Energy-saving strategy generation method based on rail transit operation mode recognition
By using multi-dimensional feature extraction and clustering algorithms to identify rail transit operation modes, a full-scenario energy-saving strategy is generated, which solves the problem of low energy-saving scheduling accuracy in existing technologies and achieves precise energy saving and safety assurance in complex scenarios.
Patent Information
- Application Number
- CN202511948523.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-23
- Publication Date
- 2026-01-23
- Estimated Expiration
- 2045-12-23
AI Technical Summary
Existing energy-saving technologies for rail transit are poorly adaptable to complex operating scenarios, unable to achieve precise energy saving, and fail to make full use of multi-source data, especially unstructured data, resulting in low accuracy of energy-saving scheduling.
By collecting multi-source data and extracting multi-dimensional features such as time, passenger flow, and operation, the improved K-Means++ and DBSCAN algorithms are used to identify typical and abnormal operating modes, generate a variety of energy-saving strategies covering all scenarios, and execute or trigger early warnings in real time.
It achieves precise energy saving in different operating scenarios, improves the targeting and accuracy of energy-saving scheduling, makes full use of the value of multi-source data, and ensures operational safety and robustness.
Smart Images

Figure CN121390799A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of rail transit operation, in particular to a kind of energy-saving strategy generation method based on rail transit operation mode identification. BACKGROUND
[0002] As the backbone system of urban public transport, urban rail transit has the advantages of large capacity, high speed, safety and punctuality, but it is also a big energy consumer. With the acceleration of urbanization and the expansion of rail transit network, the proportion of rail transit energy consumption in the total energy consumption of the city is increasing year by year, and energy saving has become an important challenge for rail transit operating companies.
[0003] Currently, energy-saving technologies in the field of rail transit are mainly divided into two categories: one is the energy-saving technology at the equipment level, such as using high-efficiency traction transmission system, regenerative energy recovery device, etc.; the other is the energy-saving strategy at the operation level, such as optimizing the operation curve, adjusting the train organization, etc. The former requires a large amount of hardware investment, and the latter is more economical and flexible.
[0004] In terms of operational energy saving, existing technologies mostly use fixed strategies or strategies based on simple rules. For example, a single energy-saving operation diagram is used during off-peak hours, and the regular operation diagram is restored during peak hours. This method is simple and easy to implement, but it has poor adaptability and cannot cope with the complex and variable conditions in actual operation. Rail transit operating scenarios are diverse, such as different scenarios on weekdays, peak, off-peak, weekends, and special event days, with significant differences in passenger flow, train density, and power grid load. Fixed strategies cannot achieve the best energy-saving effect in all scenarios.
[0005] In recent years, with the development of intelligent rail transit, some data-based energy-saving methods have been proposed. For example, Chinese Patent Publication No. CN111831658A proposes a rail transit big data analysis method and system that can preliminarily analyze and predict energy consumption data. However, such methods are mostly limited to data analysis and visualization and fail to directly translate the analysis results into executable energy-saving strategies. In addition, existing methods mostly focus on single-dimensional (such as energy consumption) data analysis and lack comprehensive multi-dimensional identification of operating modes, especially failing to fully consider the internal relationship between operating modes and energy consumption characteristics. At the same time, existing data-driven methods mostly rely on structured data, but have insufficient processing capacity for a large amount of unstructured data in the rail transit system, resulting in insufficient utilization of data value and low energy-saving scheduling precision. SUMMARY
[0006] To address the above deficiencies in the prior art, the present application provides a kind of energy-saving strategy generation method based on rail transit operation mode identification, which solves the problem of low energy-saving scheduling precision in the prior art.
[0007] In order to achieve the above object, the technical scheme adopted by the present application is as follows: a kind of energy-saving strategy generation method based on rail transit operation mode recognition, comprising the following steps: S1, collect and pretreat multi-source data, and extract time characteristics, passenger flow characteristics, operation characteristics, energy consumption characteristics and vehicle and environmental characteristics from the pretreated multi-source data; S2, standardize each feature to obtain standardized features and form a feature data set; S3, according to the probability distribution of the sample points in the feature data set, select multiple initial cluster centers from the feature data set; S4, each sample point is first assigned to the cluster where the nearest initial cluster center is located, and the cluster center is corrected, the cluster is updated to obtain a typical operation mode set; S5, classify each sample point in the feature data set, find noise points, and mark the noise points as abnormal operation modes; S6, for each cluster in the typical operation mode set, first label its operation mode, and then customize multiple energy-saving operation strategies covering all scenarios based on each labeled operation mode; S7, real-time data acquisition and standardization, matching the typical operation mode executes the corresponding energy-saving strategy, and matching the abnormal operation mode triggers the early warning strategy.
[0008] Further, the multi-source data in S1 includes train operation data, energy consumption data, passenger flow data and equipment state data.
[0009] Further, S3 includes the following steps: S31, randomly select a sample point from the feature data set as an initial cluster center, and place the initial cluster center in the initial cluster center set; S32, calculate the shortest distance between each sample point in the feature data set and the initial cluster center in the initial cluster center set; S33, calculate the probability distribution according to the shortest distance corresponding to each sample point in the feature data set; S34, according to the probability distribution, use the roulette method to select the next sample point from the feature data set as a new initial cluster center, and continue to include the new initial cluster center in the initial cluster center set; S35, repeat steps S32-S34 until K initial cluster centers are selected, and K is the upper limit number of initial cluster centers.
[0010] Further, the calculation formula of the shortest distance in S32 is: , Where D(x i ) is the shortest distance of the i-th sample point in the feature data set, x iFor the i-th sample point, c j For the j-th initial cluster center in the current initial cluster center set, k is the number of initial cluster centers in the current initial cluster center set, j is the number of initial cluster centers, || || is the Euclidean distance, min is the minimum value of k distances.
[0011] Further, the formula for calculating the probability distribution in S33 is: Where P(x i ) is the probability distribution of each sample point in the feature data set, D is the number of sample points in the feature data set, x n is the n-th sample point in the feature data set, D(x n ) is the shortest distance of the n-th sample point in the feature data set, and D is the feature data set.
