An intercity travel prediction method based on timing periodicity

Abstract

The application provides an intercity travel prediction method based on time sequence periodicity. The method comprises the following steps: S1, data preprocessing; S2, time sequence periodicity calculation; S3, prediction model adaptation; S4, model optimization and prediction output. The application calculates the time sequence periodicity through multi-index coordination, avoids the defects of single index, deeply binds the periodicity and the prediction model, solves the problems of subjective model selection and insufficient prediction accuracy in the prior art, realizes the standardized process of "periodicity calculation-model adaptation-prediction output", provides precise data support for traffic planning and transport capacity scheduling, and can be widely applied to various traffic travel prediction scenes.

Description

Technical Field

[0001] This invention relates to the field of intercity traffic data analysis and prediction technology, specifically to an intercity travel prediction method based on time-series periodicity. Background Technology

[0002] With the acceleration of urbanization and the improvement of intercity transportation networks, the demand for intercity travel is increasing day by day. Accurately identifying intercity travel volume has become a core requirement for transportation management departments and operating companies. Accurate travel volume forecasting can provide a scientific basis for capacity allocation and resource optimization, and maintain the smooth operation of the transportation system.

[0003] In existing technologies, there are two major pain points in predicting intercity travel volume: First, there is a lack of objective calculation methods for the inherent periodic characteristics of time series data. Single indicators (such as autocorrelation coefficient and periodogram method) are easily affected by noise or overestimation of local peaks, and cannot accurately identify the degree of fluctuation, periodic intensity and periodic stability of time series. Second, the selection of prediction models lacks quantitative basis and relies more on human experience, which leads to a mismatch between model performance and time series periodicity, resulting in decreased prediction accuracy and waste of computing resources.

[0004] The patent with publication number "CN116862573A" discloses a method and system for predicting short-term travel demand for intercity ride-hailing services based on incremental training. Specifically, it discloses a short-term prediction scheme for intercity ride-hailing services, including: acquiring historical ride-hailing order data, cleaning duplicate, missing, and freight data, and selecting core fields; dividing the dataset into instant orders and pre-booked orders at four different time granularities; constructing features such as day of the week, weather, holidays, and historical demand at multiple time granularities, and dividing the dataset into training and testing sets; training the model using the XGBoost algorithm, introducing an online real-time incremental training mechanism, and using known travel demand data as new samples to supplement the training (while keeping the model parameters unchanged); adding the predicted output to the pre-booked orders at the corresponding time granularity to obtain the final short-term travel demand prediction result. This aims to improve the accuracy and real-time performance of short-term ride-hailing demand prediction, but inaccuracy issues still exist.

[0005] The patent with publication number "CN118313638A" discloses a method for predicting long-term intercity ride-hailing travel demand based on SRFE-BLP-Transformer. Specifically, it discloses a prediction technology for long-term ride-hailing demand, including: acquiring historical order data for intercity ride-hailing and cleaning interfering data; dividing operating time into slices according to booking time and encoding features such as time, external environment, and historical data; filtering effective features through Spearman correlation analysis and recursive feature elimination algorithms; segmenting data by time step after normalization; capturing time dependencies using BiLSTM, adding sine and cosine combination position encoding, capturing long-distance dependencies through Transformer, and outputting long-term prediction results through linear transformation. This solves the problem that existing technologies are limited to short-term prediction and have difficulty capturing long-term dependencies, but still suffer from low accuracy.

[0006] The patent with publication number "CN117076922A" discloses a training method, prediction method, and system for intercity travel OD demand prediction models. Specifically, it discloses an OD demand prediction scheme based on multi-graph convolution, including: constructing an intercity travel database for urban agglomerations based on mobile phone signaling, weather, and POI data; using random forest to select high-importance variables, constructing round-trip relationship graphs, distance relationship graphs, and functional similarity graphs, and fusing them into a multi-graph adjacency matrix; establishing an OD-ST-AMGCN model, extracting spatial dependencies through spatial convolutional layers, capturing temporal dependencies through temporal cyclic layers, and allocating time step weights through attention layers, and outputting OD demand prediction results. It aims to solve the problems of imprecise spatiotemporal granularity and single features in large-scale regional OD demand prediction, but still suffers from strong subjectivity and insufficient accuracy.

