A subway station energy consumption working condition fine division method and system

Abstract

The application discloses a subway station energy consumption working condition fine division method and system, and relates to the technical field of subway energy consumption.The application utilizes cross-correlation function and fusion fuzzy time cognitive graph model to dynamically analyze time delay correlation between multidimensional data, adjusts a data scanning window according to subway operation characteristics, corrects data time offset, and improves data synchronism and reliability; adopts a sliding window method and a dynamic adjustment window width strategy, enhances adaptability of the model to different operation periods, and ensures accuracy and robustness of data correction; through Savitzky-Golay filtering, jump point detection and a dynamic trend model, transverse errors of multidimensional data on a time axis are eliminated, key jump points are aligned, transverse error correction is fine, and high-consistency fine correction data is provided for subsequent working condition division.

Description

Technical Field

[0001] This invention relates to the field of subway energy consumption technology, and in particular to a method and system for finely classifying the energy consumption conditions of subway stations. Background Technology

[0002] With the rapid development of subway construction, energy consumption in subway stations has become a major concern. Existing methods for classifying subway station operating conditions based on operational data typically assume that the data is synchronized in time, meaning that the time coordinates of sampled data for different variables are consistent. However, in actual subway station operation, different equipment exhibits significantly different response characteristics to passenger flow. For example, elevators respond quickly to passenger flow, while air conditioning requires a certain amount of passenger accumulation to respond. Furthermore, when passenger flow causes changes in airflow and pressure parameters, changes in variables such as temperature, CO2 concentration, and air conditioning cooling load are delayed. Therefore, even if the time coordinates of sampled data for each variable are the same, the information is lagging, causing uncorrected operating condition classifications to deviate from reality and fail to meet the refined energy-saving requirements of subway stations. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for finely classifying energy consumption conditions in subway stations, so as to improve the above-mentioned technical problems.

[0004] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0005] A method for finely classifying energy consumption conditions in subway stations is provided, which includes:

[0006] Obtain multi-dimensional hourly energy consumption data of a subway station; the multi-dimensional hourly energy consumption data includes hourly data of carbon dioxide sensing, hourly data of equipment energy consumption, hourly data of passenger flow, and hourly data of temperature;

[0007] The cross-correlation function and the fused fuzzy time cognitive graph model are used to correct the multi-dimensional hourly energy consumption data, resulting in multi-dimensional hourly energy consumption correction data.

[0008] Lateral error correction is performed on the multi-dimensional hourly energy consumption correction data to obtain multi-dimensional hourly energy consumption fine-calibration data.

[0009] The TICC algorithm is used to divide the multi-dimensional hourly energy consumption calibration data to obtain the operating condition division results.

[0010] A refined energy consumption condition classification system for subway stations is provided, comprising:

[0011] The multi-dimensional hourly energy consumption acquisition module is used to acquire multi-dimensional hourly energy consumption data of a subway station.

[0012] The data correction module is used to correct multi-dimensional hourly energy consumption data using cross-correlation functions and fused fuzzy time cognitive graph models, so as to obtain multi-dimensional hourly energy consumption corrected data.

[0013] The data calibration module is used to perform lateral error correction on multi-dimensional hourly energy consumption calibration data to obtain multi-dimensional hourly energy consumption calibration data.

[0014] The operating condition segmentation module is used to segment multi-dimensional hourly energy consumption calibration data using the TICC algorithm to obtain the operating condition segmentation results.

[0015] The beneficial effects of this invention are as follows:

[0016] This invention utilizes cross-correlation functions and a fused fuzzy time cognitive graph model to dynamically analyze the time delay correlation between multi-dimensional data. It adjusts the data scanning window according to the characteristics of subway operations to correct data time offsets and improve data synchronization and reliability. A sliding window method and a dynamic window width adjustment strategy are employed to enhance the model's adaptability to different operating periods, ensuring the accuracy and robustness of data correction. Through Savitzky-Golay filtering, abrupt change point detection, and a dynamic trend model, lateral errors in multi-dimensional data along the time axis are eliminated, key abrupt change points are aligned, and lateral error correction is refined, providing highly consistent, precisely calibrated data for subsequent operational condition segmentation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the method in an embodiment of the present invention;

[0019] Figure 2 This is a flowchart of the improved wave search algorithm in an embodiment of the present invention;

[0020] Figure 3 This is a system structure diagram in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 This embodiment provides a method for finely classifying energy consumption conditions in subway stations, which includes:

[0023] S1. Obtain multi-dimensional hourly energy consumption data for a subway station. This multi-dimensional hourly energy consumption data includes hourly data from carbon dioxide sensors, hourly data from equipment energy consumption, hourly data from passenger flow, and hourly data from temperature. All data points in each category are time series. The hourly data from equipment energy consumption includes hourly energy consumption data for equipment such as chillers, fans, lighting, elevators, and multi-split air conditioning units. Each hourly data point has a corresponding timestamp; for example, hourly energy consumption data includes both the energy consumption data and its timestamp.

