism to avoid the problem of gradient vanishing or explosion in long sequences. The prediction process of the gated recurrent unit model is as follows:

[0230] (1) Calculation formula for the update gate:

[0231] z t = σ(W z x t + U z h t-1 + b z )

[0232] In the formula: z t refers to the output of the update gate, and the output range is between [0, 1], indicating the proportion of retaining the information of the previous hidden state; σ refers to the activation function, and the output range is [0, 1]; W z refers to the input weight matrix of the forget gate; U z refers to the hidden state weight matrix of the forget gate; b z is the bias term of the forget gate; x t refers to the input feature vector at the current time t; h t-1 refers to the hidden state at the previous time t - 1.

[0233] (2) Calculation formula for the reset gate:

[0234] r t = σ(W r x t + U r h t-1 + b r )

[0235] In the formula: r t refers to the output of the reset gate, and the output range is [0, 1], indicating the proportion of discarding the information of the previous hidden state; W r represents the input weight matrix of the reset gate; U rRefers to the weight matrix for resetting the hidden state of the gate; b r Refers to the bias term of the reset gate.

[0236] (3) Calculation formula for the candidate hidden state:

[0237]

[0238] In the formula: Refers to the candidate hidden state; tanh refers to the hyperbolic sine function, used to compress the value to [-1, 1]; W h Refers to the weight matrix of the input; U h Refers to the weight matrix of the hidden state; b h Refers to the bias term; r t ⊙U h h t-1 Refers to the previous hidden state h t-1 Calculation through the control of the reset gate.

[0239] (4) Formula for updating the hidden state:

[0240]

[0241] In the formula: h t Refers to the hidden state at the current time step t; (1 - z t ) refers to controlling the candidate hidden state The influence on the current hidden state; Refers to the candidate hidden state; ⊙ refers to element-wise multiplication; z t ⊙h t-1 Refers to the influence of the update gate controlling the previous hidden state on the current hidden state.

[0242] (5) Calculation formula for the final predicted value:

[0243]

[0244] In the formula: Refers to the predicted value at time t; W out Refers to the weight matrix of the fully connected layer; b out Refers to the bias term of the fully connected layer.

[0245] Adopt a feature extraction algorithm to extract the optimal values of the two unknown parameters, the learning rate U and the number of hidden layers T, in the bidirectional long short-term memory network prediction model and the gated recurrent unit prediction model. Specifically, the raccoon algorithm is used for feature extraction, and the learning rate U and the number of hidden layers T are used as the objects to be extracted and optimized. The specific steps for extracting and optimizing the objects are as follows:

[0246] (1) Initialization stage:

[0247] Randomly initialize the "positions" of multiple raccoons as hyperparameters to be optimized. The position of each raccoon represents a combination of hyperparameters U and T. Suppose we have N raccoons, and the "position" of each raccoon represents a hyperparameter combination:

[0248] P i =(U i , T i )

[0249] where i = 1, 2, 3, …, N, representing the position of the i-th raccoon.

[0250] (2) Fitness evaluation:

[0251] Each raccoon learns at its current position and evaluates its fitness based on the prediction accuracy of the model. The fitness function can be measured using the training error, validation error, or cross-validation results of the model. Here, the mean squared error (MSE) is used as the fitness function:

[0252]

[0253] where: y j is the actual value, while represents the predicted value.

[0254] (3) Update phase:

[0255] The raccoons perform local search based on the fitness information and update their positions. This process has two stages: the exploration stage and the exploitation stage.

[0256] 1) Exploration stage:

[0257] In the exploration stage, the raccoons explore the surrounding area through random walks. Specifically, the following formula can be used to update the position:

[0258] U i new = U i + α 1 · (U best - U i ) + β 1 · rand(0, 1)

[0259] T i new = T i + α 2 · (T best - T i ) + β 2 · rand(0, 1)

[0260] where: α 1 and α 2is the learning factor, which controls the search range of the raccoon near the optimal solution. β 1 and β 2 are random perturbation factors, which allow the raccoon to perform random exploration. rand(0,1) can generate a random number from 0 to 1, U best and T best refer to the current parameters of the raccoon with the best fitness among all raccoons.