[0007] Patent publication number "CN117829375A" discloses a method, apparatus, equipment, and medium for multi-regional demand forecasting of intercity round-trip passenger transport. Specifically, it discloses a street-level multi-regional round-trip demand forecasting technology, including: acquiring a demand city-regional relationship matrix, historical demand characteristics of the two cities, regional feature vectors, and time information sequences; using a graph attention network to learn the association weights of regional nodes within and between cities to obtain intra-city graph features and inter-city graph features; inputting the concatenated features into a gated recurrent unit, combining an attention mechanism to capture spatiotemporal patterns, and obtaining regional hidden states; performing nonlinear transformations and gating filtering on regional and temporal features to obtain temporal feature vectors; fusing the hidden states and temporal feature vectors, training and optimizing the model to output multi-regional demand forecast values, thus solving the problem of complex spatiotemporal dependencies within and between cities. However, it still suffers from strong subjectivity and insufficient accuracy.

[0008] The patent with publication number "CN113642759A" discloses a method for establishing a time-varying demand prediction model for intercity railway passengers. Specifically, it discloses a method for predicting railway passenger demand during different time periods, including: analyzing the demand distribution characteristics of different stations, originating stations, round-trip directions, weekdays / non-working days, and holidays based on historical ticket sales data of intercity railways; constructing a utility function based on stochastic utility theory, introducing variables such as travel time, whether it is a weekday, and whether it is a holiday; importing the data into Nlogit software for parameter solving; verifying the effectiveness of the model through goodness ratio test and MAPE test, and outputting the passenger demand prediction results for each time period to provide a basis for railway operation scheduling. However, it still has the problems of strong subjectivity and insufficient accuracy.

[0009] Given the shortcomings of existing technologies, such as weak anti-interference capabilities and the lack of coupling between periodicity and prediction model selection, which leads to strong subjectivity and insufficient accuracy in prediction methods, there is an urgent need for an objective periodicity calculation method based on multi-indicator synergy, and to establish a quantitative adaptation mechanism between periodicity and prediction models to achieve accurate prediction of intercity travel volume. Summary of the Invention

[0010] The purpose of this invention is to propose an intercity travel prediction method based on time-series periodicity. By coordinating the calculation of time-series periodicity through multiple indicators, the method avoids the shortcomings of single indicators and deeply binds periodicity with the prediction model. This solves the problems of subjective model selection and insufficient prediction accuracy in existing technologies, and realizes a standardized process of "periodicity calculation - model adaptation - prediction output". This provides accurate data support for traffic planning and capacity scheduling and can be widely applied to various traffic travel prediction scenarios.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0012] This invention provides a method for predicting intercity travel based on time-series periodicity, comprising the following steps:

[0013] S1, Data Preprocessing;

[0014] S2, Time-series periodicity calculation;

[0015] S3, Prediction model adaptation;

[0016] S4. Model optimization and prediction output.

[0017] As a further improvement, the data preprocessing in step S1 is as follows:

[0018] Step 1: Obtaining raw data:

[0019] Collect intercity travel time-series data for n consecutive days, with the data type being the daily intercity travel population.

[0020] Step 2: Outlier Filtering

[0021] Use the 3σ principle or other commonly used outlier handling methods to screen outliers, and remove values ​​that are abnormal due to statistical reasons based on the actual situation;

[0022] Step 3: Missing value interpolation:

[0023] If data is missing, linear interpolation or spline interpolation methods are used to fill in the gaps to ensure the integrity of the data sequence.

[0024] Step 4: Mean removal:

[0025] Calculate the mean of the preprocessed data:

[0026] ;

[0027] Where n is the sequence length, i.e., the total number of days. Let be the number of people traveling on day i.

[0028] Generate a mean-free sequence:

[0029] ;

[0030] in, Let be the number of intercity trips on day k, where k = 1, 2, ..., n.

[0031] As a further improvement, the time-series periodic calculation in step S2 is performed as follows:

[0032] Step 1: Calculation of Coefficient of Variation (CV):

[0033] Calculate the mean based on the preprocessed data. and standard deviation :

[0034] ;

[0035] in, Let n be the number of people traveling on day i, and n be the total number of days.