[0012] Further, S4 includes the following steps: S41, calculate the distance between the sample points in the feature data set and the plurality of initial cluster centers, select the initial cluster center corresponding to the shortest distance, and add the corresponding sample point to the cluster; S42, traverse the sample points in the feature data set through step S41 to obtain K initial clusters; S43, take the centroid of each initial cluster as a new initial cluster center, and re-execute S41 and S42 with the new initial cluster center until convergence, to obtain K clusters; S44, form a typical operation mode set from the clusters.
[0013] Further, S5 includes the following steps: S51, randomly select an unvisited sample point p in the feature data set, and obtain the ϵ-neighborhood point set of the sample point p, where ϵ is the neighborhood radius; S52, when the number of points in the ϵ-neighborhood point set of the sample point p is greater than or equal to the minimum number of neighborhood points required by the core point minPts, the sample point p is marked as a core point p; S53, add all sample points in the neighborhood point set of the core point p to the "neighbor set"; S54, randomly select a sample point in the "neighbor set" to obtain the current neighbor point Q; S55, determine whether the current neighbor point Q is visited; S56, when not being accessed, checking the number of points in the epsilon-neighborhood point set of the current neighbor point Q, when greater than or equal to the minimum number of neighbor points required by the core point minPts, marking the current neighbor point Q as a core point, and adding the sample points in the epsilon-neighborhood point set of the current neighbor point Q which have not been clustered to the "neighbor set" of the core point p, when less than the minimum number of neighbor points required by the core point minPts, marking the current neighbor point Q as a boundary point, and clustering the boundary point into the "neighbor set" of the core point p; S57, when being accessed, jumping to S54 to select a new sample point, until the "neighbor set" of the core point p is traversed, and taking the "neighbor set" of the current core point p as a complete cluster; S58, jumping to step S51 to construct the next cluster, and traversing the sample points in the feature data set, to obtain multiple clusters, and marking the sample points which have not been included in the clusters as noise points; S59, marking the noise points as abnormal operation modes.
[0014] Further, the epsilon-neighborhood point set in S51 is: , Wherein, N ϵ (p) is the epsilon-neighborhood point set of the sample point p, dist(p,q) is the distance between the sample point p and the sample point q, D is the feature data set, and epsilon is the neighborhood radius.
[0015] Further, the types of the operation modes marked in S6 include: a "summer weekday evening peak high-density" mode, a "winter weekend flat peak low passenger flow" mode and a "spring and autumn best idling" mode. The multiple energy-saving operation strategies include: a low-load passenger long-station-distance idling optimization strategy, a peak high-density cooperative regenerative braking strategy, a seasonal temperature control and lighting self-adaptive strategy, a line topology self-adaptive traction strategy, a time-of-use electricity price-based energy storage scheduling strategy, a redundant time dynamic allocation strategy, a multi-car cooperative speed guidance strategy, a non-operating period equipment intelligent hibernation strategy and a special event dynamic response strategy.
[0016] The present application has the following advantages: 1. The present application realizes full-scene recognition of "typical operation mode + abnormal operation mode" through multi-dimensional feature extraction (5 types of features such as time, passenger flow and operation) and two clustering methods, breaks the limitations of existing fixed strategies or single-dimensional analysis, customizes exclusive energy-saving strategies for modes such as "peak high-density" and "low-load passenger long-station-distance", and executes corresponding schemes in real time to match the scene, thereby solving the problems of the prior art that cannot adapt to complex operation scenes and the energy-saving effect is unstable, making the energy-saving scheduling accurately fit the differences in passenger flow, train density and other parameters in different scenes, greatly improving the pertinence and effectiveness of the energy-saving strategy, and improving the energy-saving scheduling precision.
[0017] 2. Unlike existing data-driven methods that only focus on energy consumption analysis and visualization, this invention not only integrates multi-source data (including the potential processing capabilities of unstructured data) and mines the inherent relationship between operating modes and energy consumption characteristics, but also directly transforms the clustered patterns into executable energy-saving strategies (such as collaborative regenerative braking and seasonal temperature control), constructing a complete closed loop of "data acquisition - pattern recognition - strategy generation - real-time execution", fully releasing the value of multi-source data in rail transit and avoiding the waste of data resources.
[0018] 3. This invention identifies abnormal operating modes (such as sudden failures or a surge in passenger flow during large-scale events) and triggers early warning strategies, rather than applying conventional energy-saving solutions. This ensures both high energy efficiency in normal scenarios and prioritizes operational safety in abnormal scenarios, solving the problem of existing technologies that "emphasize energy saving over safety" or "make it difficult to balance safety and energy saving," and improving the overall robustness of rail transit systems. Attached Figure Description
[0019] Figure 1 This is a flowchart of a method for generating energy-saving strategies based on rail transit operation mode recognition. Detailed Implementation
[0020] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.
[0021] like Figure 1 As shown, a method for generating energy-saving strategies based on rail transit operation mode recognition includes the following steps: S1. Collect and preprocess multi-source data, and extract time features, passenger flow features, operation features, energy consumption features, and vehicle and environmental features from the preprocessed multi-source data; S2. Standardize each feature to obtain standardized features and form a feature dataset; S3. Select multiple initial cluster centers from the feature dataset based on the probability distribution of sample points in the feature dataset; S4. Assign each sample point to the cluster containing the nearest initial cluster center, correct the cluster center, update the cluster, and obtain a set of typical operating modes; S5. Classify each sample point in the feature dataset, find noise points, and mark the noise points as abnormal operating modes; S6. For each cluster in the typical operation mode set, first label its operation mode, and then based on each labeled operation mode, customize a variety of energy-saving operation strategies covering the whole scenario. S7. Collect and standardize data in real time. If the data matches the typical operating mode, execute the corresponding energy-saving strategy. If the data matches the abnormal operating mode, trigger the early warning strategy.