[0261] 2) Exploration phase

[0262] In the exploration phase, the raccoon will choose whether to update its position according to the current fitness. If the new position is better than the current fitness, it will be updated; otherwise, it will not be updated.

[0263] (4) Terminate the calculation

[0264] This process is repeated multiple times until the termination condition is met. The termination condition can be: reaching the maximum number of iterations and finding the optimal value.

[0265] Step S7: Input the modal component matrix obtained in Step S5 into two prediction models, namely the bidirectional long short-term memory network and the gated recurrent unit, which have been optimized with hyperparameters, for prediction.

[0266] Select two prediction models, namely the bidirectional long short-term memory network prediction model and the gated recurrent unit prediction model optimized in Step S6, to perform time series prediction on the two major categories of modal component sets in Step S5 respectively.

[0267] The modal component matrix is respectively input into two prediction models, namely the bidirectional long short-term memory network and the gated recurrent unit, which have been optimized with hyperparameters, for prediction;

[0268] First, the modal component matrix is as follows:

[0269]

[0270] In the formula: n refers to the number of all modal components; m is less than k, referring to the m-th classification of the primary mode; IMF 1k refers to the k-th modal component of the primary decomposition. The decomposed modal component matrix X is respectively input into the optimized bidirectional long short-term memory network and the gated recurrent unit.

[0271] Step S8: Restore the predicted modal component results obtained in Step S7 according to the rule of linear addition. Specifically, the restoration results are divided into two major categories: the prediction result matrix of the bidirectional long short-term memory network and the prediction result matrix of the gated recurrent unit. Specifically as follows:

[0272] For each refined modal component IMF in the matrix ij, Use a bidirectional long short-term memory network or a gated recurrent unit to predict its future k-step values respectively, and obtain a prediction value matrix:

[0273]

[0274] Where: n is the index range of the initial modal components; m is the index range of the refined modal components; Refers to the predicted value of the j-th refined component of the i-th initial modal component at t + k.

[0275] The formula for modal component restoration is as follows:

[0276]

[0277] Where: Refers to the predicted value of the j-th refined component of the i-th initial modal component at t + k; Refers to the residual component at the corresponding time.

[0278] The obtained results are divided into two categories: the single-point value prediction result P of the optimized bidirectional long short-term memory network LSTM , the single-point value prediction result P of the optimized gated recurrent unit GRU .

[0279] Step S9: Establish a comprehensive time-varying weight system according to the grey correlation coefficient method. First, calculate the time-varying weights of the two prediction result matrices in step S8 by the grey correlation coefficient method; then, multiply the prediction result matrices of the two models by the time-varying weights and add them to obtain the first prediction result of the overhead transmission line current-carrying capacity.

[0280] Among them, the formula for the prediction comprehensive variable weight system is as follows:

[0281] D i = ζ 1 ×P LSTM + ζ 2 ×P GRU

[0282] Among them, D i represents the first comprehensive prediction result of the overhead transmission line current-carrying capacity, ζ 1 is the single prediction value weight of the bidirectional long short-term memory network prediction model, ζ 2 is the single prediction value weight of the gated recurrent unit prediction model, P LSTM is the single prediction value of the bidirectional long short-term memory network prediction model, P GRU is the single prediction value of the gated recurrent unit prediction model.

[0283] The calculation formula for the single prediction value weight of the bidirectional long short-term memory network prediction model is as follows:

[0284]

[0285] Where: ξ 1 is the comprehensive error value of the bidirectional long short-term memory network prediction model, and ξ 2 is the comprehensive error value of the gated recurrent unit prediction model.

[0286] ξ 1 and ξ 2 are calculated as follows:

[0287]

[0288] Where: and represent the differences between the actual value and the predicted value of the sample respectively, while represents the average of the differences between the actual value and the predicted value of all sample points.