[0024] S2. Use the cross-correlation function (CCF) and the fusion fuzzy time cognitive graph model (FTCM) to perform data correction on the multi-dimensional hourly energy consumption data to obtain multi-dimensional hourly energy consumption correction data.

[0025] S2 includes:

[0026] S2-1. Calculate the time delay between any two data types in multi-dimensional hourly energy consumption data. Taking hourly data from carbon dioxide sensors and hourly data from equipment energy consumption as examples, when calculating the time delay between any two hourly data points from carbon dioxide sensors and any two hourly data points from equipment energy consumption, the time delay refers to the time offset of the hourly data from carbon dioxide sensors relative to the hourly data from equipment energy consumption, which can be positive or negative.

[0027] S2-2. Calculate the correlation coefficient of any two data types in multidimensional hourly energy consumption data at different time delays using the cross-correlation function.

[0028] Therefore, the formula corresponding to S2-2 is:

[0029]

[0030] Where a and b represent any two different data types in the multi-dimensional hourly energy consumption data, k represents the latency, and CCF ab (k) represents the correlation coefficient between data types a and b at time delay k, where a i b represents the i-th data point in data type a. i+k μ represents the (i+k)th data point in data type b. a μ b Let σ represent the mean values ​​corresponding to data types a and b, respectively. a σ b Let represent the standard deviations of data types a and b, respectively, n represent the data length, and ∑(·) represent the summation function.

[0031] S2-3. Sort the correlation coefficients of any two data types under different time delays from largest to smallest, select the maximum value of the corresponding correlation coefficient as the correlation result (degree of correlation) of the two data types, and take the time delay corresponding to each correlation result as the time delay estimate.

[0032] S2-4. Based on the operational characteristics of the subway station, various correlation results, and their time delay estimates, determine the data scanning range and dynamically adjust the window width. According to the operational characteristics of the subway station, the scanning range is set to one week, and the initial window width is set to 24 hours to capture daily patterns. The window width is adjusted based on the correlation results and their time delay estimates, ensuring that the window at least includes the time delay estimates corresponding to the correlation results. If the correlation exceeds a correlation threshold, the window width is reduced to improve sensitivity to time delay changes; conversely, the window width is increased. The reduction and increase values ​​of the window width are set according to actual needs.

[0033] S2-5. Based on the data scanning range, the sliding window method is used to calculate the first time delay between the multi-dimensional hourly energy consumption data in the window. This is beneficial for extracting the relationship between the first time delay and correlation results of the multi-dimensional hourly energy consumption data and time, and for analyzing the dynamic changes in the status of various equipment in the subway station.

[0034] S2-6. Based on the first time delay and its correlation coefficient, the multi-dimensional hourly energy consumption data is corrected using the Fusion Fusion Fuzzy Time Cognitive Graph Model (FTCM) to obtain multi-dimensional hourly energy consumption corrected data.

[0035] S2-6 includes:

[0036] S2-6-1. Perform trend segmentation on multi-dimensional hourly energy consumption data to obtain trend units. The process is as follows:

[0037] Obtain the changing trends of multi-dimensional hourly energy consumption data; the changing trends include rising, falling, and stable. Taking the hourly energy consumption data corresponding to wind turbines as an example, this hourly energy consumption data is defined as X=[x1,x2,…,x…]. t ,…,x T Based on the energy consumption values ​​at each time point, calculate the energy consumption difference between adjacent time points; based on each energy consumption difference, divide the hourly energy consumption data; if the energy consumption difference x within a certain continuous time period... t -x t-1 If all values ​​are positive, the trend of change in that continuous time period is upward; if all energy consumption differences in a certain continuous time period are negative, the trend of change in that continuous time period is downward; otherwise, the trend of change in the continuous time period is stable. Where T represents the total duration of the hourly energy consumption data corresponding to the wind turbine, x1, x2, x... t-1 x t x T These represent the energy consumption values ​​corresponding to time 1, time 2, time t-1, time t, and the last time, respectively.

[0038] Based on the changing trends, the starting point of the changing trend is used as the dividing point. A piecewise linear approximation method is used to segment the multi-dimensional hourly energy consumption data, resulting in trend units corresponding to hourly carbon dioxide sensing data, equipment energy consumption data, passenger flow data, and temperature data. Taking the hourly energy consumption data corresponding to wind turbines as an example, the expression for each trend unit is (t... ij,0 ,p ij ,y ij,0 ), p ij t ij,0 y ij,0 These represent the trend slope and time starting point (reflecting the rate of change) of the j-th time period in the i-th hourly energy consumption data, as well as its variable value (vertical intercept / energy consumption value).

[0039] S2-6-2. Determine the linear relationship corresponding to each trend unit; based on each linear relationship, calculate the predicted value of each trend unit. Taking the hourly energy consumption data corresponding to wind turbines as an example, the formula for the linear relationship corresponding to the trend unit is:

[0040] y ij,n (t ij,n ) = p ij (t ij,n -t ij,0 )+y ij,0 ;

[0041] Among them, t ij,n y represents the nth moment of the j-th time period in the i-th hourly energy consumption data. ij,n (t ij,n ) represents time tij,n The corresponding variable value.