[0036] ;

[0037] The standard deviation The denominator of the calculation formula is n-1, and an unbiased estimation method is used to improve the calculation accuracy.

[0038] CV calculation:

[0039] ;

[0040] Step 2: Calculation of the periodic ratio R:

[0041] Frequency axis generation:

[0042] The frequency range is limited to [0, 0.5], the number of frequency sampling points is n / 2, the period corresponding to the frequency range is not less than 2 days, and the number of frequency sampling points is rounded down;

[0043] Fourier transform:

[0044] Perform a Fourier transform on the mean-valued sequence and define the basis functions:

[0045] ;

[0046] in, Let be the basis function for the m-th frequency component, where k = 1, 2, ..., n, and j is the imaginary unit. Normalized frequency;

[0047] The actual physical frequency is The fundamental frequency is m is the harmonic coefficient;

[0048] Basis function weight calculation:

[0049] ;

[0050] The basis function weights can be calculated using Euler's formula. Simplify calculations;

[0051] The frequency domain amplitude is obtained based on the weights:

[0052] ;

[0053] Where a and b are the real and imaginary parts of F(m), and the amplitude corresponds to the intensity of the frequency component;

[0054] Step 3: Calculation of Periodic Chart Values:

[0055] ;

[0056] Normalizing the frequency domain amplitude facilitates the comparison of the intensity of different frequency domain components;

[0057] Based on the conjugate symmetry of the Fourier transform, it is only necessary to calculate the periodogram values ​​from m=0 to m=n / 2, where m is rounded down.

[0058] Step 4: Power Extraction

[0059] The preset target period is x days, corresponding to a period frequency of... ,pass The corresponding m value is obtained and rounded to the nearest integer. The target periodic power corresponding to this m value is then extracted. ;

[0060] Simultaneously, iterate through all periodic values ​​to determine the maximum periodic power. ;

[0061] Step 5: R calculation:

[0062] ;

[0063] The closer the R value is to 1, the more significant the target periodic intensity.

[0064] Step 6: Calculation of ACF-C for the consistency between the autocorrelation coefficient (ACF) and the ACF:

[0065] ACF calculation:

[0066] Take the lagged term x is the target period length. The sample must contain at least 3 periods. The target period x is a preset period to be identified, including but not limited to 7 days and 30 days. Calculate the autocorrelation coefficient (ACF) for each lag term:

[0067] ;

[0068] Wherein, the numerator is the mean-reduced sequence at the current time k and the lag. time The sum of products, with the standardized term in the denominator;

[0069] The ACF value ranges from [-1, 1] and is used to characterize the correlation of time series data in different lag terms;

[0070] The larger the absolute value of the ACF, the stronger the correlation of the data in that lag term.

[0071] ACF-C calculation:

[0072] The number m of significant ACFs with an absolute value greater than 0.2 is counted, where 0.2 is the industry-recognized significance threshold. ACF consistency:

[0073] ;

[0074] The larger the ACF-C value, the more stable the periodic correlation of the data.

[0075] Step 7: Based on the thresholds of the three indicators, calculate the periodicity of intercity travel time series, as follows:

[0076] For smooth time series, CV < 0.1 is required, indicating that the time series changes in intercity travel volume are stable and without significant fluctuations.

[0077] Strongly periodic time series must satisfy CV≥0.1, R>0.8, and ACF-C>0.4, indicating that travel volume fluctuates greatly, the target periodicity is significant, and the periodic stability is strong.

[0078] Weakly periodic time series must satisfy CV≥0.1 and R>0.4 and ACF-C>0.2 and be non-strongly periodic time series, indicating that travel volume fluctuates, the target periodicity exists but the intensity is weak, and the periodic stability is generally average.

[0079] Aperiodic time series do not meet any of the above conditions, indicating that travel volume fluctuations are irregular and lack clear periodicity.

[0080] As a further improvement, the prediction model is adapted in step S3, and the specific adaptation rules are as follows:

[0081] The smooth time series model is adapted to Simple Moving Average (SMA) or Linear Regression (LR) model; the strongly periodic time series model is adapted to STL decomposition + ARIMA or SARIMA model; the weakly periodic time series model is adapted to LSTM + Attention Mechanism model; and the non-periodic time series model is adapted to XGBoost + Sliding Window model.