[0289] When selecting the time-varying weights of the above two prediction models using the grey correlation coefficient method described in step S9, the calculation formula is as follows:

[0290]

[0291] Where, w it is the time-varying weight of the i-th single prediction model at this moment; and r i,t-1 is the grey correlation degree value of the prediction model at the previous moment, and the formula is as follows:

[0292]

[0293] Where, ξ is the monotonic coefficient, usually taken as 0.4; and m and M are the absolute minimum error and the absolute maximum error of the single prediction model respectively; Δ i (t - 1) is the absolute error of the i-th sub-prediction model at the (t - 1)-th moment.

[0294] Step S10: Calculate the difference between the first predicted value obtained in step S9 and the true value, and put this difference into the long short-term memory network in step S7 for secondary prediction to obtain the secondary difference prediction result. At the same time, consider using the error judgment threshold to screen this secondary difference prediction, and use the secondary difference prediction result to compensate and correct the first prediction result to obtain the final prediction result.

[0295] Calculate the first prediction error value, define the error E i The formula is as follows:

[0296] E i = D i - V i

[0297] Where: Di represents the first comprehensive prediction result of the current-carrying capacity of the overhead transmission line; Vi represents the true value of the current-carrying capacity of the overhead transmission line.

[0298] Put the E in step S9 i into any one of the prediction models in the prediction model in step S7 for secondary prediction, and obtain the following set of secondary prediction values:

[0299] S i ={S 1 , S 2 , S 3 ,..., S n}

[0300] Where: S i is the set of secondary prediction values.

[0301] And consider using an error judgment criterion to screen the secondary prediction values. The calculation formula of the error judgment criterion is as follows:

[0302] |S i -E i | ≤ |E i |

[0303] If the secondary prediction value satisfies the above formula, add the secondary prediction value and the first prediction value to obtain the final total prediction value. The calculation formula of the final total prediction value is as follows:

[0304] R i = S i + D i

[0305] Where: R i is the final total prediction value.

[0306] Step S11: Use a relevant evaluation index system to judge the quality of the prediction result for the final total prediction value. The relevant evaluation index system is calculated using the following formula:

[0307]

[0308] Among them: E rmse is the root mean square error, that is, the square root of the mean square error. It describes the average difference degree between the predicted value and the true value and is sensitive to outliers. E MAE is the mean absolute error. It describes the average absolute difference degree between the predicted value and the true value and is not sensitive to outliers. R 2 is the goodness of fit. It describes the goodness of fit of the prediction model.

[0309] Among the above three evaluation indexes, E rmse and E MAEThe smaller it is, the higher the prediction accuracy of the prediction model, while for R 2 the larger it is, the higher the prediction accuracy of the prediction model.

[0310] Verification example:

[0311] Figure 2 is the initial decomposition effect diagram of the prediction method of the present invention. It can be seen from Figure 2 that the original data is decomposed into 7 component forms with different degrees of stationarity. The first component form can basically reflect the overall trend of the change of the original data; while the second, third, and fourth components can reflect the periodic trend of the change of the original data; the fifth, sixth, and seventh components have relatively poor stationarity and mainly reflect the random trend of the change of the original data. At this time, it should be particularly noted that although the stationarity of components 5, 6, and 7 is relatively poor and may have less benefit in improving the overall prediction accuracy, their stationarity peaks are much smaller than those of other components. Even if they have less benefit in improving the overall prediction accuracy, they will not overly reduce the prediction accuracy. It is not difficult to see that the above 7 components are all more stationary than the original data. Even if the non-stationarity of some component data still exists, the non-stationarity peaks of this part of the component data have been significantly reduced.

[0312] Figure 3 is the secondary decomposition effect diagram of the prediction method of the present invention. It can be seen from Figure 3 that after the initial decomposition, the first component form can only basically reflect the overall trend of the change of the original data, but there is still a certain degree of non-stationarity, that is, the decomposition is not complete. In order to better achieve accurate prediction of data, it is considered to continue the secondary decomposition of the first component. The first component is decomposed into 4 sub-component forms with different degrees of stationarity. The first sub-component form can basically reflect the overall trend of the change of the first component; while the second sub-component can reflect the periodic trend of the change of the first component; the third and fourth components have relatively poor stationarity and mainly reflect the random trend of the change of the original data. Generally speaking, compared with the first component, the 4 sub-components all continue to reduce the non-stationarity to a certain extent, which is very beneficial for achieving accurate prediction of data.