[0042] Taking the hourly energy consumption data corresponding to wind turbines as an example, the formula for the predicted value is:

[0043]

[0044] Where, Δt ij,n Represents time t ij,n and the starting point of time t ij,0 The difference, where s represents the number of sampling points. This represents the predicted value corresponding to the nth moment of the j-th time period in the i-th hourly energy consumption data. Since different variables will have different time delays—for example, the time delay between carbon dioxide (CO2) and wind turbine energy consumption data is 3, and the time delay between wind turbine energy consumption and passenger flow is 2—taking wind turbine energy consumption as the benchmark, it is necessary to calculate the difference between CO2 at the current moment and the previous three moments, and the difference between passenger flow at the current moment and the previous two moments. Therefore, each moment corresponding to the time delay between different variables is used as a sampling point, resulting in the number of sampling points, s. In this embodiment, the unit of time delay is hours (h).

[0045] S2-6-3. Calculate the error between each predicted value and the multi-dimensional hourly energy consumption data at the same time, i.e., time t. ij,n The difference between the predicted and the actual variable values ​​is the error.

[0046] S2-6-4. Based on the first time delay, its correlation coefficient, and various errors, determine the nodes, directed edges, edge weights, and fuzzy relationships, and construct a fused fuzzy time cognition graph model, including:

[0047] By analyzing the temporal distribution and magnitude of each error, causal relationships between variables (multi-dimensional hourly energy consumption data) can be identified. For example, if the error of variable A changes synchronously with the error of variable B over time, it indicates a causal relationship between A and B.

[0048] Based on causal relationships, the corresponding first time delay is selected as the time delay information between variables. Nodes and directed edges are determined, and initial values ​​for edge weights are set to establish a directed time graph. Each node represents a variable (e.g., elevator energy consumption, CO2 concentration). Directed edges are established based on causal relationships and time delay information. The direction of the directed edge indicates the causal relationship, and its weight W is calculated from the correlation coefficient and error, with the corresponding formula: W = CCF × e - λ|k|. λ represents the time delay attenuation factor, e represents the error, CCF represents the correlation coefficient, and |·| represents the absolute value.

[0049] A fuzzy relation is established on the directed time graph to obtain a fused fuzzy time cognitive graph model. The trend units corresponding to the directed time graph are mapped, and the corresponding fuzzification trend intensity is determined based on the changing trend (trend slope) in each trend unit. Fuzzy logic mapping is performed on each fuzzy trend intensity to obtain the corresponding dynamic response pattern, which is used as the causal relationship rule. The causal relationship rule is superimposed on the directed time graph to obtain the fused fuzzy time cognitive graph model.

[0050] When new multi-dimensional hourly energy consumption data is input into the fused fuzzy time cognitive graph model, the trend unit needs to be recalculated and the time delay decay factor λ needs to be updated. The edge weights are then optimized through error feedback to achieve dynamic adjustment.

[0051] The process of establishing fuzzy relations is as follows:

[0052] 1. Map each trend unit to a fuzzy set and determine the fuzzy trend strength; the fuzzy trend strength includes [strong upward, weak upward, stable] or [strong downward, weak downward, stable]. Taking [strong upward, weak upward, stable] as an example, when the p of any trend unit... ij When the value is greater than 3, the fuzzy trend strength is defined as "strong upward"; when the p of any trend unit is greater than 3, the fuzzy trend strength is defined as "strong upward". ij If the value is within the range of (0.5, 3], then the fuzziness trend strength is defined as "weak upward"; when the p of any trend unit is within the range of (0.5, 3), the fuzziness trend strength is defined as "weak upward". ij If the value is within the range of (0, 0.5], the fuzziness trend strength is defined as "stationary".

[0053] 2. Define causal relationship rules based on fuzzy trend strength. The fuzzy trend strength (e.g., "strong increase" or "weak decrease") is mapped to a dynamic response pattern to the target variable (e.g., "medium-intensity delayed increase") using fuzzy logic. This dynamic response pattern is then used as the causal relationship rule and stored in the rule base. The dynamic response pattern can supplement the static association of directed graph edge weights, enhancing the ability to express nonlinear relationships. For example, rule 1 (if CO2 concentration "strongly increases," then elevator energy consumption "delays an increase") can achieve adaptive correction of causal strength by dynamically adjusting the edge weights (e.g., W = W0 * 1.5).

[0054] S2-6-5. Input the multi-dimensional hourly energy consumption data into the fusion fuzzy time cognitive graph model and output the multi-dimensional hourly energy consumption correction data.