[0082] As a further improvement, the initial parameters of the model are set to SMA window of 7-30 days, STL periodicity parameter consistent with target x periodicity parameter, LSTM hidden layer dimension of 64-128, and non-periodic model sliding window of 7-14 days.

[0083] As a further improvement, the specific adaptation rules for model optimization and prediction output in step S4 are as follows:

[0084] Step 1: Model Training and Optimization

[0085] The adaptation model is trained based on historical data, and the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are used as quantitative evaluation indicators.

[0086] If MAPE > 10%, re-validate the data preprocessing steps or supplement features, and optimize the model parameters;

[0087] Step 2: Predict the output:

[0088] Output the intercity travel volume forecast for the next 1-30 days;

[0089] Step 3: Dynamic Iteration

[0090] The time series periodicity is recalculated every 30 days based on the new data. If the periodicity changes, the adaptive model is automatically switched to ensure stable prediction accuracy.

[0091] This invention discloses a prediction method based on the periodicity of intercity travel time series, including: acquiring intercity travel time series data for n consecutive days, and performing outlier cleaning, missing value interpolation, and mean removal preprocessing; calculating three major indicators: coefficient of variation (CV), periodogram ratio (R), and autocorrelation coefficient consistency (ACF-C), to quantify the degree of time series fluctuation, periodic intensity, and stability; adapting prediction models (simple moving average / linear regression, adapted STL decomposition + ARIMA, adapted LSTM + attention mechanism, XGBoost + sliding window model) based on indicator thresholds; evaluating model performance through MAPE, and dynamically iterating and updating the model every 30 days to achieve 1-30 day intercity travel volume prediction, solving the problems of inaccurate identification of single indicator periods and low prediction accuracy caused by subjective model selection.

[0092] The main technical differences between the patents in the background art and the present invention are reflected in the following aspects:

[0093] Targeting different objectives: This invention focuses on "short-term (1-30 days) forecasting of intercity travel volume," addressing the core pain points of inaccurate time-series period identification and subjective model selection; Patent with publication number "CN116862573A" targets "short-term demand forecasting for ride-hailing services," solving the problems of real-time data updates and model adaptation; Patent with publication number "CN118313638A" focuses on "long-term demand forecasting for ride-hailing services," overcoming the challenge of capturing long-term dependencies; Patents with publication numbers "CN117076922A" and "CN117829375A" respectively target "OD pair demand forecasting" and "multi-regional round-trip demand forecasting," emphasizing spatiotemporal dependencies and adaptation to irregular regions; Patent with publication number "CN113642759A" targets "time-varying demand forecasting for railway passengers," focusing on time-period preferences and route-differentiated demand.

[0094] The data sources differ: This invention relies on time-series data of daily intercity travel (which may come from mobile phone location big data); other patents use multi-source data, including ride-hailing order data (publication numbers “CN116862573A” and “CN118313638A”), mobile phone signaling + POI + weather data (publication number “CN117076922A”), regional relationship matrix + round-trip city demand data (publication number “CN117829375A”), and railway ticketing data (publication number “CN113642759A”).

[0095] The feature parameters are different: the core parameters of this invention are "coefficient of variation (CV), periodogram ratio (R), and autocorrelation coefficient consistency (ACF-C)", focusing on the quantification of time series periods; other patent features revolve around multi-dimensional information, such as time granularity, weather, holidays (publication number "CN116862573A"), multi-source coding features (publication number "CN118313638A"), multi-graph adjacency matrix (publication number "CN117076922A"), regional function vector (publication number "CN117829375A"), and travel time preference (publication number "CN113642759A").

[0096] The key steps differ: the core of this invention is "periodic quantization + model adaptation", with the entire process revolving around the coupling of the period and the model; other patents have different focuses in their key steps, including incremental training (publication number "CN116862573A"), BiLSTM+Transformer long-term dependency capture (publication number "CN118313638A"), multi-graph convolution + spatiotemporal dependency extraction (publication number "CN117076922A"), intra-city / inter-city graph feature learning (publication number "CN117829375A"), utility function + Nlogit parameter solving (publication number "CN113642759A").