[0313] To demonstrate the superior performance of the raccoon algorithm used in the present invention in model hyperparameter optimization, it is considered to compare the optimization performance of the raccoon algorithm with the common particle swarm algorithm and genetic algorithm in the field of prediction. Figure 4 is the optimization iteration curve diagram of the raccoon algorithm of the prediction method of the present invention. It can be seen from Figure 4It can be seen that the above three algorithms all go through 40 iterations. In terms of the iteration speed, the particle swarm optimization algorithm and the genetic algorithm find the global optimal solution at the 37th and 29th iterations respectively, and the iteration stops. However, the raccoon algorithm used in the present invention stops iterating at the 21st iteration and finds the global optimal solution. Therefore, the raccoon algorithm used in the present invention has a more powerful ability to iteratively optimize the model hyperparameters and is more suitable for optimizing the model hyperparameters in the field of dynamic ampacity prediction of wires. In addition, in terms of the iteration accuracy, the minimum fitness function values obtained by optimizing the particle swarm optimization algorithm and the genetic algorithm are 5.11 and 4.78 respectively, while the minimum fitness function value obtained by optimizing the raccoon algorithm used in the present invention is 4.18. In short, the fitness function value of the raccoon algorithm is smaller, it can find a better solution, and its optimization ability is stronger.

[0314] To reflect the prediction accuracy of the present invention, it is considered to compare the prediction accuracy of the present invention with the long short-term memory network prediction model and the least squares support vector machine prediction model commonly used in the prediction field. Figure 5 is the prediction result of the prediction method of the present invention. From Figure 5 It can be intuitively seen that the prediction accuracy of the present invention is significantly improved compared with the prediction accuracies of the other two models.

[0315] To more intuitively reflect the comparison of the prediction results between the present invention and other prediction methods, it is considered to visualize the evaluation indexes of the present invention and the long short-term memory network prediction model and the least squares support vector machine prediction model commonly used in the prediction field. The visualization results are as Figure 6 shown. Analyzing Figure 6 it can be found that the R 2 evaluation index of the prediction method of the present invention is 99.51%, which is increased by 2.52% and 5.91% respectively compared with the long short-term memory network prediction model and the least squares support vector machine prediction model; in addition, the E MAE evaluation index of the prediction method of the present invention is 6.159, which is decreased by 12.446 and 21.6189 respectively compared with the long short-term memory network prediction model and the least squares support vector machine prediction model; finally, the E rmse evaluation index of the prediction method of the present invention is 8.9885, which is decreased by 13.6496 and 24.8812 respectively compared with the long short-term memory network prediction model and the least squares support vector machine prediction model.

[0316] In summary, all the evaluation indexes of the prediction method of the present invention are better than those of the other two prediction models, which shows that the prediction method of the present invention has good prediction performance and is very suitable for predicting the dynamic ampacity of the line.

Claims

1. A self-correcting prediction method for the current carrying capacity of overhead transmission lines based on multi-combination modal decomposition and integrated prediction model, characterized in that The following steps are involved: Step 1: First, the current carrying capacity influencing factors are selected using the correlation analysis method, and the initial data set is constructed based on the selected current carrying capacity influencing factors; Step 2: Then, the initial data set is preprocessed using weighted moving average method and modal decomposition type algorithm; Step 3: Finally, the preprocessed initial data set is brought into the integrated prediction model for prediction, and the constructed weight analysis system is used to assign weights to the results of the integrated prediction model to obtain the final prediction results.