[0022] In this embodiment, the multi-source data in S1 includes: train operation data, energy consumption data, passenger flow data, and equipment status data.
[0023] Data collection covers all aspects of rail transit operations, and the main data includes: Train operation data: The number of trains on the line, train position, speed, acceleration, travel time between stations, and stop time are obtained from the Automatic Train Supervision (ATS); control commands (traction level, coasting command, and braking command corresponding to different speeds) are obtained from the Automatic Train Operation (ATO).
[0024] Energy consumption data: Energy consumption data for each time period and section is obtained from the traction power supply system, including traction energy consumption and regenerative energy feedback.
[0025] Passenger flow data: Obtain passenger flow, origin-destination (OD) data, etc. for each station and time period from the Automatic Fare Collection (AFC) system.
[0026] Equipment status data: including the operating status and fault information of equipment such as tracks, overhead contact lines, and power supply systems.
[0027] The data preprocessing process includes: data cleaning, data integration, and feature engineering.
[0028] Data cleaning: handling missing values, outliers, and duplicate data. For missing values in important fields, various imputation strategies are employed (such as mean imputation, regression imputation, etc.); for outliers, they are identified and corrected in conjunction with business rules.
[0029] Data integration: Integrating and associating data from different sources and in different formats in a spatiotemporal manner to form a unified record of operational data.
[0030] Feature engineering: Extracting energy-related feature variables, such as time period characteristics (morning peak, evening peak, off-peak, etc.), passenger flow characteristics (high passenger flow, low passenger flow, etc.), and operational characteristics (speed stability, acceleration distribution, etc.).
[0031] Multi-dimensional features strongly correlated with energy consumption are extracted from the preprocessed data to form a high-dimensional feature vector for cluster analysis. These features include, but are not limited to: The time feature T includes: a time period T1 (such as dividing a day into 24 time periods or category labels such as "morning peak", "flat peak", "evening peak", "night", and the like, and performing one-hot encoding), a day state T2: a workday / weekend / holiday flag, and a season T3: spring, summer, autumn, and winter.
[0032] The passenger flow feature P includes: a line average cross-sectional passenger flow (person / hour) P1, a maximum full load rate (%) P2, and a passenger flow unevenness coefficient P3 (maximum cross-sectional passenger flow / average cross-sectional passenger flow).
[0033] The operation feature O includes: an average train density (number of trains / hour) O1, a minimum train interval (seconds) O2, an average travel speed (km / h) O3, a speed curve smoothness (measured by a standard deviation of acceleration, m² / s³) O4, and an average inter-station operation time (seconds) O5.
[0034] The energy consumption feature E includes: a unit vehicle kilometer traction energy consumption (kWh / vehicle kilometer) E1, a regenerative energy feedback rate (feedback energy / traction energy consumption, %) E2, and a unit passenger flow energy consumption (kWh / person kilometer) E3.
[0035] The vehicle and environment feature V includes: a train marshalling number (sections) V1, a vehicle type identifier (different vehicle types have different efficiencies) V2, and an average ambient temperature (°C) V3.
[0036] Because the dimensions and orders of magnitude of the feature parameters are different, standardization processing must be performed to eliminate the influence thereof. The present application adopts a Z-Score standardization method, so that the mean of each feature is 0 and the standard deviation is 1: , wherein, is a feature after standardization processing, is an original feature, is a mean of the feature on all samples, is a standard deviation of the feature.
[0037] In the present embodiment, a correspondence relationship between the data features and the operation mode recognition output is identified by using a collaborative clustering strategy, an improved K-Means++ algorithm is used as a core algorithm to identify typical operation modes of rail transit, an improved DBSCAN algorithm is used to analyze the operation data to identify abnormal operation data points that do not belong to any typical mode, and when real-time data is identified as an abnormal point, the system enables a default operation strategy and issues an alarm.
[0038] In the present embodiment, S3 includes the following steps: S31, randomly selecting a sample point from a feature data set D={x1, x2,..., x N} as an initial clustering center, and placing the initial clustering center in an initial clustering center set, x1, x2,..., xN The feature dataset contains 1 to N sample points, where N is the number of sample points in the feature dataset. S32. Calculate the shortest distance between each sample point in the feature dataset and the initial cluster center in the initial cluster center set; S33. Calculate the probability distribution based on the shortest distance corresponding to each sample point in the feature dataset; S34. Based on the probability distribution, the roulette wheel method is used to select the next sample point from the feature dataset as the new initial cluster center, and the new initial cluster center is then added to the initial cluster center set. S35. Repeat steps S32 to S34 until K initial cluster centers are selected, where K is the upper limit of the number of initial cluster centers.
[0039] In this embodiment, the formula for calculating the shortest distance in S32 is: , Where D(x) i x is the shortest distance to the i-th sample point in the feature dataset. i For the i-th sample point, c j Let be the j-th initial cluster center in the current initial cluster center set, k be the number of initial cluster centers in the current initial cluster center set, j be the index of the initial cluster center, || || be the Euclidean distance, and min be the minimum value among the k distances.
[0040] In this embodiment, the formula for calculating the probability distribution in S33 is: , Wherein, P(x i Let be the probability distribution of each sample point in the feature dataset, and D be the number of sample points in the feature dataset. n Let D(x) be the nth sample point in the feature dataset. n ) represents the shortest distance to the nth sample point in the feature dataset, and D represents the feature dataset.
[0041] this This represents the "qualification" of a point to be selected as the next center; the farther a point is from an already selected center, the lower its qualification. The larger the value, the greater the probability of being selected. This ensures that the new center point will be selected in an area far away from existing points, making the initial centers as dispersed as possible.
[0042] The formula for calculating the probability distribution is used to divide each point Converting to the probability of being selected, this formula makes the probability of a point being selected proportional to the square of its distance from the already selected center point.