[0055] An improved wave search algorithm (IWSA) is used to train the fused fuzzy temporal cognitive graph model; such as Figure 2 As shown, the IWSA algorithm includes the following process:

[0056] T1. Set optimization parameters and contribution weights. Optimization parameters include population size, maximum number of iterations, exchange period, similarity threshold, step size adjustment factor, shrinkage factor, initial step size, and weight smoothing factor. The step size adjustment factor is greater than 1, and the shrinkage factor ranges from 0 to 1. The contribution weight represents the contribution of each fractal method in the current search phase, specifically the proportion of each elementary low-dissimilarity sequence that generates a new solution in the next iteration. The updated weights are used to guide the allocation of subsequent search resources, ensuring that better-performing fractal methods receive more resources, thereby exploring the search space more efficiently.

[0057] T2. Initial solutions are generated using N sets of elementary low-discrepancy sequences (fractal method), resulting in N initial solutions and a corresponding candidate pool. In this embodiment, N is set to 4, and the elementary low-discrepancy sequences used are Sobol, Latinhypercube, Halton, and Hammersley. Each initial solution includes the optimal memory factor, coordination coefficient, and association weight matrix of the fused fuzzy temporal cognitive graph model. These initial solutions are uniformly distributed in the high-dimensional space, improving global exploration capabilities.

[0058] T3. Calculate the fitness of each initial solution by fusing the fuzzy temporal cognitive graph model, generate new solutions, and initially update each candidate pool.

[0059] The T3 includes:

[0060] T3-1. Assign the initial solutions generated from each elementary low-discrepancy sequence to the corresponding candidate pool.

[0061] T3-2. The fitness of each initial solution in each candidate pool is calculated by fusing a fuzzy temporal cognitive graph model.

[0062] T3-3, randomly select an elementary low-difference sequence.

[0063] T3-4. Select the parent solution from the candidate pool of the elementary low-dissimilarity sequence; based on the fitness and contribution weight of the initial solution, generate a new solution using the fractal strategy of the elementary low-dissimilarity sequence.

[0064] T3-5. Wave propagation perturbation is performed using an initial step size.

[0065] T3-6. Calculate the fitness of the new solution; retain the one with the highest fitness and store it in the corresponding candidate pool to complete the initial update of the candidate pool.

[0066] T3-7, repeat T3-3 to T3-7, until all elementary low-difference sequences are selected.

[0067] T4. Update contribution weights; select the solution with the highest fitness from each candidate pool and randomly insert it into other candidate pools, and update each candidate pool a second time.

[0068] T5. Merge the solutions in each candidate pool after secondary updates to obtain a global solution set; calculate the Euclidean distance between the solutions in the global solution set; merge or remove solutions in the global solution set according to the Euclidean distance to obtain an updated global solution set.

[0069] The process of merging or eliminating solutions in the global solution set based on each Euclidean distance to obtain the updated global solution set is as follows:

[0070] Two solutions whose Euclidean distance is not less than the similarity threshold are considered similar solutions and merged. If the Euclidean distance is less than the similarity threshold, the solution with higher fitness is retained, and the solution with lower fitness is discarded. In addition, if the number of solutions after merging and discarding is less than the number of initial solutions, new solutions are randomly generated to supplement them.

[0071] T6. Distribute the updated global solution set evenly into each candidate pool of the second update, and calculate the fitness improvement rate of each candidate pool of the second update; based on each fitness improvement rate, update the initial step size. The fitness improvement rate is the value obtained by dividing the difference between the fitness of the candidate pool of the second update and the fitness of the candidate pool before the update by the fitness of the candidate pool before the update.

[0072] T7. Select the solution with the highest fitness in the updated global solution set and update the global optimal solution.

[0073] T8. Repeat T3 to T7 until the maximum number of iterations is reached or the prediction error is minimized, to obtain the global optimal solution and apply it to the fused fuzzy temporal cognitive graph model.

[0074] The improved wave search algorithm complements four fractal methods, avoids the blind spots of traditional random sampling, enhances global search capabilities, and automatically allocates resources according to the search stage, balancing exploration and development and improving the optimization efficiency of FTCM.

[0075] S3. Perform lateral error correction on the multi-dimensional hourly energy consumption correction data to obtain multi-dimensional hourly energy consumption fine-calibration data. Since there is a lateral error between the hourly energy consumption data and the multi-dimensional hourly energy consumption correction data, lateral error correction needs to be performed on the multi-dimensional hourly energy consumption correction data to improve the accuracy of operating condition division. Lateral error refers to the error between the multi-dimensional hourly energy consumption data and the multi-dimensional hourly energy consumption correction data on the horizontal time axis.

[0076] Therefore, S3 includes:

[0077] S3-1. Based on the correlation coefficient and slope in S2, the multi-dimensional hourly energy consumption correction data is segmented to obtain the hourly energy consumption correction segmented data.

[0078] S3-2. Calculate the mutual information of the hourly energy consumption correction segmented data and filter them to obtain the hourly energy consumption correction characteristic data.

[0079] S3-2 includes:

[0080] S3-2-1. Set the number of bins and discretize the hourly energy consumption correction segment data to obtain hourly energy consumption correction discrete data. Number of bins N1 represents the amount of data in the hourly energy consumption correction segment.

