Traction load prediction method and device based on long and short period collaborative prediction, and storage medium
By constructing a long-short cycle collaborative prediction model, and combining a Bayesian neural network with an attention mechanism-based LSTM network, the problem of insufficient prediction accuracy in existing technologies is solved, achieving high-precision and efficient traction load prediction that can adapt to complex operating conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-03-10
AI Technical Summary
Existing traction load forecasting methods lack a long-short cycle coordination mechanism, resulting in insufficient forecasting accuracy and efficiency, and difficulty in capturing transient load changes and environmental uncertainties.
A two-layer collaborative prediction structure is adopted, combined with a Bayesian neural network, to construct a long- and short-cycle prediction model. Features are selected by Pearson correlation coefficient, and an attention-based LSTM network is introduced to dynamically adjust the sliding time window for uncertainty modeling and feature fusion.
It improves the accuracy and stability of traction load forecasting, enhances adaptability to complex operating conditions, improves the reliability and timeliness of forecasting models, and supports energy dispatching and power supply system optimization for electrified railways.
Smart Images
Figure CN121642908A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method, device, and storage medium for predicting traction load using a combined long- and short-cycle forecasting approach, belonging to the field of rail transit traction power supply technology. Background Technology
[0002] In electric locomotive traction systems, traction loads exhibit strong time non-stationarity and multimodal characteristics: they are affected not only by static factors such as train type, length, load, gradient of the running section, and operating speed, but also by dynamic factors such as temperature, humidity, traction / braking condition switching, vehicle intervals, and changes in the traction track environment. The load curve often changes abruptly between traction, braking, and no-load conditions, with large power change rates, short durations, and frequent peak-to-valley switching, making prediction highly challenging.
[0003] Currently, technologies for traction load forecasting mainly include deterministic methods based on train timetables and historical average loads, traditional time series models, and machine learning or deep learning models. However, traditional deterministic models ignore the abrupt changes in load and environmental uncertainties, making it difficult to capture transient changes. On the other hand, while conventional deep learning methods enhance nonlinear modeling capabilities, they often focus on a single forecast period, lack inter-period coordination mechanisms, and frequently fail to adequately address forecast uncertainties, resulting in insufficient forecast accuracy and adaptability. Therefore, researching a coordinated mechanism between long-period and short-period forecasts to meet the demands of modern traction power supply systems for high-precision and timely forecasting is an urgent technical problem that needs to be solved. Summary of the Invention
[0004] The purpose of this invention is to provide a traction load prediction method, device, and storage medium that combines long and short cycle prediction, which can solve the problems of existing traction load prediction methods having a single cycle, lacking a cycle coordination mechanism, and having limited prediction accuracy and efficiency.
[0005] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution.
[0006] In a first aspect, the present invention provides a traction load forecasting method based on long- and short-cycle coordinated forecasting, comprising: Acquire traction power supply system data, and construct a raw dataset of traction load based on the traction power supply system data; The original traction load dataset is preprocessed to obtain a valid dataset; Based on the effective dataset, a traction load time-series training dataset is constructed, and static and dynamic features are extracted from the traction load time-series training dataset. Combining current weather characteristics, the static and dynamic characteristics are input into a pre-built traction load prediction model, outputting a predicted traction load power value; wherein, the construction process of the traction load prediction model includes: Construct a two-layer collaborative prediction structure that includes a long-cycle prediction model and a short-cycle prediction model; A Bayesian neural network is added to the long-cycle prediction model and the short-cycle prediction model to model uncertainty.
[0007] In conjunction with the first aspect, the traction power supply system data further includes historical traction load data, real-time operating load data, line power flow data, electric locomotive operating data, track gradient data, and weather data.
[0008] In conjunction with the first aspect, the raw traction load dataset is further preprocessed, including: The original traction load dataset is cleaned to obtain a valid dataset; the data cleaning includes missing value imputation and outlier removal. Based on the traction load power time-series data in the effective dataset, calculate the traction load power change rate. The traction load power change rate The expression is: ; in, P t Indicates the first t Traction load power values at each sampling point; Indicates a time interval; Indicates the first Traction load power values at each sampling point; Indicates the rate of change of traction load power; According to the traction load power change rate The length of the sliding time window is dynamically adjusted if the traction load power change rate... Less than the preset rate of change threshold Then a 20-minute time window is used; if the traction load power change rate Not less than the preset rate of change threshold If so, a 5-minute time window is used.