[0043] In the embodiment, the target of S4 is to assign each sample point to the cluster where the nearest cluster center locates, and S4 comprises the following sub-steps: S41, calculate the distance between each sample point in the feature data set and each initial cluster center, select the initial cluster center corresponding to the shortest distance, and add the corresponding sample point to the cluster; S42, traverse all sample points in the feature data set by S41 to obtain K initial clusters; S43, take the centroid of each initial cluster as a new initial cluster center, and re-execute S41 and S42 with the new initial cluster center until convergence is achieved, thereby obtaining K clusters; S44, form a typical operation mode set with the K clusters.
[0044] The centroid formula is: wherein, is the new initial cluster center of the jth cluster, x i is the ith sample point, C j is the jth cluster, and L is the number of sample points in the jth cluster.
[0045] In the embodiment, the convergence condition comprises: 1, the position of the cluster center no longer changes significantly (i.e., ||x || is less than a preset threshold, is the old initial cluster center). 2, the members of the cluster no longer change. 3, the preset maximum iteration number is reached.
[0046] Determining the optimal cluster number K comprises two methods. The first method is to measure the compactness of the model under different K values to assist in selecting K, and the core index is the within-cluster sum of squares.
[0047] The formula of the within-cluster sum of squares is: , wherein, is the total within-cluster sum of squares when the cluster number is K, is the sum of squares of the Euclidean distance of all sample points in the cluster C j to its c j , which measures the compactness of the cluster.
[0048] Calculate SSE(K) corresponding to K from 1 to a maximum value (such as 10). Draw the relationship curve of K and SSE(K). Observe the curve to find the elbow point. Before the point, SSE(K) rapidly decreases with the increase of K; after the point, the decrease speed suddenly slows down. The K value corresponding to the inflection point is the recommended value.
[0049] The second method comprehensively evaluates the clustering quality from two dimensions of within-cluster cohesion and between-cluster separation.
[0050] For sample point i (belongs to cluster C a ), calculate the average intra-cluster distance: , where, is the average distance of the ith sample point to all other points in the same cluster C a , is the number of elements in the set, The smaller the value is, the more the point i should belong to this cluster, and the higher the intra-cluster cohesion is.
[0051] Calculate the average nearest cluster distance: , where b(i) is the average nearest cluster distance of the ith sample point, C b is another cluster, and x m is the mth sample point.
[0052] The formula for calculating the average nearest cluster distance is used to calculate the average distance of the ith sample point to all points in all non-belonging clusters C a , and then take the minimum value of these average distances. b(i) can be understood as the average distance of point i to the "suboptimal belonging cluster". The larger b(i) is, the better the point i is separated from other clusters.
[0053] Calculate the silhouette coefficient of sample point i: , where s(i) is the silhouette coefficient of the ith sample point, and max is the maximum value.
[0054] In the formula for calculating the silhouette coefficient of sample point i, the numerator b(i)−a(i): measures the difference between the separation degree and the cohesion degree. The denominator max[a(i),b(i)]: is the normalization factor. The value range of s(i) is [−1,1]. s(i) close to 1: indicates that the clustering of point i is reasonable. s(i) close to 0: indicates that point i is on the boundary of two clusters. s(i) close to -1: indicates that the sample point i may be assigned to the wrong cluster.
[0055] The process of determining the optimal K value includes: clustering for each K value. Calculate the average value S(K) of the silhouette coefficient s(i) of all sample points. Select the K value that makes the average silhouette coefficient S(K) maximum as the optimal clustering number.
[0056] In this embodiment, S5 includes the following steps: S51, randomly select an unvisited sample point p in the feature data set, and obtain the ε-neighborhood point set of the sample point p, where ε is the neighborhood radius; S52, when the number of points in the ε-neighborhood point set of the sample point p is greater than or equal to the minimum number of neighbor points minPts required by the core point, the sample point p is marked as a core point p; S53, add all sample points in the neighborhood point set of the core point p to the "neighbor set"; S54, randomly select a sample point in the "neighbor set" to obtain a current neighbor point Q; S55, determine whether the current neighbor point Q has been visited; S56, when the current neighbor point Q has not been visited, check the number of points in the ε-neighborhood point set of the current neighbor point Q, when the number is greater than or equal to the minimum number of neighborhood points minPts required by the core point, mark the current neighbor point Q as a core point, and add the sample points in the ε-neighborhood point set of the current neighbor point Q that have not been clustered to the "neighbor set" of the core point p, when the number is less than the minimum number of neighborhood points minPts required by the core point, mark the current neighbor point Q as a boundary point, and cluster the boundary point into the "neighbor set" of the core point p; S57, when the current neighbor point Q has been visited, jump to S54 to select a new sample point until the "neighbor set" of the core point p is traversed, and the "neighbor set" of the current core point p is taken as a complete cluster; S58, jump to step S51 to construct the next cluster, and traverse all sample points in the feature data set to obtain multiple clusters, and mark the remaining sample points that have not been included in the clusters as noise points; S59, mark the noise points as abnormal operation modes.
[0057] These noise points correspond to sudden failures, large-scale activities, extreme weather, and other atypical abnormal operation states. When the system identifies online, once it is found that the real-time data is determined as noise by DBSCAN, a warning is triggered, and a conservative default operation strategy is adopted instead of any energy-saving strategy, thereby ensuring the operation safety. This greatly enhances the intelligence and robustness of the entire system.
[0058] In the embodiment, the ε-neighborhood point set in S51 is: , wherein, N ϵ (p) is the ε-neighborhood point set of the sample point p, dist(p, q) is the distance between the sample point p and the sample point q, D is the feature data set, and ε is the neighborhood radius.
[0059] The algorithm process in step S5 is summarized: given an object (sample point) p and a radius ε (specified by the parameter eps), the ε-neighborhood of point p refers to the set of all points within the hypersphere (circle in two-dimensional space) centered at p with radius ε.
[0060] Core point, boundary point and noise point: The type of a point depends on the number of points in its ε-neighborhood (including the point itself).
[0061] Core point: A point p is a core point if its ϵ-neighborhood contains at least minPts points (i.e., the density reaches the threshold), i.e., ∣N ϵ (p)∣≥minPts.