[0081] S3-2-2, Calculate the mutual information of each discrete data point in each hourly energy consumption correction discrete dataset. Use energy consumption-related data as the target variable and the remaining data as input features. For example, the input features could be temperature and passenger flow data in the hourly energy consumption correction discrete dataset, and the target variable could be the energy consumption data of each piece of equipment in the hourly energy consumption correction discrete dataset.

[0082] Mutual information I(X) d ;Y d The corresponding formula for ) is:

[0083]

[0084] Where p(·) represents the marginal probability distribution function, log(·) represents the logarithmic function with a base of a constant, and X d Y d Let x represent the input features and the target variable, respectively. d y d These represent the discrete data corresponding to the input features and the target variable, respectively.

[0085] Mutual information can measure the statistical dependence between two variables, which is helpful for selecting key feature variables.

[0086] S3-2-3. Remove the hourly energy consumption correction discrete data that are less than the mutual information threshold to obtain the hourly energy consumption correction feature data. In this embodiment, the mutual information threshold is 0.3.

[0087] S3-3. Perform Savitzky-Golay filtering and transition point detection on the hourly energy consumption correction feature data to obtain the set of transition point locations.

[0088] S3-3 includes:

[0089] S3-3-1. Perform Savitzky-Golay filtering on the hourly energy consumption correction characteristic data to obtain hourly energy consumption correction filtered data, thus avoiding the influence of noise on the data.

[0090] S3-3-2. Traverse the hourly energy consumption correction filtering data and determine the jump point based on the jump point rule.

[0091] The jump point rule is that if the value of the middle data point among three consecutive data points in the hourly energy consumption correction filter data is the maximum or minimum value, then the middle data point is considered the jump point, and the corresponding expression is:

[0092]

[0093] Among them, P i P i+1 P i+2 Y represents the i-th, (i+1)-th, and (i+2)-th data points in the hourly energy consumption correction filter data. t This indicates the threshold for the transition.

[0094] S3-3-3. Determine the position of each transition point to obtain the set of transition point positions.

[0095] S3-4. Divide the trend segments of hourly energy consumption correction characteristic data based on the set of jump point locations and establish the corresponding dynamic trend model.

[0096] S3-4 includes:

[0097] S3-4-1. Based on the set of transition point locations, divide the trend segment boundaries. Regions within a continuous window with consistent slope direction (both positive / both negative) and amplitude variation not exceeding the amplitude threshold are divided into the same trend segment.

[0098] S3-4-2. Based on the trend segment boundaries, divide the hourly energy consumption correction feature data to obtain several trend segments.

[0099] S3-4-3. Establish the initial dynamic trend model for each trend segment. Taking one trend segment as an example, the corresponding initial dynamic trend model formula is: (t jump,start ,t jump,end ,p jump ,τ jump ), where t jump,start t jump,end p represents the start time of this trend segment. jump τ represents the average rate of change of this trend segment. jump This indicates the duration of the trend segment.

[0100] Average rate of change p jump The corresponding formula is:

[0101]

[0102] Where, N jump P(t) represents the length of the trend segment. jump,i+1 ), P(t jump,i ) represent the i-th time t of the trend segment respectively. jump,i At time i+1, t jump,i+1 The corresponding data value.

[0103] S3-4-4: Use the same method as S2 to S3-4-2 to obtain the historical trend segment corresponding to the multi-dimensional hourly energy consumption historical data.

[0104] S3-4-5. Match each historical trend segment and each trend segment to obtain the matching results; update each initial dynamic trend model based on the matching results to obtain the corresponding dynamic trend model.

[0105] The process of S3-4-5 is as follows:

[0106] Calculate the difference in average rate of change and the difference in magnitude of change between historical trend segments and trend segments. Historical trend segments whose average rate of change difference and magnitude of change difference fall within a threshold range are classified as trend segments of the same type. The threshold range is set according to the actual situation.

[0107] Calculate the average duration of each trend segment and its corresponding similar trend segment; based on each average duration, adjust the duration τ. jump The adjusted duration is applied to each initial dynamic trend model to obtain the corresponding dynamic trend model, i.e., (t jump,start ,t jump,end ,p jump ,τ current ).

[0108] Taking any initial dynamic trend model as an example, the adjusted duration τ current The corresponding formula is:

[0109]

[0110] Where, τ hist p represents the average rate of change of similar trend segments in the initial dynamic trend model. current p hist ΔP represents the average rate of change of the trend segment and the similar trend segment in the initial dynamic trend model, respectively. hist ΔP current These represent the magnitude changes of similar trend segments and trend segments in the initial dynamic trend model, respectively.

[0111] S3-5. Perform lateral error correction on the multi-dimensional hourly energy consumption correction data using a dynamic trend model to obtain multi-dimensional hourly energy consumption fine-calibration data.