[0009] In conjunction with the first aspect, further, after extracting the static and dynamic features, the process includes: Based on the extracted static and dynamic features, a feature matrix is constructed; The feature matrix was subjected to correlation analysis using the Pearson correlation coefficient to obtain the Pearson correlation coefficient matrix. Rnn Based on the Pearson correlation coefficient being greater than a preset correlation threshold r threshold The filtering criteria were used to exclude the Pearson correlation coefficient matrix. R nn Redundant features in the matrix are used to obtain the filtered feature matrix. By performing eigenvalue decomposition on the filtered feature matrix, a matrix Q composed of eigenvectors is obtained; The filtered feature matrix is projected onto matrix Q to obtain the principal component matrix; Select the top contributors from the principal component matrix based on their cumulative contribution rates. k z Each principal component characteristic quantity; The final static and dynamic features are determined based on the selected principal component features and used as input data for the traction load prediction model.
[0010] In conjunction with the first aspect, the expression for the Pearson correlation coefficient is further as follows: ; ; ; in, Indicates the first j The first feature and the second k Pearson correlation coefficient between the features; Indicates the number of samples; m represents the number of samples. i The sample at the th j The values that can be taken on each feature; Indicates the first j The mean of all samples for each feature; Indicates the first k The mean of all samples for each feature; i represents the i-th sample.
[0011] In conjunction with the first aspect, the training process of the traction load prediction model further includes: Based on the aforementioned traction load prediction model, a loss function L is constructed; wherein, the expression for the loss function L is: ; Where L represents the loss function; Indicates the number of training samples; Indicates the training sample index; Indicates the first One predicted traction load power value; Indicates the first Actual value of traction load power; Represents the regularization coefficient; In the traction load prediction model, the first... d Each trainable parameter includes the weight parameters and bias parameters of each neural network layer; D Represents the total number of trainable parameters; j represents the j-th feature; The weight coefficient of the j-th feature is represented by ; n represents the number of features; and i represents the i-th sample. The traction load prediction model is optimized based on the loss function L, and upper and lower limits of traction load, power change rate constraints, and sliding time window stability constraints are set. The optimized traction load prediction model was trained and tested using the k-fold cross-validation method.
[0012] In conjunction with the first aspect, the predicted output traction load power value further includes: The historical operating data of the traction load, the final static characteristics, and the dynamic characteristics are input into the long-term prediction model to obtain the long-term trend prediction value of the traction load; wherein, the training process of the long-term prediction model includes: By dynamically adjusting the sliding time window, daily target training samples are constructed with a step size of 1 hour. A long short-term neural network with an attention mechanism is used as the prediction model. The prediction model is trained using the daily target training samples to learn the long-term trend of traction load. Based on the output of the trained prediction model, probabilistic modeling is performed on its output layer to output the mean and variance of the long-term prediction results, and the mean and variance of the long-term prediction results are used as long-term distribution parameters. The historical operating data of the traction load, the final static and dynamic characteristics, and the long-term trend prediction value of the traction load are input into the short-cycle prediction model to obtain the future traction load power sequence; wherein, the training process of the short-cycle prediction model includes: By utilizing the historical operating data, static characteristics, and dynamic characteristics of the traction load, and integrating the long-term trend prediction value of the traction load, an intraday rolling prediction training sample is constructed. Local transient features in the intraday rolling prediction training samples are extracted using a one-dimensional convolutional neural network to obtain the feature mapping of the time series; By assigning weights to the features at each time step of the feature map in the time dimension through an attention mechanism, a weighted feature vector reflecting the short-term dynamic power change characteristics is obtained.
[0013] The weighted feature vector is probabilistically modeled to output the mean and variance of the short-cycle prediction results, and the mean and variance of the short-cycle prediction results are used as short-cycle distribution parameters. The long-term trend forecast of traction load and the future traction load power sequence are jointly optimized through a weighted fusion mechanism to output the traction load power forecast.
[0014] In conjunction with the first aspect, the long-term prediction model is further defined with a 99% confidence interval, and the short-term model is defined with a 90% confidence interval. The expression for calculating the confidence interval is: ; in, This represents the lower limit of the short-cycle traction load forecast value; This indicates the upper limit of the short-cycle traction load forecast value; This represents the lower limit of the long-cycle traction load forecast value; This indicates the upper limit of the long-cycle traction load forecast value; This indicates that the actual traction load value is predicted using a short-cycle forecasting model. The probability within the range; This indicates that when using a long-period prediction model, the actual value of the traction load falls within... The probability within the range; This represents the actual value of the future traction load.
[0015] Secondly, a traction load prediction device for combined long- and short-cycle prediction includes: The dataset construction module is used to acquire traction power supply system data and construct the original traction load dataset based on the traction power supply system data. The preprocessing module is used to preprocess the original traction load dataset to obtain a valid dataset; The feature extraction module is used to construct a traction load time-series training dataset based on the effective dataset, and to extract static and dynamic features from the traction load time-series training dataset. The load power prediction module is used to combine current weather characteristics, input the static and dynamic characteristics into a pre-built traction load prediction model, and output the traction load power prediction value; wherein, the construction process of the traction load prediction model includes: Construct a two-layer collaborative prediction structure that includes a long-cycle prediction model and a short-cycle prediction model; A Bayesian neural network is added to the long-cycle prediction model and the short-cycle prediction model to model uncertainty.