2. The self-correcting prediction method for overhead transmission line current carrying capacity based on multi-combination modal decomposition and integrated prediction model according to claim 1 is characterized in that: The step 1 comprises the following steps: Step 1.1: First, obtain all meteorological data of the influencing factors of the current carrying capacity of overhead transmission lines and the current carrying capacity data of conductors, so as to construct an initial data set including meteorological data and current carrying capacity data of conductors; Step 1.2: Use the Person correlation coefficient method and the Spearman correlation coefficient method to conduct correlation analysis on the meteorological influencing factors of the current carrying capacity data of overhead transmission lines; Step 1.3: Determine the screening threshold of the influencing factors of the current carrying capacity of overhead transmission lines according to the actual engineering requirements, and eliminate the weakly correlated influencing factors according to the screening threshold; Step 1.4: Take the intersection of the results obtained by the Person correlation coefficient method and the Spearman correlation coefficient method to finally determine the set of key influencing factors of the current carrying capacity of the transmission line.

3. The self-correcting prediction method for overhead power line current carrying capacity based on multi-combination modal decomposition and integrated prediction model according to claim 2 is characterized by: In step 1.2, the formula of the Person correlation coefficient method is: Where: and represents the sample mean; r represents the correlation between variables. The closer the absolute value of r is to 1, the stronger the correlation between the two variables is; X i represents the true value of the sample of the i-th meteorological factor; Y i represents the true value of the sample of the current carrying capacity of the i-th overhead transmission line; n represents the total number of samples; The formula for the Spearman correlation coefficient method is: Where: d i It indicates the difference between the line current carrying capacity data and the meteorological data after aligning their positions; P represents the correlation between samples. The closer the absolute value of P is to 1, the more monotonic correlation there is between the two variables.

4. The self-correcting prediction method for overhead power line current carrying capacity based on multi-combination modal decomposition and integrated prediction model according to claim 2 is characterized by: In the step 1.3, the screening threshold of the meteorological influencing factors is determined, and the meteorological influencing factors with weak correlation are eliminated according to the screening threshold; wherein the screening threshold of the meteorological influencing factors is as follows: Where: ε i It represents the correlation value of the i-th meteorological influencing factor; m represents the number of types of meteorological influencing factors.

5. According to claim 1, the self-correcting prediction method for overhead transmission line current carrying capacity based on multi-combination modal decomposition and integrated prediction model is characterized by: The step 2 comprises the following steps: Step 2.1: Use the Z-score method to preprocess the impact factor set data, thereby eliminating abnormal data in the impact factor set data; Step 2.2: Divide the influencing factor set data into a training set and a validation set according to a certain ratio; Step 2.3: Use the ensemble empirical mode decomposition to perform a preliminary decomposition of the training set data to obtain multiple primary modal components; Step 2.4: Use the variational mode decomposition algorithm to perform a secondary decomposition of the high-frequency modal components that still exist after the primary decomposition to obtain the residual modal components; Step 2.5: The multiple primary modal components and the remaining modal components obtained in step 2.3 and step 2.4 are arranged in order and are ready to be sent to the subsequent prediction model for prediction.

6. The self-correcting prediction method for overhead power transmission line current carrying capacity based on multi-combination modal decomposition and integrated prediction model according to claim 5 is characterized by: In step 2.1, the impact factor set data is expressed as follows: X={X1,X2,…,X n } In the formula, X1 represents the first meteorological factor data; X2 represents the second meteorological factor data; and X n Represents the nth meteorological factor data; The calculation formula of the Z-score method is as follows: Where: X n Represents the nth meteorological factor data; μ n represents the mean value of the nth meteorological factor data; J n Represents the standard deviation of the nth meteorological factor data.