[0062] Border point: A point q is a border point if its ϵ-neighborhood contains less than minPts points, but it falls within the ϵ-neighborhood of some core point, i.e., ∣N ϵ (q)∣<minPts, and ∃p∈D such that p is a core point and q∈N ϵ (p).
[0063] Noise point: A point that is neither a core point nor a border point. These points are in low-density regions and do not belong to any cluster. The decision condition is that it does not satisfy the conditions of core points and border points.
[0064] The specific implementation process starts from any unvisited point. If the number of points in its ϵ-neighborhood exceeds minPts, a new cluster is created, and all density-reachable points are recursively added to the cluster. Points that cannot be assigned to any cluster are considered noise.
[0065] In this embodiment, the types of the running modes marked in S6 include: a "summer weekday evening peak high-density" mode, a "winter weekend flat peak low passenger flow" mode, and a "spring and autumn best coasting" mode; The various energy-saving running strategies include: a low-load passenger long-station-distance coasting optimization strategy, a peak high-density cooperative regenerative braking strategy, a seasonal temperature control and lighting adaptive strategy, a line topology adaptive traction strategy, a storage energy scheduling strategy based on time-of-use electricity prices, a redundant time dynamic allocation strategy, a multi-car cooperative speed guidance strategy, an intelligent hibernation strategy for equipment during non-operating periods, and a special event dynamic response strategy.
[0066] Business interpretation is performed on each cluster, and a label with a clear energy-saving meaning is assigned to it, for example: "Summer weekday evening peak high-density" mode: characterized by T1= evening peak, T3= summer, O1> 30 trains / hour, V3> 30℃, and E1 and E2 are both high.
[0067] "Winter weekend flat peak low passenger flow" mode: characterized by T1= flat peak, T2= weekend, P1< 50% full load rate, and V3< 10℃.
[0068] "Spring and autumn best coasting" mode: characterized by T3= spring / autumn, O1 is moderate, O4 (speed smoothness) is good, and V3 is suitable (no need for high-power air conditioning).
[0069] At the same time, detailed feature portraits are generated for each mode, i.e. the statistical distribution of each feature parameter of all sample points in the cluster (such as mean, quantile, etc.), providing accurate data support for subsequent customization of energy-saving strategies.
[0070] This embodiment sets 9 different energy-saving operation modes.
[0071] 1. Low passenger load long station distance coasting optimization strategy Target mode: suitable for suburban lines, night period, characterized by low passenger flow, long station spacing, no urgent on-time pressure.
[0072] Feature parameters: passenger flow < full load rate 30%, station spacing > 2km, planned redundancy time > 1 minute.
[0073] Strategy content: significantly reduce the maximum running speed (e.g. from 100km / h to 80km / h) and extend the coasting phase in advance on the premise of ensuring the minimum travel time. The system will calculate the optimal "acceleration-coasting-braking" curve to maximize the use of train kinetic energy and reduce braking energy consumption.
[0074] 2. Peak high-density coordinated regenerative braking strategy Target mode: suitable for morning and evening peak hours, characterized by maximum train density (minimum interval < 2 minutes) and frequent train start-stop.
[0075] Feature parameters: train density > 30 pairs / hour, large energy consumption fluctuation per unit time, frequent regenerative energy generation.
[0076] Strategy content: dynamically optimize the coordinated operation of trains within the same power supply partition. When a train brakes, the system will slightly adjust the traction of adjacent accelerating trains, or preferentially use the regenerative energy through the energy management system (EMS) rather than feeding it back to the power grid, increasing the "immediate utilization rate" of regenerative energy from the usual 30-40% to more than 50%.
[0077] 3. Seasonal temperature control and lighting adaptive strategy Target mode: suitable for all operation modes, but with parameter fine-tuning according to season and external environment.
[0078] Feature parameters: ambient temperature, humidity, light intensity, car full load rate.
[0079] Strategy content: Summer: within the acceptable comfort range of passengers, dynamically set the target temperature of the car (e.g. from 24°C to 26°C), and optimize the operation mode of the fresh air system.
[0080] Winter: use passenger body heat and equipment heat dissipation to reduce electric heating power.
[0081] Lighting: automatically adjust the brightness of platform and carriage lighting according to sunlight intensity and whether the train is entering or leaving the station.
[0082] 4. Line topology adaptive traction strategy Target mode: suitable for lines with many slopes and large curves.
[0083] Characteristic parameters: real-time train position, line slope curve, and current train load.
[0084] Strategy content: the system pre-stores the accurate slope map of the line. Before climbing a slope, the train is allowed to slightly exceed the speed limit on a flat section to accumulate kinetic energy; before descending a slope, the train enters coasting or slight braking mode in advance to convert potential energy and reduce net traction energy consumption.
[0085] 5. Energy storage scheduling strategy based on time-of-use electricity prices Target mode: suitable for vehicle depots, parking lots, and sites on the line equipped with energy storage devices (such as flywheels or batteries).
[0086] Characteristic parameters: real-time electricity prices (peak / flat / valley), energy storage device SOC (state of charge), and predicted next peak period energy consumption.
[0087] Strategy content: charge the energy storage device during the electricity price valley period (such as late at night) and discharge it during the electricity price peak period (such as the evening rush hour) to power the train or station equipment, directly reducing electricity costs.
[0088] 6. Dynamic allocation of redundant time strategy Target mode: suitable for timetables with planned redundant time (buffer time).
[0089] Characteristic parameters: current delay status, remaining time to the next site, and scheduled redundant time.
[0090] Strategy content: if the train is on time or slightly ahead of schedule, the system will automatically enable the energy-saving operation curve to smoothly allocate redundant time instead of waiting for a long time before the station. If the train is delayed, switch to the on-time priority mode.
[0091] 7. Multi-train coordinated speed guidance strategy Target mode: suitable for medium to high density lines and with continuous communication capabilities between vehicles (V2V) or between vehicles and the ground.
[0092] Characteristic parameters: distance between front and rear trains, speed, and real-time operation status.