[0016] Thirdly, a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method described in the first aspect.
[0017] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: (1) Traction load is affected by multiple factors such as train operation plan, section gradient, traction and braking conditions, and exhibits both regularity and periodicity, as well as randomness and abrupt changes. To improve the traction load prediction model, a two-layer structure is adopted to achieve the synergistic fusion of information at different time scales. The long-cycle model effectively captures the overall trend of power changes in the traction system, while the short-cycle model enhances the ability to identify transient fluctuations and abrupt changes in operating conditions. The synergistic fusion of long-cycle and short-cycle predictions improves prediction accuracy and stability, enhances the adaptability and reliability of the prediction model to complex operating conditions, and provides high-precision prediction support for energy scheduling and power supply system optimization of electrified railways.
[0018] (2) To address the characteristics of long-term dependence, uneven distribution of feature weights and easy dilution of information at key moments in traction load time series, an Attention mechanism is introduced into the LSTM (Long Short-Term Memory) network, enabling the model to automatically identify and strengthen the influence weight of key time segments on load changes, thereby improving the model's ability to capture complex features, improving the feature dilution problem in traditional LSTM in long sequence modeling, and enhancing the identification and prediction performance of load change time points.
[0019] (3) Due to the high volatility and randomness of traction load, adjusting the sliding time window length in real time according to the power change rate can reduce data redundancy in the load stable phase, improve the resolution of the model in the load fluctuating phase, dynamically balance prediction accuracy and computational efficiency, improve the model's adaptability to power changes and its response to non-stationary characteristics, thereby further enhancing the timeliness and reliability of prediction results. Attached Figure Description
[0020] Figure 1 The diagram shown is a structural diagram of an electric locomotive passing through a traction substation according to an embodiment of the present invention. Figure 2 The figure shows the day-ahead load curve of the traction substation provided in an embodiment of the present invention; Figure 3 The figure shown is a voltage curve of a power supply line provided in an embodiment of the present invention; Figure 4 The figure shown is a power supply line current curve provided in an embodiment of the present invention; Figure 5 The image shown is a data diagram of electric locomotive models provided in an embodiment of the present invention; Figure 6 The image shown is a weather data map of the traction area provided in an embodiment of the present invention; Figure 7 The diagram shows a flowchart of the traction load prediction method provided in an embodiment of the present invention. Detailed Implementation
[0021] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.
[0022] Example 1 See Figure 7 This embodiment introduces a traction load forecasting method based on combined long- and short-cycle forecasting, including the following steps: Step S1: Obtain traction power supply system data and construct the original dataset of traction load based on the traction power supply system data; wherein, the traction power supply system data includes historical traction load operation data, real-time operation load data, line power flow data (power supply line voltage, current, reactive power, active power), electric locomotive operation data (operation number, type of locomotive, length of locomotive), track slope data and weather data.
[0023] For example, when an electric locomotive is running, it needs a huge traction power when going uphill, the high temperature in summer affects power consumption, and strong winds increase the load.
[0024] Step S2: Preprocess the original traction load dataset to obtain a valid dataset; Specifically, the raw traction load dataset undergoes data cleaning, including missing value imputation and outlier removal, to obtain a valid dataset. For example, fault data caused by sensor malfunctions is removed.
[0025] Based on the time-series data of traction load power in the valid dataset, the rate of change of traction load power is calculated. The expression is: ; in, P t Indicates the first t Traction load power values at each sampling point; Indicates a time interval; Indicates the first Traction load power values at each sampling point; Indicates the rate of change of traction load power; Based on the traction load power change rate The length of the sliding time window is dynamically adjusted if the traction load power change rate... Less than the preset rate of change threshold If the traction load power change rate is... Not less than the preset rate of change threshold If so, a 5-minute time window is used.
[0026] Step S3: Based on the effective dataset, construct the traction load time series training dataset, and extract static and dynamic features from the traction load time series training dataset; Specifically, the process of constructing the traction load time series training dataset includes: Step S31: Construct a feature matrix based on the extracted static and dynamic features; Static features include feature parameters that do not change or change at a very low frequency during the prediction period, specifically represented as follows: (2) Where M represents the model category of electric locomotive; Indicates the length of the electric locomotive train; Indicates the length of the gradient used for train operation; Indicates the gradient of the track on which the train runs; This indicates the type of road section, including flat roads, slopes, curves, etc. Indicates the driving section number; A set representing static features.
