A short-term prediction method for railway traction load

Abstract

The present invention discloses a short-term prediction method for railway traction load, which belongs to the field of power data processing technology. The present invention includes data cleaning, data primary decomposition, data secondary decomposition, sample entropy clustering, model training, and data prediction. The present invention combines the intrinsic time scale decomposition algorithm, the robust local mean decomposition algorithm, sample entropy clustering, the time convolution network, and the bidirectional gated recurrent network to establish a short-term prediction method for traction load based on a combination of secondary decomposition reconstruction and deep learning double-layer optimization. This method can reduce the complexity of high-frequency components. At the same time, it effectively solves the defect that existing research cannot deeply capture the real-time dynamic characteristics of the load, thereby improving the accuracy of traction load prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of power data processing, and in particular to a short-term prediction method for railway traction load. Background Art

[0002] Currently, mainstream power system load forecasting methods fall into three categories: traditional time series methods, deep learning methods, and combined forecasting methods. Chinese Patent Publication No. CN118260652A discloses a power system load forecasting model selection method, system, and load forecasting method. This method selects several candidate power system load forecasting models, trains and annotates them, and constructs a forecasting model consisting of several meta-learners. The optimal power system load forecasting model is selected based on actual load data and fitting formulas. For example, Patent Publication No. CN111600296A discloses an edge computing-based power load forecasting system and method. This method incorporates an edge computing model and combines it with real-time monitoring of the power load at each node by power load metering units to perform calculations and forecasts. Given the high volatility, sudden change, and strong dynamic characteristics of traction loads, as well as the ultra-short-term load forecasting requirements of energy management systems within the "grid-source-storage-vehicle" coordinated energy supply system, these methods suffer from difficulties in determining hyperparameters and are prone to underfitting and overfitting. Furthermore, these methods struggle to extract hidden features between nonlinear time series data, making them inappropriate for direct application to traction load forecasting. In view of this, the existing technology needs to be further improved. Summary of the Invention

[0003] In order to address the defects of the above-mentioned prior art, the present invention proposes a short-term prediction method for railway traction load. The method is based on quadratic decomposition and reconstruction and deep learning double-layer optimization, which reduces the complexity of historical load data and improves the prediction accuracy of traction load.

[0004] The technical solution of the present invention is achieved as follows:

[0005] A short-term prediction method for railway traction load includes the following steps:

[0006] Step 1: Data cleaning,

[0007] Collect historical load data, input abnormal event time period, remove abnormal data from historical load data based on the abnormal event time period, and generate historical sample data;

[0008] Step 2: Decompose the data once,

[0009] Iteratively decompose the historical sample data into the first fluctuation component and the baseline component based on the time scale until the baseline component becomes a monotonic trend component, generating eigenmode components of different frequencies;

[0010] Step 3: Secondary decomposition of data,

[0011] Iteratively decomposing the envelope signal and the pure frequency signal from the eigenmode component and calculating the second fluctuation component and the residual signal until the residual signal is monotonic data or a constant, and obtaining multiple groups of second fluctuation components after eliminating interference terms;

[0012] Step 4: Sample entropy clustering,

[0013] Calculating the sample entropy of the second fluctuation component, and clustering the plurality of groups of second fluctuation components into sample high-frequency components, sample medium-frequency components, and sample low-frequency components according to the sample entropy;

[0014] Step 5: Model training,

[0015] Collect future load data corresponding to each historical load data, feed the high-frequency component, medium-frequency component, and low-frequency component of the sample into a short-term time series prediction model respectively, and then train at least one model parameter of the short-term time series prediction model in combination with the traction load data;

[0016] Step 6: Data prediction,

[0017] Collect current load data, and obtain the current high-frequency component, current medium-frequency component, and current low-frequency component of the current load data according to steps 1 to 4. Input them into the short-term time series prediction model to obtain the traction load prediction data corresponding to the current load data.

[0018] In the present invention, in step 1, the historical load data includes photovoltaic power generation data, wind power generation data, and train energy consumption data.