[0097] The application scenarios differ: This invention is applicable to short- and medium-term travel volume forecasting for various intercity transportation modes (highway, railway, aviation, etc.), serving the capacity allocation and resource allocation of transportation management departments; other patent scenarios include ride-hailing operation scheduling (publication number "CN116862573A", publication number "CN118313638A"), urban agglomeration OD demand planning (publication number "CN117076922A"), intercity round-trip passenger transport multi-regional vehicle configuration (publication number "CN117829375A"), and railway time-sharing capacity arrangement (publication number "CN113642759A").

[0098] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention constructs a comprehensive analysis system for time series periodicity, employing a joint calculation method using three major indicators: coefficient of variation, periodogram ratio, and ACF consistency. This avoids the limitations of single indicators. Simultaneously, it adapts the optimal prediction model based on the periodicity analysis results, combining a differentiated strategy of "simple models for smooth time series and precise models for complex time series." Relying on a clear and operable indicator calculation process and a conventional compilation tool implementation path, it avoids the limitations of traditional time series periodicity analysis where single indicators are susceptible to noise interference and inaccurate characterization. It solves the problems of insufficient prediction accuracy, poor operability of traditional methods, narrow applicability, and difficulty in balancing prediction effect and computational efficiency caused by "model-periodity disconnection." It effectively reduces the impact of noise interference on the characterization of time series periodicity, reduces unnecessary computational costs, and avoids the problems of insufficient prediction accuracy for complex time series and wasted computational resources for smooth time series. While accurately characterizing time series periodicity, it significantly improves prediction accuracy and method operability. This innovative approach employs a multi-indicator joint anti-interference design and a differentiated model adaptation strategy, overcoming the shortcomings of traditional single-indicator methods, such as weak anti-interference capabilities and blind model selection. It effectively suppresses noise interference in time series data, balancing prediction performance with computational efficiency. This results in more accurate characterization of time series periodicity and more reliable predictions. Furthermore, the calculation methods for each indicator are clearly defined and the steps are straightforward, easily implemented using common compilers such as Python and MATLAB. Its applicability covers common modes of transportation, including intercity highways, railways, and aviation. In addition, through a differentiated strategy of "adapting simple models to smooth time series and precise models to complex time series," combined with a clear operating procedure and broad applicability scenarios, it effectively solves the problems of traditional time series analysis methods struggling to balance computational cost and prediction accuracy, and limited applicability. It demonstrates stable analysis and prediction capabilities under the dual challenges of complex noise environments and diverse traffic scenarios, ultimately providing reliable, efficient, and accurate time series analysis and prediction support for transportation planning, real-time scheduling, capacity allocation, and other transportation-related applications. Attached Figure Description

[0099] Figure 1 Logic diagram for selecting time-series periodic calculation and prediction models. Detailed Implementation

[0100] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention:

[0101] like Figure 1 As shown, this invention provides a method for predicting intercity travel based on time-series periodicity, comprising the following steps:

[0102] S1, Data Preprocessing;

[0103] S2, Time-series periodicity calculation;

[0104] S3, Prediction model adaptation;

[0105] S4. Model optimization and prediction output.

[0106] Step S1 is as follows:

[0107] Step 1: Obtaining raw data:

[0108] Collect intercity travel time-series data for n consecutive days, with the data type being the daily intercity travel population.

[0109] Step 2: Outlier Filtering

[0110] Use the 3σ principle or other commonly used outlier handling methods to screen outliers, and remove values ​​that are abnormal due to statistical reasons based on the actual situation;

[0111] Step 3: Missing value interpolation:

[0112] If data is missing, linear interpolation or spline interpolation methods are used to fill in the gaps to ensure the integrity of the data sequence.

[0113] Step 4: Mean removal:

[0114] Calculate the mean of the preprocessed data:

[0115] ;

[0116] Where n is the sequence length, i.e., the total number of days. Let be the number of people traveling on day i.

[0117] Generate a mean-free sequence:

[0118] ;

[0119] in, Let be the number of intercity trips on day k, where k = 1, 2, ..., n.