[0027] Dynamic features include real-time monitoring data that changes significantly over time, as specifically represented below: (3) (4) (5) in, Indicates the active power of the traction load; Indicates the reactive power of the traction load; Indicates the traction load condition category; This indicates the real-time power supply of the overhead contact line feeder; This indicates the real-time temperature of the operating section; This indicates the real-time relative humidity of the operating section, which can affect line resistance and locomotive air conditioning load; Indicates three-phase current; Indicates the current in phase A; Indicates the current in phase B; Indicates the C-phase current; Indicates three-phase voltage; Indicates the voltage of phase A; This represents the voltage of phase B; Indicates the voltage of phase C; A set representing dynamic characteristics.
[0028] Furthermore, all static and dynamic features are combined into a feature matrix, where each row represents a sample data point and each column represents a feature. The expression for this matrix is: (6) in, Represents the characteristic matrix; Indicates the first i The sample at the th j Values can be taken in each feature dimension; m Indicates the number of samples; n Indicates the number of features.
[0029] Step S32: Perform correlation analysis on the feature matrix using the Pearson correlation coefficient, and calculate the correlation coefficient between two features, i.e., the Pearson correlation coefficient matrix. R nn , where the Pearson correlation coefficient matrix R nn The expression is: (7) in, R nn express symmetric matrix; This represents the Pearson correlation coefficient between the first feature and itself; The Pearson correlation coefficient between the first feature and the nth feature; The Pearson correlation coefficient between the nth feature and the 1st feature is represented. This represents the Pearson correlation coefficient between the nth feature and itself; Let represent the Pearson correlation coefficient between the j-th feature and itself; Specifically, the Pearson correlation coefficient is used to calculate the linear dependence between features (static and dynamic features). The expression for the Pearson correlation coefficient is as follows: (8) (9) (10) in, Indicates the first j The first feature and the second k The Pearson correlation coefficient between the features is used to determine key and redundant features and to provide a basis for subsequent principal component analysis; m represents the number of samples. Indicates the first i The sample at the th j The values that can be taken on each feature; Indicates the first j The mean of all samples for each feature; Indicates the first k The mean of all samples for each feature; i represents the i-th sample; Indicates the first i The sample at the th k The values that can be taken on each feature.
[0030] It should be noted that, if , indicating a positive correlation, if , indicating a negative correlation, if , indicating no correlation, and The closer to 1, the stronger the correlation.
[0031] Step S33, based on Greater than the preset correlation threshold r threshold The selection criteria are used to extract highly relevant features, i.e., to remove redundant features, resulting in a filtered feature matrix.
[0032] Step S34: By performing eigenvalue decomposition on the filtered feature matrix, a matrix Q composed of eigenvectors is obtained. The filtered feature matrix is then projected onto matrix Q to obtain the principal component matrix. Based on the cumulative contribution rate, the top-performing features are selected from the principal component matrix. k z Each principal component characteristic quantity; Specifically, by performing eigenvalue decomposition on the selected feature matrix C, we obtain the eigenvalue matrix V and the matrix Q composed of eigenvectors. The eigenvalue matrix V is obtained from the eigenvalue matrix C. k column vector q k For the first k The orientation of each principal component is used to identify and transform redundant information between features to achieve dimensionality reduction, and its expression is: (11) (12) (13) in, V Represents the eigenvalue matrix, which is composed of eigenvalues. to The diagonal matrix constructed from the main diagonal elements corresponds to the variance of its principal components; The transpose matrix representing the eigenvectors; This represents the first eigenvalue; This represents the nth eigenvalue; This indicates the first primitive feature in the first principal component. Load on; This indicates that the first primitive feature is in the nth principal component. Load on; This indicates that the nth original feature is in the first principal component; This indicates that the nth primitive feature is in the nth principal component. The load on it.
[0033] It should be noted that, generally, before selection k z The principal component with the largest eigenvalue is used for modeling, while k z The value is determined based on the variance contribution rate. The expression for selection is: (14) in, Indicates the preceding k z Cumulative variance contribution rate of each principal component; n Indicates the number of features; Indicates the first i z Each feature value.
[0034] when Reaching the preset contribution rate threshold η threshold At this time, take this moment k z Value, and retain the previous value. k z Each principal component characteristic quantity; k z Indicates the number of principal components; Indicates the first Each feature value.
[0035] Step S35: Determine the final static and dynamic features based on the selected principal components, and input them into the traction load prediction model.
[0036] Furthermore, if the i-th data point in the feature set x in formula (8) When the traction load condition is in operation, the types of traction load conditions include no-load condition, traction condition and braking condition.
[0037] The steps for identifying this traction condition include: Step S331: Based on the collected historical traction load power data, construct a traction load curve, extract the continuous period with small power fluctuation range in the traction load curve as the no-load operation period, and calculate the average active power of the traction load during the no-load operation period. Its expression is: (15) in, This refers to the start time of the no-load operating condition time sequence. This is the end time of the no-load operating condition time sequence. This represents the average active power of the traction load during the no-load operation period.