[0019] In the present invention, in step 2, the baseline extraction factor F is calculated according to X t =H t +(R+F)L t = The historical sample data X t Decomposed into the first wave component H t and the baseline component L t , where R is the rotation component extraction operator, is the first fluctuation component of the kth layer, is the monotonic trend component, and n is the number of layers of data decomposition.

[0020] In the present invention, in step 3, the local mean and local envelope estimation values ​​are iteratively calculated with the eigenmode component as the initial value, and then the local mean function and the local envelope function are obtained, and then the frequency modulated signal and the zero mean signal are estimated until the frequency modulated signal is a pure frequency signal or the iteration upper limit is reached, and the envelope signal is calculated from multiple groups of local envelope estimation values.

[0021] In the present invention, in step 4, the vector sequence of the second fluctuation component is sequentially composed: X m (i)={x(i),x(i+1),…,x(i+m-1)}, 1≤i≤N-m+1, statistics m (j) and X m (i) The number of B whose distance is less than the similarity tolerance r i , sample entropy ,in, , , , 1≤i≤Nm, N is the length of the second fluctuation component, and m is the embedding dimension.

[0022] In the present invention, in step 5, the short-term time series prediction model includes a temporal convolutional network sub-model and a bidirectional gated recurrent network sub-model, the temporal convolutional network sub-model has a causal convolution unit, a dilated convolution unit and a residual block, and the bidirectional gated recurrent network sub-model has a forward GRU layer and a backward GRU layer.

[0023] In the present invention, in step 6, a high-frequency prediction component, a medium-frequency prediction component, and a low-frequency prediction component are respectively obtained based on the time convolutional network sub-model and the bidirectional gated recurrent network sub-model, and then the high-frequency prediction component, the medium-frequency prediction component, and the low-frequency prediction component are fused into traction load prediction data.

[0024] The short-term prediction method for railway traction load according to the present invention has the following beneficial effects:

[0025] First, to address the challenges of historical load data, such as high volatility and strong non-stationary time series, resulting in incomplete single-shot data decomposition and a high degree of complexity in the resulting sequence, which is unfavorable for traction load forecasting, this paper proposes a second decomposition of historical load data combined with an optimization algorithm. The ITD algorithm is first used to decompose the historical load data into a series of intrinsic mode components with different frequencies. The high-frequency intrinsic mode components are then secondarily decomposed and reconstructed using the RLMD algorithm. The sample entropy of the second fluctuation component is calculated and clustered to reduce the complexity of the high-frequency components.

[0026] At the same time, the present invention organically combines the temporal convolutional network TCN, the bidirectional gated recurrent network BiGRU and the data optimization algorithm to perform two-layer optimization, highlighting the advantages of thorough data decomposition and accurate traction load prediction, effectively solving the defect that existing research cannot deeply capture the real-time dynamic characteristics of historical load data, and improving the prediction accuracy of traction load from the model end. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flow chart of the short-term prediction method for railway traction load of the present invention;

[0028] Figure 2 This is a data flow diagram of the short-term prediction method for railway traction load of the present invention;

[0029] Figure 3 is a sequence diagram of the original signal of the present invention;

[0030] Figure 4 It is a sequence diagram of the first layer PF component of the present invention;

[0031] Figure 5 It is a sequence diagram of the second layer PF component of the present invention;

[0032] Figure 6 It is a sequence diagram of the third layer PF component of the present invention;

[0033] Figure 7 is a sequence diagram of the fourth layer PF component of the present invention;

[0034] Figure 8 It is a sequence diagram of the fifth layer PF component of the present invention;

[0035] Figure 9 is a sequence diagram of the sixth layer PF component of the present invention;

[0036] Figure 10 This is a sequence diagram of the seventh layer PF component of the present invention;

[0037] Figure 11 This is a sequence diagram of the eighth layer PF component of the present invention;

[0038] Figure 12 This is the last sequence diagram of the residual signal of the present invention. DETAILED DESCRIPTION