[0120] Step S2 is as follows:

[0121] Step 1: Calculation of Coefficient of Variation (CV):

[0122] Calculate the mean based on the preprocessed data. and standard deviation :

[0123] ;

[0124] in, Let n be the number of people traveling on day i, and n be the total number of days.

[0125] ;

[0126] The standard deviation The denominator of the calculation formula is n-1, and an unbiased estimation method is used to improve the calculation accuracy.

[0127] CV calculation:

[0128] ;

[0129] Step 2: Calculation of the periodic ratio R:

[0130] Frequency axis generation:

[0131] The frequency range is limited to [0, 0.5], the number of frequency sampling points is n / 2, the period corresponding to the frequency range is not less than 2 days, and the number of frequency sampling points is rounded down;

[0132] Fourier transform:

[0133] Perform a Fourier transform on the mean-valued sequence and define the basis functions:

[0134] ;

[0135] in, Let be the basis function for the m-th frequency component, where k = 1, 2, ..., n, and j is the imaginary unit. Normalized frequency;

[0136] The actual physical frequency is The fundamental frequency is m is the harmonic coefficient;

[0137] Basis function weight calculation:

[0138] ;

[0139] The basis function weights can be calculated using Euler's formula. Simplify calculations;

[0140] The frequency domain amplitude is obtained based on the weights:

[0141] ;

[0142] Where a and b are the real and imaginary parts of F(m), and the amplitude corresponds to the intensity of the frequency component;

[0143] Step 3: Calculation of Periodic Chart Values:

[0144] ;

[0145] Normalizing the frequency domain amplitude facilitates the comparison of the intensity of different frequency domain components;

[0146] Based on the conjugate symmetry of the Fourier transform, it is only necessary to calculate the periodogram values ​​from m=0 to m=n / 2, where m is rounded down.

[0147] Step 4: Power Extraction

[0148] The preset target period is x days, corresponding to a period frequency of... ,pass The corresponding m value is obtained and rounded to the nearest integer. The target periodic power corresponding to this m value is then extracted. ;

[0149] Simultaneously, iterate through all periodic values ​​to determine the maximum periodic power. ;

[0150] Step 5: R calculation:

[0151] ;

[0152] The closer the R value is to 1, the more significant the target periodic intensity.

[0153] Step 6: Calculation of ACF-C for the consistency between the autocorrelation coefficient (ACF) and the ACF:

[0154] ACF calculation:

[0155] Take the lagged term x is the target period length. The sample must contain at least 3 periods. The target period x is a preset period to be identified, including but not limited to 7 days and 30 days. Calculate the autocorrelation coefficient (ACF) for each lag term:

[0156] ;

[0157] Wherein, the numerator is the mean-reduced sequence at the current time k and the lag. time The sum of products, with the standardized term in the denominator;

[0158] The ACF value ranges from [-1, 1] and is used to characterize the correlation of time series data in different lag terms;

[0159] The larger the absolute value of the ACF, the stronger the correlation of the data in that lag term.

[0160] ACF-C calculation:

[0161] The number m of significant ACFs with an absolute value greater than 0.2 is counted, where 0.2 is the industry-recognized significance threshold. ACF consistency:

[0162] ;

[0163] The larger the ACF-C value, the more stable the periodic correlation of the data.

[0164] Step 7: Based on the thresholds of the three indicators, calculate the periodicity of intercity travel time series, as follows:

[0165] For smooth time series, CV < 0.1 is required, indicating that the time series changes in intercity travel volume are stable and without significant fluctuations.

[0166] Strongly periodic time series must satisfy CV≥0.1, R>0.8, and ACF-C>0.4, indicating that travel volume fluctuates greatly, the target periodicity is significant, and the periodic stability is strong.

[0167] Weakly periodic time series must satisfy CV≥0.1 and R>0.4 and ACF-C>0.2 and be non-strongly periodic time series, indicating that travel volume fluctuates, the target periodicity exists but the intensity is weak, and the periodic stability is generally average.

[0168] Aperiodic time series do not meet any of the above conditions, indicating that travel volume fluctuations are irregular and lack clear periodicity.