[0038] Step S332: Calculate the power jump between two adjacent time lengths, the expression of which is: (16) in, ΔP(t) represents the active power value of the traction load at time t-1; ΔP(t) represents the power jump variable used to capture the instant when the operating condition begins or ends.
[0039] It should be noted that ΔP(t) is very small during smooth operation, but it will suddenly increase when the locomotive starts to traction or brake.
[0040] Step S333, if ,and The condition is determined to be traction mode, and the start time of this mode is set to... ;like ,and This is determined to be a braking condition, and the start time of this condition is set to... ;like ,and The traction condition is determined to have ended, and the end time for this condition is set to [time value missing]. ;like ,and The braking condition is determined to have ended, and the end time of this condition is set to [time value missing]. .
[0041] Step S4: Combining current weather characteristics, input static and dynamic features into the pre-constructed traction load prediction model and output the traction load power prediction value; wherein, a two-layer collaborative prediction structure including a long-cycle prediction model and a short-cycle prediction model is constructed. In view of the high volatility and randomness of traction load, a Bayesian neural network is added to the long-cycle prediction model and the short-cycle prediction model to perform uncertainty modeling, thereby realizing the quantitative representation of the reliability of the prediction results.
[0042] Specifically, the training process for the traction load prediction model includes: Step S41: Based on the traction load prediction model, construct the loss function L (i.e., the optimization objective during model training); where the expression for the loss function L is: (17) Where L represents the loss function; Indicates the number of training samples; Represents the training sample index, i.e., the first... Each training sample corresponds to a set of predicted and actual traction load power values; Indicates the first i One predicted traction load power value; Indicates the first i Actual value of traction load power; Represents the regularization coefficient; In the traction load prediction model, the first... d Each trainable parameter includes the weight parameters and bias parameters of each neural network layer; D This represents the total number of trainable parameters. The trainable parameters are updated by minimizing the loss function, enabling the model to learn the changing patterns of traction load from the training data. Indicates mean square error; This indicates regularization.
[0043] Step S42: Optimize the traction load prediction model based on the loss function L, and set upper and lower limits for traction load, power change rate constraints, and sliding time window stability constraints. The constraints are as follows: Load upper and lower limit constraints: (18) in, This indicates the minimum traction load power. P max This indicates the maximum traction load power.
[0044] Power change rate constraint: (19) in, This represents the threshold for the rate of change of traction load power. It represents the instantaneous rate of change of the active power of the traction load at time t.
[0045] Sliding time window stability constraints: (20) in, This represents the standard deviation of the traction load power sequence within the sliding time window; w This is the length of the sliding window, i.e., the number of sampling points within the sliding time window; This is the upper limit threshold for stability; Indicates the index of the sampling time point within the sliding window; This indicates the first time within the sliding time window. The time for each sampling point Indicates the first time within the sliding time window The predicted traction load power at each sampling point; i represents the cyclic index of the sampling point within the sliding window. Indicates in The predicted value of traction load power at any given time.
[0046] Step S43: Use the k-fold cross-validation method to train and test the optimized traction load prediction model to evaluate the model's generalization ability.
[0047] The collected dataset was divided into training and test sets. Multi-window parallel training was performed on the training set, and dynamic k-fold cross-validation was used, with the k value dynamically adjusted according to the power change rate. Evaluation metrics for the test set included RMSE, MAE, and R0. 2 With confidence coverage. Input electric locomotive operation data, weather data, and line data into the trained prediction model, and output the traction load prediction value and its confidence interval.
[0048] Furthermore, the process of outputting the predicted traction load power and training the model includes the following steps: Step S51: Input the historical operating data of traction load, the final static characteristics, and the dynamic characteristics into the long-term prediction model to obtain the long-term trend prediction value of traction load; wherein, the training process of the long-term prediction model includes: Based on the dynamic sliding time window method described above, daily target training samples are constructed with a step size of 1 hour to learn the overall trend of traction load changes; An Attention-LSTM (Long Short-Term Neural Network with an attention mechanism) is used as the prediction model. The model is trained using the daily target training samples to learn the long-term trend of traction load. The weight calculation expression for the attention mechanism is as follows: (twenty one) in, w t express t Hidden state of time Attention weights; v This means mapping the attention vector to the projection vector of the weight scalar, i.e., the weight parameters to be trained; W represents the transpose of the projection vector. h The linear transformation matrix represents the encoder's hidden state, which is the weight parameter to be trained. Indicates the encoder is in t Hidden state at any moment; W s This represents the linear transformation matrix of the query vector, i.e., the weight parameters to be trained; s Indicates the current state of the decoder; This represents the time step index, used to traverse all the time steps in the input sequence. T Each time step is used to normalize the attention weights; express The hidden state at any given moment.