[0039] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0040] This paper combines the intrinsic time-scale decomposition (ITD) algorithm, the robust local mean decomposition (RLMD) algorithm, sample entropy clustering, a temporal convolutional network (TCN), and a bidirectional gated recurrent unit (BiGRU) network to establish a short-term traction load forecasting method based on a two-layer optimization combination of quadratic decomposition and reconstruction and deep learning. Specifically, it can be divided into four stages: data preprocessing, data optimization decomposition and reconstruction, clustering processing, construction of a short-term time series forecasting model, and traction load forecasting. First, historical load data on the power supply arm is collected at the high-speed railway traction station. Abnormal data is cleaned to obtain historical sample data. Second, to address the high volatility and strong randomness of the historical sample data, the intrinsic time-scale decomposition algorithm and the robust local mean decomposition algorithm are used to quickly pre-decompose the historical sample data into multiple relatively stable and regular time series mode functions (eigenmode components of different frequencies). The complex time series mode functions are further reconstructed to highlight the time series characteristics of the historical load data. Next, the sample entropy of the decomposed time series modal function is calculated and clustered using the K-means algorithm to form sample high-frequency components, sample mid-frequency components, and sample low-frequency components. Finally, the sample high-frequency components, sample mid-frequency components, and sample low-frequency components are respectively fed into a short-term time series prediction model composed of a time convolutional network sub-model and a bidirectional gated recurrent network sub-model to predict the traction load. The short-term prediction method for railway traction load of the present invention can effectively grasp the changing trend of traction load and significantly improve the accuracy of traction load prediction. Example 1

[0041] like Figure 1 and Figure 2 As shown, the short-term prediction method for railway traction load of the present invention includes the following steps.

[0042] Step 1: Data cleaning: Collect historical load data, enter the abnormal event time period, and remove abnormal data from the historical load data based on the abnormal event time period to generate historical sample data. Historical load data includes photovoltaic power generation data, wind power generation data, and train energy consumption data.

[0043] Step 2: Decompose the data once, iteratively decompose the historical sample data into the first fluctuation component and the baseline component based on the time scale until the baseline component is a monotonic trend component, and generate eigenmode components of different frequencies. For example, calculate the baseline extraction factor F, according to X t =H t +(R+F)L t = The historical sample data X t Decomposed into the first wave component H t and the baseline component L t , where R is the rotation component extraction operator, is the first fluctuation component of the kth layer is the monotonic trend component, and n is the number of layers of data decomposition. The specific steps of data decomposition are described in Example 2.

[0044] Step 3: Data secondary decomposition, iteratively decompose the envelope signal and pure frequency signal from the eigenmode component and calculate the second fluctuation component and the residual signal until the residual signal is monotonic data or a constant, and obtain multiple groups of second fluctuation components after removing the interference term. For example, iteratively calculate the local mean and local envelope estimation value with the eigenmode component as the initial value, then obtain the local mean function and the local envelope function, and then estimate the frequency modulation signal and the zero mean signal until the frequency modulation signal is a pure frequency signal or reaches the iteration limit, calculate the envelope signal from the multiple groups of local envelope estimation values, calculate the pure frequency signal from the multiple groups of frequency modulation signals, calculate the second fluctuation component and the residual signal based on the frequency modulation signal and the envelope signal, and repeat the above steps until the residual signal is a monotonic function or a constant. The specific steps of data secondary decomposition are described in Example 3.

[0045] Step 4: Sample entropy clustering, calculate the sample entropy of the second fluctuation component, and cluster multiple groups of second fluctuation components into sample high-frequency components, sample medium-frequency components and sample low-frequency components according to the sample entropy. For example, the vector sequence of the second fluctuation component is: X m (i)={x(i),x(i+1),…,x(i+m-1)}, 1≤i≤N-m+1. Define vector X m (i) With X m (j) The distance d[X m (i),X m (j)] is the absolute value of the maximum difference between the two corresponding elements, that is: d[X m (i),X m (j)]=max(|x(i+k)-x(j+k)|), 0≤k≤m-1. Statistics X m (j) and X m (i) The number of B whose distance is less than the similarity tolerance r i , sample entropy ,in, , , , 1≤i≤Nm, N is the length of the second fluctuation component, and m is the embedding dimension.