[0112] S3-5 includes:

[0113] S3-5-1. Calculate the set of state transition time differences between multi-dimensional hourly energy consumption correction data and multi-dimensional hourly energy consumption data; the formula corresponding to the state transition time difference is:

[0114]

[0115] Where, Δt i t represents the time difference between the state transition of the i-th multi-dimensional hourly energy consumption correction data and the corresponding multi-dimensional hourly energy consumption data. jump,pre,i t represents the time corresponding to the transition position of the i-th multi-dimensional hourly energy consumption correction data. jump,real,i This represents the time corresponding to the jump position of the i-th multi-dimensional hourly energy consumption data. When Δt i If the value is greater than 0, then the relationship between the multi-dimensional hourly energy consumption correction data and the multi-dimensional hourly energy consumption data is a prediction lag; when Δt i If the value is less than 0, then the relationship between the multi-dimensional hourly energy consumption correction data and the multi-dimensional hourly energy consumption data is that the prediction is ahead.

[0116] S3-5-2. Fit the probability distribution of the set of time differences between state transitions using the kernel density estimation method, and calculate the corresponding mean deviation and confidence interval.

[0117] S3-5-3. Shift the multi-dimensional hourly energy consumption data along the time axis by the magnitude of the deviation mean, aligning the jump points of the multi-dimensional hourly energy consumption correction data with the jump points of the original multi-dimensional hourly energy consumption data, thus obtaining the updated multi-dimensional hourly energy consumption data. If data breaks occur after the shift, cubic spline interpolation is used to fill in the missing values.

[0118] S3-5-4. Based on the set of state transition time differences and confidence intervals, update the parameters of the dynamic trend model, i.e., update the adjusted duration τ. current .

[0119] When Δt i If it is greater than 0, then τ needs to be shortened. current , making τ current More closely approximating the measured jump point; when Δt i If it is less than 0, then τ needs to be extended. current To match the measured data.

[0120] τ current The update formula is:

[0121] τ new =ατ current+(1-α)τ avg ;

[0122] Where, τ new α represents the updated duration; α represents the weighting coefficient, which is determined by the confidence interval of Δt or the distribution characteristics of the kernel density estimate.

[0123] S3-5-5. Based on the parameters of the updated dynamic trend model, the updated multi-dimensional hourly energy consumption data is corrected to obtain multi-dimensional hourly energy consumption fine-calibration data.

[0124] S4. The TICC algorithm (Inverse Covariance Clustering Algorithm) is used to divide the multi-dimensional hourly energy consumption calibration data to obtain the working condition division results.

[0125] S4 includes:

[0126] S4-1. Set a time window and divide the multi-dimensional hourly energy consumption calibration data into time windows to obtain the corresponding hourly energy consumption window data. The length of the time window is W; the multi-dimensional hourly energy consumption calibration data is divided into T1-W+1 overlapping windows. The sliding step size of adjacent windows is set to 1 hour to ensure time sequence continuity and avoid missing the gradual changes in transition periods. T1 is the sequence length of the multi-dimensional hourly energy consumption calibration data.

[0127] S4-2. Set TICC parameters. TICC parameters include the number of clusters K and the sparsity parameter λ. s and the maximum number of iterations N TICC . λ s The value is 0.1, N TICC The value is 500. The elbow method is used to determine the number of clusters K, which is to calculate the sum of squares within the cluster (WCSS), find the critical point (elbow) where the WCSS decreasing trend changes from steep to gentle, and the value corresponding to the critical point is taken as the number of clusters K.

[0128] Calculate the sum of squares within the cluster WCSS(k) TICC The corresponding formula is:

[0129]

[0130] Where, k TICC C represents the number of clusters. i Let μ represent the time-to-time energy consumption window data set of the i-th cluster. i Let x represent the centroid of the i-th cluster, ||·|| represent the norm, and x TICC This represents hourly energy consumption window data.

[0131] S4-3. Calculate the inverse covariance matrix of each hourly energy consumption window. First, calculate the covariance matrix of each hourly energy consumption window; then, invert each covariance matrix to obtain the inverse covariance matrix. The formula for the covariance matrix is:

[0132]

[0133] Where M1 represents the number of samples in the time window, i.e., the number of samples in the hourly energy consumption window data, x mi x mj Let represent the values ​​of the i-th and j-th dimensions of the m-th hourly energy consumption window data, respectively. Let σ represent the mean of the i-th dimension and the j-th dimension, respectively. ij This represents the covariance matrix corresponding to the i-th and j-th dimensions of the hourly energy consumption window data.

[0134] The diagonal and off-diagonal elements in each covariance matrix reflect the variability and correlation of that covariance matrix, respectively.

[0135] S4-4. With the goal of minimizing the inverse covariance matrix, the TICC algorithm is used to iterate over the inverse covariance matrix, alternating between E-step cluster allocation and M-step inverse covariance matrix updates until convergence or the maximum number of iterations N is reached. TICC The clustering results are obtained.

[0136] For each time window, in the E-step, calculate the cost of belonging to a cluster, and use dynamic programming (Viterbi algorithm) to find the optimal cluster sequence, minimizing the total cost and smoothing term.