[0169] Step S3 is as follows:

[0170] (1) Smooth time series: suitable for simple moving average (SMA) or linear regression (LR) models

[0171] Adaptation criteria: The data has small fluctuations and stable trends. Simple models can meet the prediction accuracy requirements, and the computational efficiency is high, making it suitable for real-time prediction scenarios.

[0172] Key parameters: SMA window size is set to 7-30 days (adjust according to data length); LR model only uses timestamps as input features, no additional feature engineering is required.

[0173] (2) Strongly periodic time series: adapted to STL decomposition + ARIMA (or SARIMA) model

[0174] Adaptation criteria: Stable periodicity exists; STL decomposition can separate trend, periodic, and residual components; ARIMA model can accurately capture linear periodic patterns.

[0175] Key parameters: The STL period parameter is kept consistent with the target period x; the ARIMA order is optimized using the AIC criterion.

[0176] (3) Weakly periodic time series: Adapted to LSTM + attention mechanism model

[0177] Adaptation criteria: Periodic instability and nonlinearity; LSTM can capture long-term dependencies; attention mechanism can strengthen local periodic weights; balance pattern capture and anti-interference ability.

[0178] Key parameters: LSTM hidden layer dimension set to 64-128; attention window set to 2x (x is the target period).

[0179] (4) Aperiodic timing: XGBoost + sliding window model

[0180] Adaptation criteria: There is no clear pattern, the fluctuations are random, the ensemble model can adapt to nonlinear characteristics, and the sliding window can capture short-term trends;

[0181] Key parameters: The sliding window size is set to 7-14 days; input features include daily travel volume, the average of the previous 3 days, and holiday identifiers (external features such as weather and large-scale events can be added according to the scenario).

[0182] Step S4 is as follows:

[0183] Step 1: Model Training and Optimization

[0184] The adaptation model is trained based on historical data, and the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are used as quantitative evaluation indicators.

[0185] If MAPE > 10%, re-validate the data preprocessing steps or supplement features, and optimize the model parameters;

[0186] Step 2: Predict the output:

[0187] Output the intercity travel volume forecast for the next 1-30 days;

[0188] Step 3: Dynamic Iteration

[0189] The time series periodicity is recalculated every 30 days based on the new data. If the periodicity changes, the adaptive model is automatically switched to ensure stable prediction accuracy.

[0190] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for predicting intercity travel based on time-series periodicity, characterized in that, Includes the following steps: S1, Data Preprocessing; S2, Time-series periodicity calculation; S3, Prediction model adaptation; S4. Model optimization and prediction output.

2. The intercity travel prediction method based on time-series periodicity according to claim 1, characterized in that, The data preprocessing in step S1 is as follows: Step 1: Obtaining raw data: Collect intercity travel time-series data for n consecutive days, with the data type being the daily intercity travel population. Step 2: Outlier Filtering Use the 3σ principle or other commonly used outlier handling methods to screen outliers, and remove values ​​that are abnormal due to statistical reasons based on the actual situation; Step 3: Missing value interpolation: If data is missing, linear interpolation or spline interpolation methods are used to fill in the gaps to ensure the integrity of the data sequence. Step 4: Mean removal: Calculate the mean of the preprocessed data: ； Where n is the sequence length, i.e., the total number of days. Let be the number of people traveling on day i. Generate a mean-free sequence: ； in, Let be the number of intercity trips on day k, where k = 1, 2, ..., n.