7. The self-correcting prediction method for overhead power line current carrying capacity based on multi-combination modal decomposition and integrated prediction model according to claim 5 is characterized by: EMD is a multi-step EMD algorithm with Gaussian white noise added: Y n (t)=X n (t)+U n (t) Where: X n (t) is the aggregate data of various impact factors, U n (t) is Gaussian white noise signal, Y n (t) is the aggregate data of various impact factors after adding Gaussian white noise; The specific steps are as follows: First, use the empirical mode decomposition algorithm (EMD) to n (t) is decomposed into n (t) decomposed into a series of modal components and a sum of residual components; The calculation formula is: Where: are the modal components of the original signal function after white noise is added; A represents the number of EMD algorithm iterations; Q represents the number of algorithm integrations; res(t) represents the residual component; M represents the number of algorithm iteration stops; Repeat the above steps, add white noise signals with different amplitudes to each decomposition, and then perform mean calculation on the component signal set to obtain multiple primary modal components; Finally, we get the set empirical mode decomposition After the ensemble empirical mode decomposition, some components may still have rich information but not completely decomposed. Such components are high-frequency IMF1 components. Therefore, the sub-signal components IMF1 of some modal components that still contain rich information are secondary decomposed, and the secondary decomposition method uses the variational mode decomposition algorithm in the completely non-recursive method. The variational mode decomposition algorithm is used to further decompose the sub-signal components IMF1 into sub-sub-signal components of different frequencies. Among them, the sub-signal component set obtained after variational mode decomposition of IMF1 is shown in the following formula: IMF1K={IMF 11 ,IMF 12 ,IMF 13 ,.....,IMF 1K } Where: IMF1K represents the sub-sub-signal component set; IMF 11 represents the first component in the sub-sub-signal component set; and so on, until the Kth; IMF 1K represents the Kth component in the sub-sub-signal component set; The set of multiple modal components is shown in the following formula: IMF={IMF 11 ,IMF 12 ,IMF 13 ,IMF 14 ,......,IMF 1K ,IMF2,IMF3,......IMFn} In the formula, IMF n Represents the n components of the ensemble empirical mode decomposition; IMF 1K Represents the Kth component obtained by the secondary decomposition of the first high-frequency component of the ensemble empirical mode decomposition.

8. The self-correcting prediction method for overhead power line current carrying capacity based on multi-combination modal decomposition and integrated prediction model according to claim 1 is characterized by: The step three comprises the following steps: Step 3.1: Use an artificial intelligence optimization algorithm to extract the optimal hyperparameters of the bidirectional long short-term memory network prediction model and the gated recurrent unit prediction model; Step 3.2: Input the multiple modal components obtained in step 2.5 into the bidirectional long short-term memory network and the gated recurrent unit after hyperparameter optimization for prediction in two prediction models; Step 3.3: The predicted modal component results obtained in step 3.2 are restored according to the rule of linear addition; the restoration results are divided into two categories: the prediction results of the bidirectional long short-term memory network and the prediction results of the gated recurrent unit; Step 3.4: Establish a comprehensive time-varying weight system based on the grey correlation coefficient method: First, the time-varying weights of the two prediction result matrices in step 3.3 are calculated using the grey correlation coefficient method; Then, the two prediction result matrices are multiplied by the time-varying weights and added together to obtain the first prediction result of the current carrying capacity of the overhead transmission line; Step 3.5: Calculate the difference between the first predicted value and the true value obtained in step 3.4, and put the difference into the bidirectional long short-term memory network in step 3.2 for secondary prediction to obtain the second difference prediction result; Step 3.6: Consider using the error judgment threshold to screen the secondary difference prediction, and use the second difference prediction result to compensate and correct the first prediction result to obtain the final prediction result; Step 3.7: Based on the second prediction results, use the evaluation index system to judge the quality of the prediction results.