[0137] In step M, the inverse covariance matrix is ​​updated. For each cluster, the mean of all window covariance matrices belonging to that cluster is calculated to represent the global features of that cluster. This mean is then input into the GraphicalLASSO algorithm, where weakly correlated clusters are removed using L1 regularization, generating a sparse inverse covariance matrix.

[0138] S4-5. Calculate objective indicators for clustering results; based on these objective indicators, evaluate the clustering results using the analytic hierarchy process (AHP) and entropy weighting method to obtain the evaluation results. Objective indicators include the silhouette coefficient, the Calinski-Harabasz index, and the Davies-Bouldin coefficient.

[0139] S4-5 includes:

[0140] S4-5-1. The clustering results are evaluated in multiple dimensions using the silhouette coefficient, Calinski-Harabasz index, and Davies-Bouldin coefficient to obtain the multi-dimensional evaluation results.

[0141] The three evaluation indicators are weighted using the entropy weight method, and multi-dimensional evaluation results are calculated based on these weights. Specifically, the silhouette coefficient quantifies the cluster compactness and separation by calculating the average distance between a sample point and other data points within the same cluster, as well as the minimum average distance to the nearest cluster; its value ranges from -1 to 1. The Calinski-Harabasz index reflects the balance between inter-cluster separation and intra-cluster compactness by the ratio of inter-cluster dispersion to intra-cluster dispersion. The Davies-Bouldin index measures inter-cluster discriminability based on the ratio of the average intra-cluster distance to the cluster center distance; a smaller value indicates better clustering performance.

[0142] S4-5-2. Using the analytic hierarchy process and based on the professional knowledge of operators regarding energy consumption conditions, a judgment matrix is ​​constructed and the subjective weights of the three indicators are calculated to obtain the subjective evaluation results.

[0143] S4-5-3. Based on the multi-dimensional evaluation results and subjective evaluation results, a weighted average is used to calculate the evaluation result. In this embodiment, the subjective weight is 0.4 and the objective weight is 0.6.

[0144] S4-6. Based on the evaluation results, determine the working condition classification results. Select the clustering result corresponding to the evaluation result with the highest comprehensive coefficient as the working condition classification result.

[0145] In summary, this invention utilizes the cross-correlation function (CCF) and the fused fuzzy temporal cognitive graph model (FTCM) to dynamically analyze the time-delay correlation between multi-dimensional data (such as carbon dioxide concentration, equipment energy consumption, passenger flow, and temperature). It also adjusts the data scanning window according to the characteristics of subway operation to correct data time offsets and improve data synchronization and reliability. Furthermore, it employs a sliding window method and a dynamic window width adjustment strategy to optimize the scanning range based on data fluctuations and correlation results, enhancing the model's adaptability to different operating periods and ensuring the accuracy and robustness of data correction. Finally, it utilizes Savitzky-Golay filtering, transition point detection, and dynamic... The state trend model eliminates lateral errors in multi-dimensional data along the time axis, aligns key transition points, and provides highly consistent, refined data for subsequent operating condition classification. The TICC algorithm (Inverse Covariance Clustering) is used to perform multi-dimensional time-series clustering on the corrected data. The clustering results are optimized using the analytic hierarchy process (AHP) and entropy weighting method, accurately identifying typical energy consumption conditions in subway stations and improving energy efficiency. The wave search algorithm is improved by using a multi-fractal strategy and adaptive resource allocation to accelerate the training process of the fuzzy time cognitive graph model, balancing global exploration and local development capabilities to ensure rapid convergence and optimal solution acquisition in complex high-dimensional data.

[0146] like Figure 3 As shown, a refined energy consumption condition classification system for subway stations includes:

[0147] The multi-dimensional hourly energy consumption acquisition module is used to acquire multi-dimensional hourly energy consumption data of a subway station.

[0148] The data correction module is used to correct multi-dimensional hourly energy consumption data using cross-correlation functions and fused fuzzy time cognitive graph models, so as to obtain multi-dimensional hourly energy consumption corrected data.

[0149] The data calibration module is used to perform lateral error correction on multi-dimensional hourly energy consumption calibration data to obtain multi-dimensional hourly energy consumption calibration data.

[0150] The operating condition segmentation module is used to segment multi-dimensional hourly energy consumption calibration data using the TICC algorithm to obtain the operating condition segmentation results.

[0151] In summary, this system is designed with modules for data acquisition, correction, fine calibration, and operating condition classification. It supports flexible expansion to different subway stations, adapts to diverse operating environments and equipment configurations, and provides a general solution for intelligent energy-saving management of urban rail transit.