[0093] Strategy content: not only single train optimization, but also overall optimization of multiple trains as a system. By slightly adjusting the speed of the leading train, create more favorable coasting or braking conditions for the following train, form a "green wave", and reduce the average energy consumption of the entire section.
[0094] 8. Intelligent hibernation strategy for non-operation period Target mode: Suitable for yards and stations during non-operation period at night.
[0095] Characteristic parameters: Operation timetable, last use time of equipment, equipment inspection plan.
[0096] Strategy content: Implement hierarchical and time-based hibernation for non-critical equipment (such as partial lighting, advertising screens, automatic ticket vending machines, and station hall ventilation), rather than simply turning off all equipment, to balance energy saving and equipment wake-up preparation, safety monitoring.
[0097] 9. Dynamic response strategy for special events Target mode: Suitable for sudden or planned special events such as large-scale activities, sports events, and extreme weather.
[0098] Characteristic parameters: AFC abnormal passenger flow data, dispatching instructions, and warning information released by social media or transportation departments.
[0099] Strategy content: After the system recognizes the special event mode, it dynamically overrides the regular energy saving strategy to prioritize capacity and safety. For example, during the dispersal peak, a high-acceleration performance timetable is used, and empty cars are dispatched in advance, with energy saving as a secondary goal.
[0100] The matching details of the nine different operation energy saving modes, characteristic parameters, and clustering algorithms are shown in Table 1.
[0101] Table 1 details the characteristic parameters corresponding to each operation mode and the core logic of the clustering algorithm for identifying the mode. Table 1: Matching table of energy saving operation modes, characteristic parameters, and clustering identification Operation mode Corresponding characteristic parameters Clustering algorithm execution and optimization Mode labeling and identification description 1. Low passenger load and long station distance inertial optimization strategy Passenger flow characteristics: P1 (average cross-section passenger flow) low, P2 (maximum full load rate) < 30%. Operating characteristics: O5 (average station-to-station running time) long (corresponding to a station spacing > 2 km), O3 (average travel speed) possibly lower. Time characteristics: T1 (time period) is night or flat peak, T2 (weekday / weekend) can be a low demand day. Supplementary parameters: the "station spacing" parameter is derived from the data Algorithm optimization: KMeans++ algorithm, because the mode characteristics are relatively concentrated (such as low passenger flow and long station distance). The number of clusters K is determined by the elbow rule and the silhouette coefficient (such as K=35 may include this mode). Execution process: the algorithm clusters based on characteristics such as P1, P2, and O5, and finds a cluster with low P1 and P2 values and high O5 values. The silhouette coefficient is used to evaluate the tightness within the cluster. Business labeling: after clustering, the cluster is labeled as "low passenger load and long station distance mode", with characteristics of low passenger flow and long inter-station running time. The energy-saving strategy triggers the inertial optimization. 2. Peak high-density collaborative regenerative braking strategy Time feature: T1 (period) is early or late peak. Operational feature: O1 (average vehicle density) > 30 pairs / hour, O2 (minimum vehicle interval) < 2 minutes. Energy consumption feature: E2 (regenerative energy feedback rate) is high, E1 (unit vehicle-kilometer traction energy consumption) fluctuates greatly. Supplementary parameter: "Vehicle density maximum" flag derived from O1 and O2. Algorithm optimization: KMeans++ or DBSCAN. DBSCAN may be more optimal, as the peak mode may form a dense cluster (high O1 and O2), and can handle noise (such as sudden events). The value of K is determined by maximizing the silhouette coefficient. Execution process: the algorithm clusters based on T1, O1, O2, and E2, and identifies a cluster with high density and high feedback rate. The silhouette coefficient evaluates the separation degree of the cluster. Business labeling: labeled as "peak high-density mode", with characteristics of high train density and frequent regenerative energy. The strategy triggers collaborative regenerative braking. 3. Seasonal temperature control and lighting adaptive strategy Time characteristics: T3 (season) is winter, summer, spring and autumn. Vehicle and environmental characteristics: V3 (average environmental temperature) is high or low, and V1 (train formation) may affect air conditioning load. Passenger flow characteristics: P2 (maximum full load rate) affects car heat load. Supplementary parameters: add "light intensity" parameter (from environmental data). Algorithm optimization: KMeans++ algorithm, because the seasonal mode characteristics are distributed in a spherical shape. The number of clusters K is divided by season (such as K=4 corresponds to four seasons). The elbow rule helps determine K. Execution process: the algorithm clusters based on T3, V3, and P2, and generates multiple clusters such as "summer high temperature and high passenger flow" and "winter low temperature and low passenger flow". The silhouette coefficient evaluates the quality of the seasonal cluster. Business labeling: labeled as "summer high temperature mode" or "winter low temperature mode", etc. The strategy triggers temperature control and lighting adaptation. 4. Line topology adaptive traction strategy Running characteristics: O4 (speed curve smoothness) is poor (due to slope changes) and O3 (average travel speed) fluctuates greatly. Supplementary parameters: need to clearly add "line slope" and "radius of curvature" parameters (derived from line topology data, mentioned in high-dimensional feature vector but not listed in the table, can be considered as an extension of running characteristics). Vehicle characteristics: V2 (vehicle type identification) affects traction performance. Algorithm optimization: KMeans++ or hierarchical clustering. Hierarchical clustering is suitable for the hierarchical structure of line topology (such as different sections). The value of K is determined by business knowledge (such as the number of line sections). Execution process: the algorithm clusters based on slope, O4, and O3, and finds a cluster with large slope changes (such as a continuous uphill section). The silhouette coefficient evaluates the reasonableness of the cluster. Business labeling: labeled as "large slope line mode" or "bend-intensive mode". The strategy triggers traction optimization. 