9. The self-correcting prediction method for overhead power transmission line current carrying capacity based on multi-combination modal decomposition and integrated prediction model according to claim 8 is characterized by: The modal component matrices are input into the bidirectional long short-term memory network and the gated recurrent unit after hyperparameter optimization for prediction. First, the modal component matrix looks like this: Where: n refers to the number of all modal components; m is less than k, which refers to the mth classification of the first mode; IMF 1k refers to the kth modal component of the initial decomposition; The decomposed modal component matrix X is input into the optimized bidirectional long short-term memory network and gated recurrent unit respectively; the modal component matrix X is input into the bidirectional long short-term memory network after hyperparameter optimization, and the bidirectional long short-term memory network includes an input gate, a forget gate, and an output gate to ensure the long-term dependency of the data; (1) The calculation formula of the forget gate is: f t =σ(W f x t +U f h t-1 +b f ) Where: f t Refers to the output of the forget gate, the output range is [0,1], indicating the previous state C t-1 The degree of memory; σ refers to the activation function, and the output range is [0,1]; W f Refers to the input weight matrix of the forget gate; U f Refers to the hidden state weight matrix of the forget gate; b f Bias term of forget gate; x t Refers to the input feature vector at the current time t; h t-1 Refers to the hidden state at the previous moment t-1; (2) In addition, the calculation formula of the input gate is: i t =σ(W i x t +U i h t-1 +b i ) Among them, i t Refers to the output of the input gate, indicating that the current information is written into the cell state C t the extent of; refers to the candidate memory state, which is calculated by the current input and the previous hidden state; tanh() refers to the hyperbolic sine function, which compresses the value to [-1,1]; W i and W C Refers to the input weight matrix of the input gate and candidate state; U i and U C refers to the hidden state weight matrix of the candidate state; b i and b C Refers to the bias term of the candidate state; (3) Again, the calculation formula of the cell memory unit is as follows: Where: C t refers to the cell state at the current time t; ⊙ refers to the element-by-element multiplication; C t-1 Refers to the cell state at the previous moment; f t It refers to the forget gate that controls the degree of memory forgetting; It refers to the ratio of new information written into the cell state determined by the input gate; (4) Calculation formula of output gate: the t =σ(W o x t +U o h t-1 +b o ) h t =o t ⊙tanh(C t ) In the formula: o t Refers to the output of the output gate, the output range is between [0,1], and controls the cell state C t Output to hidden state h t The ratio of h t Refers to the hidden state at the current time t, which serves as the output and input of the next layer; W o Refers to the input weight matrix of the output gate; U o refers to the hidden state weight matrix of the output gate; b o Refers to the bias term of the output gate; (5) Calculation formula of the final prediction value: Where: is the predicted value at time t; W out is the weight matrix of the fully connected layer; b out is the bias term of the fully connected layer; The prediction process of the gated recurrent unit model is as follows: 1) Update the calculation formula of the gate: z t =σ(W z x t +U z h t-1 +b z ) Where: z t refers to the output of the update gate, the output range is between [0,1], indicating the proportion of the previous hidden state information retained; σ refers to the activation function, the output range is [0,1]; W z Refers to the input weight matrix of the forget gate; U z Refers to the hidden state weight matrix of the forget gate; b z Bias term of forget gate; x t Refers to the input feature vector at the current time t; h t-1 Refers to the hidden state at the previous moment t-1; 2) Reset gate calculation formula: r t =σ(W r x t +U r h t-1 +b r ) Where: r t Refers to the output of the reset gate, the output range is [0,1], indicating the proportion of discarding the previous hidden state information; W r Represents the input weight matrix of the reset gate; U r refers to the hidden state weight matrix of the reset gate; b r refers to the bias term of the reset gate; 3) Calculation formula for candidate hidden states: Where: refers to the candidate hidden state; tanh refers to the hyperbolic sine function, which is used to compress the value to [-1,1]; W h Refers to the input weight matrix; U h refers to the weight matrix of the hidden state; b h refers to the bias term; r t ⊙U h h t-1 refers to the previous hidden state h t-1 After the calculation of the reset gate control; 4) Formula for hidden state update: Where: h t refers to the hidden state at the current time step t; (1-z t ) refers to controlling the candidate hidden state Impact on the current hidden state; refers to the candidate hidden state; ⊙ refers to the element-by-element multiplication; z t ⊙h t-1 It refers to the update gate controlling the influence of the previous hidden state on the current hidden state; 5) Calculation formula for the final prediction value: Where: Refers to the predicted value at time t; W out Refers to the weight matrix of