CN-121998186-A - Subway energy consumption prediction method based on situation enhancement and block convolution converter

CN121998186ACN 121998186 ACN121998186 ACN 121998186ACN-121998186-A

Abstract

The invention discloses a subway energy consumption prediction method based on situation enhancement and block convolution transformers, and relates to the technical field of artificial intelligence time sequence prediction and urban rail transit energy management intersection. The method comprises the steps of obtaining subway related data, preprocessing the subway related data to construct an input feature matrix containing passenger flow grading features, obtaining a single line feature sequence according to a line splitting matrix, inputting the single line feature sequence into a PCformer model after block modeling processing, enhancing self-attention extraction features through multi-scale convolution, adopting a combined loss function training model fusing mean square error and average absolute error, dynamically adjusting parameters until convergence, and obtaining an actual energy consumption value by utilizing independent reasoning and inverse normalization of the trained model. The method can be used for energy consumption prediction of the multi-line subway network and supports fine power supply scheduling. Through the cooperation of multiple technical features, the prediction precision under a complex operation scene is improved, and the generalization capability and stability of the model are enhanced.

Inventors

LONG LIHENG
ZHANG LIJIE
LI LINLIN
HU LIJUN
CHEN ZHIYAO
WU JUNQIAN
LIU LIPING
WU DIANHUA

Assignees

广州地铁设计研究院股份有限公司

Dates

Publication Date: 20260508
Application Date: 20260122

Claims (10)

1. A subway energy consumption prediction method based on situation enhancement and block convolution transformers is characterized by comprising the following steps: acquiring subway related data, preprocessing the subway related data, and constructing an input feature matrix based on the preprocessed subway related data; Splitting the input feature matrix according to the subway line to obtain a plurality of single-line feature sequences; carrying out block modeling processing on each single-line characteristic sequence to obtain a block sequence; Inputting the block sequence into PCformer model, and extracting multi-scale convolution enhanced self-attention characteristic of the block sequence; training the PCformer model by adopting a joint loss function until the model converges; And independently reasoning each single-line block sequence by using the PCformer model which is completed through training, and performing inverse normalization processing on the reasoning result to obtain an actual energy consumption value.
2. The method of claim 1, wherein the step of determining the position of the substrate comprises, When the subway related data is acquired, historical energy consumption data, passenger flow data, external situation data and station attribute data are collected; when preprocessing the subway related data, filling the missing values through linear interpolation, removing abnormal peak values by adopting an abnormal value recognition criterion, replacing the abnormal peak values with adjacent time step average values, carrying out normalization processing on continuous data, and encoding the classified data; When the input feature matrix is constructed, the historical energy consumption main sequence, the passenger flow sequence, the weather related sequence, the date attribute identification, the peak time identification and the passenger flow grading feature are integrated, and the passenger flow grading feature is generated by setting critical points, constructing a fuzzy connection function, adapting to multiple scene influence coefficients and dynamic adjustment parameters.
3. The method of claim 1, wherein the input feature matrix is split according to subway lines, cross-line feature stitching is not performed, and only subsequent model parameter sharing is reserved, so that the single-line feature sequence corresponding to each line is obtained.
4. The method according to claim 1, wherein when each single line feature sequence is subjected to block modeling processing, feature extraction and channel expansion are performed on the single line feature sequence through one-dimensional convolution, the expanded features are subjected to block operation through one-dimensional depth separable convolution, the blocked features are subjected to linear weighted fusion, and the fused features are subjected to channel compression through one-dimensional convolution, so that the block sequence containing operation period semantic segments is obtained.
5. The method of claim 1, wherein after the block sequence is input into the PCformer model, the multi-scale convolution kernel is utilized to perform feature extraction on keys and values in the self-attention mechanism respectively, the extracted multi-scale features are fused, the fused features are sent into a probability sparse attention mechanism to perform long Cheng Yilai modeling, and multi-scale convolution enhanced self-attention feature extraction is completed.
6. The method of claim 1, wherein when the joint loss function is adopted for training, a mean square error and an average absolute error are fused, weight parameters are set in combination with line type, time period and passenger flow intensity differentiation, and a passenger flow change duration matching error is introduced to construct a double optimization objective function; and training the PCformer model by an optimizer, and dynamically adjusting passenger flow classification critical points, fuzzy connection function parameters and duration matching weights based on verification set errors after each round of training until the dual optimization objective function converges.
7. The method of claim 2, wherein the external context data comprises weather-related data, date attribute-related data, operation period-related data, and city event-related data, and the station attribute data comprises station type labeling information.
8. The method of claim 2, wherein the step of determining the position of the substrate comprises, Setting a passenger flow variable quantity critical point according to the type and the time period of the station and setting a passenger flow continuous time critical point according to the equipment regulation characteristic when setting the critical point; when the fuzzy connection function is constructed, a triangular fuzzy function is established between adjacent passenger flow variable quantity critical points; When the multi-scene influence coefficient is adapted, a combined scene is built by integrating the passenger flow variable quantity, the passenger flow duration time, the station type and the time period, and a corresponding energy consumption influence coefficient matrix is built; When parameters are dynamically adjusted, the parameters of all critical points, fuzzy connection functions and time length matching weights are adjusted through a gradient descent method.
9. The method of claim 4, wherein when the one-dimensional depth separable convolution partitioning operation is performed, a convolution kernel size and a step size are set, and the single-line feature sequence is segmented into partitions corresponding to operation semantic segments, so as to obtain the partitioned features.
10. The method of claim 1, wherein the PCformer models are trained by dividing the preprocessed subway-related data into a training set, a verification set and a test set, setting training rounds, batch sizes and learning rates, adjusting the learning rates by a learning rate attenuation strategy, and outputting the energy consumption prediction results of future time periods of each line by utilizing the PCformer models after training reasoning, and recovering the original scale of the data by inverse normalization to obtain the directly applicable actual energy consumption values.