[0046] Step 5: Model training. Collect future load data corresponding to each historical load data set, feed the high-frequency, mid-frequency, and low-frequency components of the samples into a short-term time series prediction model. Combined with the traction load data, at least one model parameter of the short-term time series prediction model is trained. Preferably, the short-term time series prediction model includes a temporal convolutional network model sub-model and a bidirectional gated recurrent network sub-model. The temporal convolutional network model sub-model comprises a causal convolution unit, a dilated convolution unit, and a residual block, and the bidirectional gated recurrent network sub-model comprises a forward GRU layer and a backward GRU layer.

[0047] Step 6: Data Forecasting: Collect current load data and, following Steps 1 through 4, obtain the current high-frequency, medium-frequency, and low-frequency components of the current load data. These components are then input into the short-term time series forecasting model to obtain the corresponding traction load forecast data. For example, the high-frequency, medium-frequency, and low-frequency forecast components are derived from the temporal convolutional network sub-model and the bidirectional gated recurrent network sub-model, respectively. These components are then fused to form the traction load forecast data. Example 2

[0048] This embodiment further discloses the preferred method for one-step data decomposition of the present invention.

[0049] Step 201: Historical sample data is X t =[x1,x2,…,x t ], define the baseline extraction factor F. Perform an ITD decomposition on the historical sample data, X t =FX t +(1-F)X t =L t +H t . H t =(1-F)X t , H t is the PR component, i.e. the first fluctuation component, L t =FX t , L t is the baseline component.

[0050] Step 202: Determine historical sample data X t The extreme value x k and the corresponding time τ k , according to the extreme value x k and the corresponding time τ k Compute baseline components and baseline extraction factors, where k=1,2,…,n.

[0051] , where: t∈(τ k ,τ k +1), Fk is the extreme value of the kth baseline extraction factor. a is the gain control parameter of the inherent rotation component amplitude, which is a constant, a∈(0,1). Usually, when a is 0.5, the low-frequency component extraction effect is better.

[0052] Step 203: The obtained baseline component L t As the original data, repeat steps 201 to 203 until the baseline component becomes a monotonic trend component, generating multiple layers of first fluctuation components and monotonic trend components. t =H t +(R+F)L t = , where R is the rotation component extraction operator, is the first fluctuation component of the kth layer, is the monotonic trend component, and n is the number of layers of data decomposition. Example 3

[0053] like Figures 3 to 12 As shown, this embodiment further discloses the preferred method for secondary decomposition of data of the present invention.

[0054] Step 301: Signal processing. The eigenmode component after decomposition in step 2 is X n =[x1,x2,…,x n ], using the mirror extension method to expand the signal X n The original signal s(n) is obtained from the left and right ends of the original signal, and all local extreme points m of the original signal are found. i (i=1,2,…). Figure 3 is the sequence diagram of the original signal s(n).

[0055] Step 302: Signal solution. Calculate the local mean p based on the local extreme points i and the local envelope value α i , According to p i and α i Determine the fixed subset and use the moving average algorithm to calculate p i and α i Smoothing process makes the processed data and X n The data length is consistent, by the local mean p i and the local envelope value α i Fitting is performed to obtain the local mean function p 11 (n) and the local envelope function α 11 (n). Estimation of the frequency modulated signal f 11 (n) and the zero-mean signal z 11 (n), where .

[0056] Step 303: Data iteration. Set the target function t' and determine the frequency modulation signal f 11 (n) Is it a pure frequency signal? If not, calculate the value of the objective function t' ij , f 11 (n) is used as a new signal, and steps 301 to 303 are repeated until three consecutive objective function values ​​t' are obtained. ij , t' i(j+1) and t' i(j+2) Satisfy t' i(j+1) >t' ij And t' i(j+1) >t' i(j+2) When , the iteration stops, otherwise it continues.

[0057] , where: b n is the local envelope signal with zero baseline, for b n The average value of is the local envelope value corresponding to the nth local extreme point.

[0058] Step 304: Signal separation: According to the iteration stop condition in step 303, the envelope signal a1(n) and the pure frequency signal f1(n) are calculated in two cases.

[0059] ① When the objective function iteration stop criterion is met after p iterations, .

[0060] ②When the number of iterations reaches the set upper limit H, .

[0061] Multiply the envelope signal a1(n) and the pure frequency signal f1(n) to obtain the first product function PF1(n), PF1(n) = f1(n)a1(n).