the fully connected layer; b out Refers to the bias term of the fully connected layer; The predicted modal components are superimposed according to the restoration principle. For each refined modal component IMF in the matrix ij , use the bidirectional long short-term memory network or gated recurrent unit to predict the value of the next k steps respectively, and get the predicted value matrix: Where: n is the index range of the initial modal component; m is the index range of the refined modal component; It refers to the predicted value of the jth refined component of the i-th initial modal component at t+k; The formula for modal component restoration is as follows: Where: It refers to the predicted value of the jth refined component of the i-th initial modal component at t+k; refers to the residual component at the corresponding moment; The results obtained can be divided into two categories: the single point value prediction results of the optimized bidirectional long short-term memory network P LSTM , the single-point value prediction result P of the optimized gated recurrent unit GRU ; The formula of the comprehensive time-varying weight system is as follows: D i =ζ1×P LSTM +ζ2×P GRU Where: D i represents the first comprehensive prediction result of the current carrying capacity of overhead transmission lines, ζ1 is the single prediction value weight of the bidirectional long short-term memory network prediction model; ζ2 is the single prediction value weight of the gated cyclic unit prediction model; P LSTM is the single prediction value of the bidirectional long short-term memory network prediction model; P GRU Predict a single prediction value for the gated recurrent unit prediction model; The single prediction value weight calculation formula of the bidirectional long short-term memory network prediction model is as follows: Where: ξ1 is the comprehensive error value of the bidirectional long short-term memory network prediction model, ξ2 is the comprehensive error value of the gated recurrent unit prediction model; The calculation formulas for ξ1 and ξ2 are as follows: Where: and Represent the difference between the actual value and the predicted value of the sample; It represents the average of the difference between the actual value and the predicted value of all sample points; When the grey correlation coefficient method is used to select the time-varying weights of the two prediction models, the calculation formula is as follows: Where: w it is the time-varying weight of the i-th single prediction model at this moment; r i,t-1 is the grey relational value of the prediction model at the previous moment, and the formula is as follows: Where: ξ is the monotonic coefficient; m and M are the minimum and maximum absolute error values ​​of the single prediction model, respectively; Δ i (t-1) is the absolute value of the error of the i-th sub-prediction model at the t-1th time; Calculate the first prediction error value and define the error E i The formula is as follows: E i =D i -V i Where: D i represents the first comprehensive prediction result of the current carrying capacity of overhead transmission lines; V i Represents the actual value of the current carrying capacity of the overhead transmission line; E i Put any prediction model in the prediction model for secondary prediction, and get the secondary prediction value set as follows: <h2 style=";text-align:left;direction:ltr">S<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> (S1,S2,S3,...,S)<h2 style=";text-align:left;direction:ltr"> n <h2 style=";text-align:left;direction:ltr">} Where: S i is a set of secondary prediction values, i=1,2,3...n; Consider using the error judgment standard to screen the secondary prediction value; the error judgment standard calculation formula is as follows: |S i -E i |≤|E i | If the second prediction value satisfies the above formula, the second prediction value is added to the first prediction value to get the final prediction total value. The calculation formula for the final prediction total value is as follows: R i =S i +D i Where: R i The final predicted total value.

10. The self-correcting prediction method for overhead power transmission line current carrying capacity based on multi-combination modal decomposition and integrated prediction model according to claim 9 is characterized in that: In step 3.7, a comprehensive evaluation index system is considered to judge the accuracy of the prediction results; specifically, the composition of the comprehensive evaluation index system is expressed by the following formula: Where: E rmse is the root mean square error, which is the square root of the mean square error; it describes the average difference between the predicted value and the true value and is sensitive to outliers; E MAE is the mean absolute error, which describes the absolute degree of the average difference between the predicted value and the true value. Although it is not as sensitive to outliers as Ermse, since it uses absolute values, positive and negative errors will not offset each other, and it can accurately reflect the actual situation of the error; R 2 is the goodness of fit, which describes how well the prediction model fits; In summary, among the above three evaluation indicators, E rmse and E MAE The smaller the value, the higher the prediction accuracy of the prediction model. 2 The larger it is, the higher the prediction accuracy of the prediction model.

Images