[0049] Based on the calculated attention weights, a context vector is generated by weighted summing of the hidden states of the encoder at all time steps. c Its expression is: (twenty two) in, c This represents the weighted summation of the context vector, which integrates information from the hidden states at all time steps. t Indicates the current moment.
[0050] Based on the output of the trained prediction model, probabilistic modeling is performed on its output layer to output the mean and variance of the long-term prediction results, and these mean and variance are used as long-term distribution parameters. The long-term prediction model is set with a 99% confidence level, and its confidence interval is calculated as follows: (twenty three) in, This represents the lower limit of the long-cycle traction load forecast value; This indicates the upper limit of the short-cycle traction load forecast value; This indicates that the actual traction load value is predicted using a short-cycle forecasting model. The probability within the range; This represents the actual value of the future traction load.
[0051] Furthermore, the confidence interval of the prediction model is obtained by passing the predicted mean output by the Bayesian neural network. With the predicted standard deviation Calculated during the inference phase, the confidence interval is used to characterize the confidence range of the prediction result. Although the confidence interval does not directly participate in the backpropagation training of the model, its construction depends on the uncertainty parameter (prediction variance) learned during training. The confidence interval reflects the risk level of traction load forecasting, providing a basis for scheduling decisions, safety margin settings, and abnormal condition identification in the traction power supply system, thereby significantly improving the reliability and interpretability of the forecast results in engineering applications. This is achieved through the mean of long-term forecast results. With variance Calculate the upper and lower limits of the confidence interval .
[0052] For example: Model output =100MW, =2MW. Therefore, the 99% confidence interval is [100 - 2.576*2, 100 + 2.576*2] = [94.85, 105.15]MW. This means the model is 99% confident that the actual load will fall within this interval.
[0053] Step S52: Input the historical operating data of traction load, the final static and dynamic characteristics, and the long-term trend prediction value of traction load into the short-cycle prediction model to obtain the future traction load power sequence; wherein, the training process in the short-cycle prediction model includes: By utilizing historical operating data, static and dynamic characteristics of traction load, and integrating the long-term trend forecast value of traction load, an intraday rolling forecast training sample is constructed. Local transient features in the intraday rolling prediction training samples are extracted using a one-dimensional convolutional neural network to obtain the feature mapping of the time series; By assigning weights to the features at each time step of the feature map in the time dimension through an attention mechanism, a weighted feature vector reflecting the short-term dynamic power change characteristics is obtained.
[0054] Probabilistic modeling is performed on the weighted eigenvectors to output the mean of the short-cycle prediction results. and variance As a short-term distribution parameter, and based on the distribution parameter, the confidence interval of the short-term prediction result is determined; wherein, the short-term prediction model is set with a 90% confidence level, and its confidence interval calculation expression is: (twenty three) in, This represents the lower limit of the short-cycle traction load forecast value; This indicates the upper limit of the short-cycle traction load forecast.
[0055] Furthermore, the upper and lower limits of the confidence interval are calculated using the mean and variance of the long-term prediction results. .
[0056] Step S53: The long-term trend prediction value of traction load and the future traction load power sequence are optimized by weighted fusion mechanism to output the traction load power prediction value.
[0057] Specifically, the short-cycle forecasting model is triggered every 5 minutes to predict the traction load sequence for the next 5-30 minutes. The results are compared and fused with the output of the long-cycle forecasting model, and a weighted fusion mechanism is used to collaboratively optimize the forecast results. For example, at very close time points (such as the next 5 minutes), the predictions of the short-cycle model are given more weight; at slightly further time points (such as the next 25-30 minutes), the trend of the long-cycle model is given more weight. Through this intelligent weighting, the predicted traction load power value is output.
[0058] Step S53: The long-term trend prediction value of traction load and the future traction load power sequence are optimized by weighted fusion mechanism to output the traction load power prediction value.
[0059] Specifically, the short-cycle forecasting model is triggered every 5 minutes to predict the traction load sequence for the next 5-30 minutes. The results are compared and fused with the output of the long-cycle forecasting model, and a weighted fusion mechanism is used to collaboratively optimize the forecast results. For example, at very close time points (such as the next 5 minutes), the predictions of the short-cycle model are given more weight; at slightly further time points (such as the next 25-30 minutes), the trend of the long-cycle model is given more weight. Through this intelligent weighting, the predicted traction load power value is output. Example 2
[0060] See Figure 1 Based on the actual operating data of a traction substation, predictive analysis is performed. Using the fusion of long and short cycle collaborative prediction method proposed in this invention, day-ahead and intraday rolling predictions of electric locomotive traction loads are conducted. The implementation process of this method begins with the collection of multi-source data, including: Figure 2 The data shown includes historical traction load operating data and real-time traction load operating data measured in real-time by the traction substation SCADA system; such as... Figure 3 and Figure 4 The power flow data shown includes line voltage and current; such as... Figure 5 The data for the electric locomotive shown is as follows.