5. Energy storage scheduling strategy based on time-of-use electricity price Time feature: T1 (period) corresponds to electricity price peak valley flat (e.g. valley period is midnight). Energy consumption feature: E1 (unit vehicle kilometer traction energy consumption) is high in peak period. Supplementary parameter: add "time-of-use electricity price" parameter (obtained from grid data, mentioned in high-dimensional feature vector). Operating feature: O1 (traffic density) is low in valley period. Algorithm preference: KMeans++ algorithm, as the electricity price pattern has periodicity. The number of clusters K is divided according to the electricity price period (e.g. K=3 corresponds to peak, flat, valley). The silhouette coefficient is used for verification. Execution process: the algorithm clusters based on T1, E1, and electricity price parameters, and identifies the "electricity price valley period low energy consumption" cluster. The elbow rule determines K. Business annotation: annotated as "electricity price valley period mode" or "high peak high cost mode". The strategy triggers energy storage scheduling. 6. Dynamic allocation strategy of redundant time Operating feature: O5 (average inter-station operating time) is compared with "graph scheduling time" to derive the "redundant time" parameter. Time feature: T1 (period) may be flat peak or period with buffer. Supplementary parameter: add "delay level" or "schedule deviation" parameter (obtained from scheduling data). Algorithm preference: DBSCAN or KMeans++. DBSCAN may be more optimal, as the redundant time pattern may be an outlier (e.g. delay event). The value of K is selected through the silhouette coefficient. Execution process: the algorithm clusters based on redundant time, O5, and delay parameters, and finds the "planned redundancy" cluster. The silhouette coefficient evaluates the tightness of the cluster. Business annotation: annotated as "redundant time mode" or "on-time buffer mode". The strategy triggers dynamic allocation. 7. Multi-vehicle coordinated speed guidance strategy Operating feature: O1 (average traffic density) is high, O2 (minimum traffic interval) is small. Energy consumption feature: E2 (regenerative energy feedback rate) is high. Supplementary parameter: add "vehicle-vehicle distance" or "coordination coefficient" parameter (derived from V2V communication data). Time feature: T1 (period) is peak. Algorithm preference: KMeans++ algorithm, as the multi-vehicle coordination feature is relatively continuous. The number of clusters K is determined by the elbow rule (e.g. number of medium-high density sections). Execution process: the algorithm clusters based on O1, O2, vehicle-vehicle distance, etc., and identifies the "small vehicle distance, synchronized speed" cluster. The silhouette coefficient evaluates the coordination. Business annotation: annotated as "high-density coordination mode" or "green wave mode". The strategy triggers speed guidance. 8. Intelligent sleep strategy of equipment during non-operating period Time feature: T1 (period) is the night non-operating period. Operating feature: O1 (average traffic density) is close to zero, O5 (inter-station operating time) is none. Energy consumption feature: E1 (unit vehicle kilometer traction energy consumption) is low, E3 (unit passenger flow energy consumption) is low. Supplementary parameter: add "equipment status" parameter (obtained from inspection data). Algorithm preference: KMeans++ algorithm is simple and effective, as non-operating period data is obviously separated. K=2 or 3 (operating / non-operating). The elbow rule is obvious. Execution process: the algorithm clusters based on T1, O1, and E1, and directly separates the "non-operating" cluster. The high silhouette coefficient indicates good separation. Business annotation: Annotated as "non-operating period mode". Strategy trigger device dormancy. 9. Special event dynamic response strategy Time characteristics: T2 (weekday / holiday) is a special holiday or event day. Passenger flow characteristics: P1 (average cross-section passenger flow) is abnormally high, and P3 (passenger flow unevenness coefficient) is high. Supplementary parameters: add "event warning" parameters (obtained from AFC or social media data). Running characteristics: O1 (train running density) may temporarily increase. Algorithm optimization: DBSCAN algorithm is most suitable, because special events are abnormal points (noise), and DBSCAN can automatically identify abnormal clusters. K value is not fixed. Execution process: The algorithm clusters based on T2, P1, and event parameters. DBSCAN will cluster normal points into clusters, and special events will be identified as noise points. The contour coefficient is used to evaluate normal clusters. Business annotation: Annotated as "special event mode" or "abnormal passenger flow mode". Strategy trigger dynamic response. The present application first proposes and systematically implements a method based on big data clustering analysis to automatically identify typical operation modes of rail transit and dynamically generate and recommend customized energy saving strategies for each mode. It breaks through the current static energy saving methods that rely on fixed timetables or simple rules (such as only dividing into peak and off-peak periods), and solves the pain point of their inability to adapt to complex and variable operation scenarios.
[0102] The present application establishes a complete learning closed loop of "data collection → mode discovery → strategy generation → effect evaluation → strategy optimization". The system can continuously optimize itself as data accumulates, continuously improving energy saving effects. It changes the traditional mode of relying on manual experience adjustment, lacking quantitative feedback and continuous improvement capabilities, and realizes the automation and intelligentization of energy saving management.
[0103] The present application defines an unprecedented multi-dimensional feature parameter system dedicated to energy consumption pattern recognition. The system not only contains conventional time series, passenger flow, and operation data, but also creatively introduces line topology (slope, curvature), time-of-use electricity price, environmental temperature, equipment status, and other parameters that have a significant impact on energy consumption but have been overlooked before. This makes pattern recognition more accurate and is the key to distinguishing it from other general big data analysis solutions.
[0104] The present application uses a first clustering algorithm to efficiently identify major and conventional operation patterns (such as peak and flat peaks). A second clustering algorithm is specifically used to discover abnormal operation patterns (such as emergencies and faults), which is the key to ensuring system robustness.
[0105] The present application first constructs a strategy library containing nine fine energy-saving strategies. Each strategy accurately corresponds to an operation pattern identified by clustering and provides specific solutions from control instructions (such as speed regulation) to scheduling suggestions (such as coordinated braking).
[0106] The above is only a preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included within the scope of protection of the present application.