[0062] Step 305: Extract the PF components. Separate the first PF1(n) from S(n) to obtain the first residual signal μ1(n). Use the first residual signal as the new signal and repeat steps 301 to 305 to extract all PF components until the last residual signal is monotonic or constant. Figures 4 to 11 It is a sequence diagram of PF components with different decomposition levels. Figure 12 This is the sequence diagram of the last margin signal.

[0063] Step 306: Signal generation. Decompose S(n) into q PF components and the qth residual signal μ q (n), i.e. The RLMD algorithm is used to eliminate the decomposed interference term PF1(n) and obtain the PF component (the second fluctuation component).

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A short-term prediction method for railway traction load, characterized in that: The following steps are involved: Step 1: Data cleaning, Collect historical load data, input abnormal event time period, remove abnormal data from historical load data based on the abnormal event time period, and generate historical sample data; Step 2: Decompose the data once, Iteratively decompose the historical sample data into the first fluctuation component and the baseline component based on the time scale until the baseline component becomes a monotonic trend component, generating eigenmode components of different frequencies; Step 3: Secondary decomposition of data, Iteratively decomposing the envelope signal and the pure frequency signal from the eigenmode component and calculating the second fluctuation component and the residual signal until the residual signal is monotonic data or a constant, and obtaining multiple groups of second fluctuation components after eliminating interference terms; Step 4: Sample entropy clustering, Calculating the sample entropy of the second fluctuation component, and clustering the plurality of groups of second fluctuation components into sample high-frequency components, sample medium-frequency components, and sample low-frequency components according to the sample entropy; Step 5: Model training, Collect future load data corresponding to each historical load data, feed the high-frequency component, medium-frequency component, and low-frequency component of the sample into a short-term time series prediction model respectively, and then train at least one model parameter of the short-term time series prediction model in combination with the traction load data; Step 6: Data prediction, Collect the current load data, and obtain the current high-frequency component, current medium-frequency component, and current low-frequency component of the current load data according to steps 1 to 4. Input them into the short-term time series prediction model to obtain the traction load prediction data corresponding to the current load data, where: In step 2, the baseline extraction factor F is calculated based on X t =H t +(R+F)L t = The historical sample data X t Decomposed into the first fluctuation component H t and the baseline component L t , where R is the rotation component extraction operator, is the first fluctuation component of the kth layer, is the monotonic trend component, n is the number of layers of data decomposition, In step 3, the local mean and local envelope estimation values ​​are iteratively calculated with the eigenmode component as the initial value, and then the local mean function and the local envelope function are obtained, and the frequency modulated signal and the zero mean signal are estimated again until the frequency modulated signal is a pure frequency signal or the iteration limit is reached, and the envelope signal is calculated from multiple sets of local envelope estimation values. In step 4, the vector sequence of the second fluctuation component is composed in order: X m (i)={x(i),x(i+1),…,x(i+m-1)}, 1≤i≤N-m+1, statistics m (j) and X m (i) The number of B whose distance is less than the similarity tolerance r i , sample entropy ,in, , , , 1≤i≤Nm, N is the length of the second fluctuation component, and m is the embedding dimension.

2. The short-term prediction method for railway traction load according to claim 1, characterized in that: In step 1, the historical load data includes photovoltaic power generation data, wind power generation data, and train energy consumption data.

3. The short-term prediction method for railway traction load according to claim 1, characterized in that: In step 5, the short-term time series prediction model includes a temporal convolutional network sub-model and a bidirectional gated recurrent network sub-model, the temporal convolutional network sub-model has a causal convolution unit, a dilated convolution unit and a residual block, and the bidirectional gated recurrent network sub-model has a forward GRU layer and a backward GRU layer.

4. The short-term prediction method for railway traction load according to claim 3, characterized in that: In step 6, based on the temporal convolutional network sub-model and the bidirectional gated recurrent network sub-model, high-frequency prediction component, medium-frequency prediction component, and low-frequency prediction component are obtained respectively, and then the high-frequency prediction component, medium-frequency prediction component, and low-frequency prediction component are fused into traction load prediction data.

Images