[0061] The electric locomotive model data for this section includes four types, denoted by 1, 2, 3, and 4, each available in 8-car and 16-car formations. To clearly illustrate some of the raw data, traction power supply system operating data for certain times are provided in Table 1 below: Table 1
[0062] In addition, the slope data of the operating section and such Figure 6 Weather data for the area where the traction depot is located, as shown, is also included in the data collection scope. After data collection is completed, through key steps such as data cleaning, feature extraction, and uncertainty modeling, high-precision traction load time-series prediction results are finally output through the coordinated operation of long- and short-period prediction models.
[0063] In addition, the slope data of the operating section and such Figure 6Weather data for the area where the traction depot is located, as shown, is also included in the data collection scope. After data collection is completed, through key steps such as data cleaning, feature extraction, and uncertainty modeling, high-precision traction load time-series prediction results are finally output through the coordinated operation of long- and short-period prediction models. Example 3
[0064] This embodiment introduces a traction load prediction device that combines long and short cycle prediction, including: The dataset construction module is used to acquire traction power supply system data and construct the original traction load dataset based on the traction power supply system data. The preprocessing module is used to preprocess the original traction load dataset to obtain a valid dataset; The feature extraction module is used to construct a traction load time-series training dataset based on the effective dataset, and to extract static and dynamic features from the traction load time-series training dataset. The load power prediction module is used to combine current weather characteristics, input the static and dynamic characteristics into a pre-built traction load prediction model, and output the traction load power prediction value; wherein, the construction process of the traction load prediction model includes: Construct a two-layer collaborative prediction structure that includes a long-cycle prediction model and a short-cycle prediction model; A Bayesian neural network is added to the long-cycle prediction model and the short-cycle prediction model to model uncertainty. Example 4
[0065] A computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the steps of the method described in Embodiment 1.
[0066] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0067] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0068] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0069] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0070] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A traction load prediction method of long-short cycle collaborative prediction, characterized by, The method comprises the following steps: acquiring traction power supply system data and constructing a traction load original data set according to the traction power supply system data; preprocessing the traction load original data set to obtain an effective data set; based on the effective data set, constructing a traction load time series training data set, and extracting static features and dynamic features from the traction load time series training data set; combining the current weather features, inputting the static features and dynamic features into a pre-constructed traction load prediction model, and outputting a traction load power prediction value; wherein the construction process of the traction load prediction model comprises: constructing a double-layer collaborative prediction structure including a long-period prediction model and a short-period prediction model; adding a Bayesian neural network to the long-period prediction model and the short-period prediction model for uncertainty modeling.
2. The long-short cycle collaborative prediction of tractive effort prediction method of claim 1, wherein, The traction power supply system data includes traction load historical operation data, real-time operation load data, line power flow data, electric locomotive operation data, operating section slope data, and weather data.
3. The long-short cycle collaborative prediction of tractive effort prediction method of claim 2, wherein, The preprocessing of the traction load original data set comprises: data cleaning of the traction load original data set to obtain an effective data set; wherein the data cleaning includes missing value filling and outlier removal. Based on the traction load power time series data in the effective data set, a traction load power change rate is calculated , and an expression of the traction load power change rate is ； wherein, P t represents the traction load power value of the first t sample point; represents the time interval; represents the traction load power value of the first sample point; represents the traction load power change rate; According to the traction load power change rate The length of the sliding time window is dynamically adjusted, if the traction load power change rate is less than a preset change rate threshold , a time window length of 20 minutes is adopted; if the traction load power change rate is not less than the preset change rate threshold , a time window length of 5 minutes is adopted.
4. The long-short cycle collaborative prediction of tractive effort prediction method of claim 2, wherein, After extracting the static features and dynamic features, the method comprises the following steps: based on the extracted static features and dynamic features, constructing a feature matrix; The feature matrix was subjected to correlation analysis using the Pearson correlation coefficient to obtain the Pearson correlation coefficient matrix. R nn Based on the Pearson correlation coefficient being greater than a preset correlation threshold r threshold The filtering criteria were used to exclude the Pearson correlation coefficient matrix. R nn Redundant features in the matrix are used to obtain the filtered feature matrix. performing eigenvalue decomposition on the filtered feature matrix to obtain a matrix Q composed of eigenvectors; projecting the filtered feature matrix onto the matrix Q to obtain a principal component matrix; selecting, from the principal component matrix, the first k z principal component feature quantity according to a cumulative contribution rate determining the final static features and dynamic features as input data of the traction load prediction model according to the selected principal component feature quantities.