Description

Subway energy consumption prediction method based on situation enhancement and block convolution converter Technical Field The invention relates to the technical field of intersection of artificial intelligent time sequence prediction and urban rail transit energy management, and is particularly applied to energy consumption prediction of a multi-line subway network. Background With the acceleration of the urban process, the ground traffic is difficult to meet the rapidly-increased travel demands of urban residents, and the urban rail traffic gradually becomes an important component of the urban public transportation system by virtue of the advantages of large transportation capacity, high efficiency, low energy consumption, stable operation and the like. Along with the development of a subway system to a networked and high-density operation mode, the energy consumption problem is increasingly remarkable, the operation cost is increased, and the subway system is directly related to an urban carbon reduction development target, so that the accurate prediction of the energy consumption of the urban subway line network becomes a key for realizing refined power supply scheduling, improving the operation efficiency and reducing the carbon emission. The energy consumption prediction of the subway line network has the challenge that the energy consumption covers a plurality of parts such as train traction energy consumption, station illumination energy consumption, ventilation energy consumption, air conditioner energy consumption and the like, passenger flow portraits, peak time periods, running modes and conditions of different subway lines are different, the energy consumption modes show obvious heterogeneity, and when a prediction object is expanded to the multi-line interweaved subway network, the prediction difficulty is greatly increased. The existing subway energy consumption prediction method mainly comprises a statistical model, a machine learning model and a deep learning model. The traditional statistical model is based on linear assumption, can not capture nonlinear fluctuation caused by passenger flow peak, holiday effect and burst scheduling change, and has insufficient prediction precision in complex subway operation scenes. The machine learning model relies on feature engineering to extract key variables, is sensitive to noise and has limited generalization capability. The deep learning method is outstanding in the field of time sequence prediction, the convolutional neural network can effectively extract local features, the cyclic neural network can capture time sequence dependence, but the sequence recursion structure of the cyclic neural network is difficult to process in parallel, and gradient disappearance is easy to occur in long sequence prediction. The Transformer and the improved model thereof realize long-range dependency modeling by using a self-attention mechanism and are excellent in sequence prediction, but the models are designed for general time sequence scenes in multiple directions, and are not optimized for the characteristics of periodicity, line heterogeneity, influence by external factors and the like of subway energy consumption data. Some researches try to improve prediction accuracy through methods such as pattern decomposition, spectrum analysis and signal decomposition, but a unified modeling framework specially suitable for multi-line subway network energy consumption prediction is still lacking. The subway energy consumption sequence is different from industrial load or meteorological data, and has strong periodicity, complex non-stationarity and obvious inter-line structure difference. If a multi-line mixed modeling mode is simply adopted, personalized energy consumption characteristics of each line can be ignored, false correlation relations are easily introduced, the model is over-fitted, and the prediction effect is limited. Meanwhile, the existing prediction method only simply correlates passenger flow data, the specificity of subway passenger flow and an environmental control system is not considered, the influence of the passenger flow on energy consumption is not in a continuous linear relation, key energy consumption equipment such as an air conditioner and an elevator has an energy consumption stability threshold, the energy consumption can be obviously changed only after the passenger flow reaches a specific critical point, the equipment configuration and the space layout of different subway stations are different, the energy consumption critical points of the same subway station in different time periods can be dynamically changed, the existing model lacks specific modeling on the characteristic, and the prediction precision is further insufficient. In addition, the existing self-attention mechanism focuses on capturing global long-range dependence of a sequence, has limited capability of describing local energy consumption fluctuation, is difficult to simu