3. The intercity travel forecasting method based on time-series periodicity according to claim 1, characterized in that, The time-series periodic calculation in step S2 is performed as follows: Step 1: Calculation of Coefficient of Variation (CV): Calculate the mean based on the preprocessed data. and standard deviation : ； in, Let n be the number of people traveling on day i, and n be the total number of days. ； The standard deviation The denominator of the calculation formula is n-1, and an unbiased estimation method is used to improve the calculation accuracy. CV calculation: ； Step 2: Calculation of the periodic ratio R: Frequency axis generation: The frequency range is limited to [0, 0.5], the number of frequency sampling points is n / 2, the period corresponding to the frequency range is not less than 2 days, and the number of frequency sampling points is rounded down; Fourier transform: Perform a Fourier transform on the mean-valued sequence and define the basis functions: ； in, Let be the basis function for the m-th frequency component, where k = 1, 2, ..., n, and j is the imaginary unit. Normalized frequency; The actual physical frequency is The fundamental frequency is m is the harmonic coefficient; Basis function weight calculation: ； The basis function weights can be calculated using Euler's formula. Simplify calculations; The frequency domain amplitude is obtained based on the weights: ； Where a and b are the real and imaginary parts of F(m), and the amplitude corresponds to the intensity of the frequency component; Step 3: Calculation of Periodic Chart Values: ； Normalizing the frequency domain amplitude facilitates the comparison of the intensity of different frequency domain components; Based on the conjugate symmetry of the Fourier transform, it is only necessary to calculate the periodogram values ​​from m=0 to m=n / 2, where m is rounded down. Step 4: Power Extraction The preset target period is x days, corresponding to a period frequency of... ,pass The corresponding m value is obtained and rounded to the nearest integer. The target periodic power corresponding to this m value is then extracted. ; Simultaneously, iterate through all periodic values ​​to determine the maximum periodic power. ; Step 5: R calculation: ； The closer the R value is to 1, the more significant the target periodic intensity. Step 6: Calculation of ACF-C for the consistency between the autocorrelation coefficient (ACF) and the ACF: ACF calculation: Take the lagged term x is the target period length. The sample must contain at least 3 periods. The target period x is a preset period to be identified, including but not limited to 7 days and 30 days. Calculate the autocorrelation coefficient (ACF) for each lag term: ； Wherein, the numerator is the mean-reduced sequence at the current time k and the lag. time The sum of products, with the standardized term in the denominator; The ACF value ranges from [-1, 1] and is used to characterize the correlation of time series data in different lag terms; The larger the absolute value of the ACF, the stronger the correlation of the data in that lag term. ACF-C calculation: The number m of significant ACFs with an absolute value greater than 0.2 is counted, where 0.2 is the industry-recognized significance threshold. ACF consistency: ； The larger the ACF-C value, the more stable the periodic correlation of the data; Step 7: Based on the thresholds of the three indicators, calculate the periodicity of intercity travel time series, as follows: For smooth time series, CV < 0.1 is required, indicating that the time series changes in intercity travel volume are stable and without significant fluctuations. Strongly periodic time series must satisfy CV≥0.1, R>0.8, and ACF-C>0.4, indicating that travel volume fluctuates greatly, the target periodicity is significant, and the periodic stability is strong. Weakly periodic time series must satisfy CV≥0.1 and R>0.4 and ACF-C>0.2 and be non-strongly periodic time series, indicating that travel volume fluctuates, the target periodicity exists but the intensity is weak, and the periodic stability is generally average. Aperiodic time series do not meet any of the above conditions, indicating that travel volume fluctuations are irregular and lack clear periodicity.

4. The intercity travel prediction method based on time-series periodicity according to claim 1, characterized in that, In step S3, the prediction model is adapted, and the specific adaptation rules are as follows: The smooth time series model is adapted to Simple Moving Average (SMA) or Linear Regression (LR) model; the strongly periodic time series model is adapted to STL decomposition + ARIMA or SARIMA model; the weakly periodic time series model is adapted to LSTM + Attention Mechanism model; and the non-periodic time series model is adapted to XGBoost + Sliding Window model. The initial parameters of the model are set as follows: SMA window 7-30 days, STL periodicity parameter consistent with target x periodicity parameter, LSTM hidden layer dimension 64-128, and non-periodic model sliding window 7-14 days.

5. The intercity travel prediction method based on time-series periodicity according to claim 1, characterized in that, The specific adaptation rules for model optimization and prediction output in step S4 are as follows: Step 1: Model Training and Optimization The adaptation model is trained based on historical data, and the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are used as quantitative evaluation indicators. If MAPE > 10%, re-validate the data preprocessing steps or supplement features, and optimize the model parameters; Step 2: Predict the output: Output the intercity travel volume forecast for the next 1-30 days; Step 3: Dynamic Iteration The time series periodicity is recalculated every 30 days based on the new data. If the periodicity changes, the adaptive model is automatically switched to ensure stable prediction accuracy.

Images