5. The long-short cycle coordinated prediction's tractive load prediction method according to claim 4, characterized in that, The expression of the Pearson correlation coefficient is: ； ； ； in, Indicates the first j The first feature and the second k Pearson correlation coefficient between the features; Indicates the number of samples; m represents the number of samples. i The sample at the th j The values that can be taken on each feature; Indicates the first j The mean of all samples for each feature; Indicates the first k The mean of all samples for each feature; i represents the i-th sample.
6. The long-short cycle coordinated prediction's tractive load prediction method according to claim 1, characterized in that, The training process of the traction load prediction model comprises: based on the traction load prediction model, constructing a loss function L; wherein the expression of the loss function L is: ； Where L represents the loss function; Indicates the number of training samples; Indicates the training sample index; Indicates the first One predicted traction load power value; Indicates the first Actual value of traction load power; Represents the regularization coefficient; In the traction load prediction model, the first... d Each trainable parameter includes the weight parameters and bias parameters of each neural network layer; D Represents the total number of trainable parameters; j represents the j-th feature; The weight coefficient of the j-th feature is represented by ; n represents the number of features; and i represents the i-th sample. optimizing the traction load prediction model according to the loss function L, and setting traction load upper and lower limit constraints, power change rate constraints, and sliding time window stability constraints; training and testing the optimized traction load prediction model using the k-fold cross-validation method.
7. The long-short cycle collaborative prediction's tractive load prediction method according to claim 4, characterized in that, The output of the traction load power prediction value comprises: inputting the traction load historical operation data, the final static features and dynamic features into the long-period prediction model to obtain a traction load long-term trend prediction value; wherein the long-period prediction model training process comprises: by means of the dynamic adjustment of the sliding time window, constructing daily target training samples with 1 hour as the step; using a long short-term neural network with an attention mechanism as a prediction model, training the prediction model using the daily target training samples to learn the long-term trend of the traction load; based on the output of the trained prediction model, probabilistic modeling is performed on the output layer, the mean and variance of the long-period prediction result are output, and the mean and variance of the long-period prediction result are taken as long distribution parameters; The traction load historical operation data, the final static characteristics and dynamic characteristics, and the long-term trend prediction value of the traction load are input into the short-period prediction model to obtain a future traction load power sequence; wherein the training process in the short-period prediction model includes: The traction load historical operation data, the static characteristics and the dynamic characteristics are utilized, and the long-term trend prediction value of the traction load is fused to construct an intra-day rolling prediction training sample; Local transient characteristics in the intra-day rolling prediction training sample are extracted by a one-dimensional convolutional neural network to obtain a feature mapping of a time sequence; Each time step feature of the feature mapping in the time dimension is assigned a weight by an attention mechanism to obtain a weighted feature vector reflecting short-term dynamic power change characteristics; The weighted feature vector is probabilistically modeled to output a mean and a variance of a short-period prediction result, and the mean and the variance of the short-period prediction result are taken as short distribution parameters; The long-term trend prediction value of the traction load and the future traction load power sequence are synergistically optimized by a weighted fusion mechanism to output a traction load power prediction value.
8. The long-short cycle collaborative prediction's tractive load prediction method according to claim 7, characterized in that, The long-period prediction model sets a 99% confidence interval, and the short-period model sets a 90% confidence interval; The calculation expression of the confidence interval is: ； wherein, represents a lower limit of the short-period tractive effort prediction value; represents an upper limit of the short-period tractive effort prediction value; represents a lower limit of the long-period tractive effort prediction value; represents an upper limit of the long-period tractive effort prediction value; represents a probability that the tractive effort true value falls in the range of when predicted using the short-period prediction model; represents a probability that the tractive effort true value falls in the range of when predicted using the long-period prediction model; represents the future tractive effort true value.
9. A long-short cycle cooperative prediction-based traction load prediction device characterized by comprising: It includes: A data set construction module configured to obtain traction power supply system data and construct a traction load original data set based on the traction power supply system data; A preprocessing module configured to preprocess the traction load original data set to obtain an effective data set; A feature extraction module configured to construct a traction load time series training data set based on the effective data set, and extract static characteristics and dynamic characteristics from the traction load time series training data set; A load power prediction module configured to input the static characteristics and the dynamic characteristics into a pre-constructed traction load prediction model in combination with current weather characteristics to output a traction load power prediction value; wherein the construction process of the traction load prediction model includes: Constructing a double-layer collaborative prediction structure including a long-period prediction model and a short-period prediction model; Adding a Bayesian neural network in the long-period prediction model and the short-period prediction model to perform uncertainty modeling.
10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by a processor, implements the steps of the method of any one of claims 1-8.