Claims
1. A method for generating energy-saving strategies based on rail transit operation mode recognition, characterized in that, Includes the following steps: S1. Collect and preprocess multi-source data, and extract time features, passenger flow features, operation features, energy consumption features, and vehicle and environmental features from the preprocessed multi-source data; S2. Standardize each feature to obtain standardized features and form a feature dataset; S3. Select multiple initial cluster centers from the feature dataset based on the probability distribution of sample points in the feature dataset; S4. Assign each sample point to the cluster containing the nearest initial cluster center, correct the cluster center, update the cluster, and obtain a set of typical operating modes; S5. Classify each sample point in the feature dataset, find noise points, and mark the noise points as abnormal operating modes; S6. For each cluster in the typical operation mode set, first label its operation mode, and then based on each labeled operation mode, customize a variety of energy-saving operation strategies covering the whole scenario. S7. Collect and standardize data in real time. If the data matches the typical operating mode, execute the corresponding energy-saving strategy. If the data matches the abnormal operating mode, trigger the early warning strategy.
2. The energy-saving strategy generation method based on rail transit operation mode recognition according to claim 1, characterized in that, The multi-source data in S1 includes: train operation data, energy consumption data, passenger flow data, and equipment status data.
3. The energy-saving strategy generation method based on rail transit operation mode recognition according to claim 1, characterized in that, S3 includes the following steps: S31. Randomly select a sample point from the feature dataset as the initial cluster center, and place the initial cluster center in the initial cluster center set; S32. Calculate the shortest distance between each sample point in the feature dataset and the initial cluster center in the initial cluster center set; S33. Calculate the probability distribution based on the shortest distance corresponding to each sample point in the feature dataset; S34. Based on the probability distribution, the roulette wheel method is used to select the next sample point from the feature dataset as the new initial cluster center, and the new initial cluster center is then added to the initial cluster center set. S35. Repeat steps S32 to S34 until K initial cluster centers are selected, where K is the upper limit of the number of initial cluster centers.
4. The energy-saving strategy generation method based on rail transit operation mode recognition according to claim 3, characterized in that, The formula for calculating the shortest distance in S32 is: , Where D(x) i x is the shortest distance to the i-th sample point in the feature dataset. i For the i-th sample point, c j Let be the j-th initial cluster center in the current initial cluster center set, k be the number of initial cluster centers in the current initial cluster center set, j be the index of the initial cluster center, || || be the Euclidean distance, and min be the minimum value among the k distances.
5. The energy-saving strategy generation method based on rail transit operation mode recognition according to claim 3, characterized in that, The formula for calculating the probability distribution in S33 is as follows: , Wherein, P(x i Let be the probability distribution of each sample point in the feature dataset, and D be the number of sample points in the feature dataset. n Let D(x) be the nth sample point in the feature dataset. n ) represents the shortest distance to the nth sample point in the feature dataset, and D represents the feature dataset.
6. The energy-saving strategy generation method based on rail transit operation mode recognition according to claim 1, characterized in that, S4 includes the following steps: S41. Calculate the distance between sample points in the feature dataset and multiple initial cluster centers, select the cluster containing the initial cluster center corresponding to the shortest distance, and add the corresponding sample points to that cluster; S42. By traversing all sample points in the feature dataset in step S41, K initial clusters are obtained; S43. Take the centroid of each initial cluster as the new initial cluster center. With the new initial cluster centers, re-execute S41 and S42 until convergence, and obtain K clusters. S44. Form a set of typical operating modes for each cluster.
7. The energy-saving strategy generation method based on rail transit operation mode recognition according to claim 1, characterized in that, S5 includes the following steps: S51. Randomly select an unvisited sample point p in the feature dataset and obtain the ϵ-neighborhood point set of sample point p, where ϵ is the neighborhood radius. S52. When the number of points in the ϵ-neighborhood set of sample point p is greater than or equal to the minimum number of neighborhood points minPts required for the core point, the sample point p is marked as the core point p. S53. Add all sample points in the neighborhood point set of the core point p to the "neighbor set"; S54. Take any sample point from the "neighbor set" to obtain the current neighbor point Q; S55. Determine whether the current neighbor point Q has been visited; S56. When not visited, check the number of points in the ϵ-neighborhood set of the current neighbor point Q. If it is greater than or equal to the minimum number of neighborhood points minPts required for the core point, mark the current neighbor point Q as the core point and add the unclustered sample points in the ϵ-neighborhood set of the current neighbor point Q to the "neighborhood set" of the core point p. If it is less than the minimum number of neighborhood points minPts required for the core point, mark the current neighbor point Q as a boundary point and assign the boundary points to the "neighborhood set" of the core point p. S57. When a point has already been visited, jump to S54 to select a new sample point until the "neighbor set" of the core point p has been traversed. Treat the "neighbor set" of the current core point p as a complete cluster. S58. Jump to step S51 to construct the next cluster. After traversing all the sample points in the feature dataset, multiple clusters are obtained. The remaining sample points that are not included in the cluster are marked as noise points. S59. Mark the noise points as abnormal operating modes.
8. The energy-saving strategy generation method based on rail transit operation mode recognition according to claim 7, characterized in that, The set of ϵ-neighborhood points in S51 is: , Where, N ϵ (p) is the set of ϵ-neighborhood points of sample point p, dist(p,q) is the distance between sample point p and sample point q, D is the feature dataset, and ϵ is the neighborhood radius.
9. The energy-saving strategy generation method based on rail transit operation mode recognition according to claim 1, characterized in that, The types of operating modes marked in S6 include: "High-density evening peak on summer weekdays" mode, "Low passenger flow on weekends in winter" mode, and "Optimal inertia mode in spring and autumn" mode. Multiple energy-saving operation strategies include: low-passenger-load long-station-distance coasting optimization strategy, peak-hour high-density collaborative regenerative braking strategy, seasonal temperature control and lighting adaptive strategy, line topology adaptive traction strategy, energy storage scheduling strategy based on time-of-use pricing, redundant time dynamic allocation strategy, multi-vehicle collaborative speed guidance strategy, equipment intelligent hibernation strategy during non-operation periods, and special event dynamic response strategy.
Citation Information
Patent Citations
Rail transit big data analysis method and system
CN111831658A
Urban traffic prediction method and system based on spatio-temporal data flow fusion analysis
CN111179592A
Method and system for offline customer arriving and path planning based on geographic position
CN118396517A