[0152] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for fine classification of energy consumption operating conditions of a subway station, characterized in that, The method comprises the following steps: acquiring multi-dimensional hourly energy consumption data of a subway station; the multi-dimensional hourly energy consumption data comprises carbon dioxide sensing hourly data, equipment energy consumption hourly data, passenger flow hourly data and temperature hourly data; performing data correction on the multi-dimensional hourly energy consumption data by using a cross-correlation function and a fusion fuzzy time cognitive graph model to obtain multi-dimensional hourly energy consumption corrected data; performing horizontal error correction on the multi-dimensional hourly energy consumption corrected data to obtain multi-dimensional hourly energy consumption fine correction data; dividing the multi-dimensional hourly energy consumption fine correction data by using a TICC algorithm to obtain working condition division results; the data correction process comprises the following steps: calculating time delays between any two data types in the multi-dimensional hourly energy consumption data; calculating correlation coefficients of any two data types in the multi-dimensional hourly energy consumption data at different time delays by using a cross-correlation function; sorting the correlation coefficients of any two data types at different time delays from large to small, selecting the maximum value of the corresponding correlation coefficient as the correlation result of the two data types, and taking the time delay corresponding to each correlation result as a time delay estimation value; determining a data scanning range and dynamically adjusting the width of a window based on the operation characteristics of the subway station, the correlation results and the time delay estimation values thereof; calculating a first time delay between each multi-dimensional hourly energy consumption data in the window by using a sliding window method based on the data scanning range; performing data correction on the multi-dimensional hourly energy consumption data by using a fusion fuzzy time cognitive graph model based on the first time delay and the correlation coefficient to obtain multi-dimensional hourly energy consumption corrected data; the multi-dimensional hourly energy consumption corrected data is obtained by the following steps: performing trend segmentation on the multi-dimensional hourly energy consumption data to obtain trend units; determining linear relationships corresponding to the trend units; calculating prediction values of the trend units based on the linear relationships; calculating errors between each prediction value and the multi-dimensional hourly energy consumption data at the same time; determining nodes, directed edges, edge weights and fuzzy relationships based on the first time delay, the correlation coefficient and the errors, and constructing a fusion fuzzy time cognitive graph model; inputting the multi-dimensional hourly energy consumption data into the fusion fuzzy time cognitive graph model to output the multi-dimensional hourly energy consumption corrected data.

2. The method for fine classification of energy consumption operating conditions of a subway station according to claim 1, characterized in that, The fusion fuzzy time cognitive graph model is trained by using an improved wave search algorithm, which comprises the following processes: T1, setting optimization parameters and contribution weights; T2, generating initial solutions corresponding to N groups of elementary low-discrepancy sequences to obtain N groups of initial solutions, and generating corresponding candidate pools; T3, calculating the fitness of each initial solution by using the fusion fuzzy time cognitive graph model, generating a new solution, and updating each candidate pool; T4, updating the contribution weights; selecting the solution with the highest fitness in each candidate pool and randomly inserting it into other candidate pools to update each candidate pool twice; T5, merging the solutions in each twice-updated candidate pool to obtain a global solution set and updating the global solution set to obtain an updated global solution set; T6, dividing the updated global solution set into each twice-updated candidate pool, calculating the fitness improvement rate of each twice-updated candidate pool, and updating the initial step length based on the fitness improvement rates. T7, selecting the solution with the highest fitness in the updated global solution set, updating the global optimal solution; T8, repeating T3 to T7 until the maximum number of iterations is reached or the prediction error is minimized, obtaining the global optimal solution and applying it to the fused fuzzy time cognitive graph model.

3. The method for fine classification of energy consumption operating conditions of a subway station according to claim 1, characterized in that, The process of transverse error correction is: Segmenting the multi-dimensional hourly energy consumption correction data to obtain hourly energy consumption correction segmented data; Calculating the mutual information of the hourly energy consumption correction segmented data and performing screening to obtain hourly energy consumption correction feature data; Performing Savitzky-Golay filtering and jump point detection on the hourly energy consumption correction feature data to obtain a set of jump point positions; Dividing the trend segments of the hourly energy consumption correction feature data based on the set of jump point positions and establishing corresponding dynamic trend models; Performing transverse error correction on the multi-dimensional hourly energy consumption correction data through the dynamic trend models to obtain multi-dimensional hourly energy consumption fine correction data.

4. The method for fine division of energy consumption operating conditions of a subway station according to claim 3, characterized in that, The process of dividing the trend segments of the hourly energy consumption correction feature data based on the set of jump point positions and establishing corresponding dynamic trend models includes: Based on the set of jump point positions, dividing the trend segment boundaries; Based on the trend segment boundaries, dividing the hourly energy consumption correction feature data to obtain several trend segments; Establishing initial dynamic trend models for each trend segment; Obtaining historical trend segments corresponding to the multi-dimensional hourly energy consumption historical data; Matching each historical trend segment with each trend segment to obtain a matching result; updating each initial dynamic trend model based on the matching result to obtain the corresponding dynamic trend model.

5. The method for fine classification of energy consumption operating conditions of a subway station according to claim 4, characterized in that, The process of performing transverse error correction on the multi-dimensional hourly energy consumption correction data through the dynamic trend models to obtain multi-dimensional hourly energy consumption fine correction data includes: Calculating the state jump time difference set of the multi-dimensional hourly energy consumption correction data and the multi-dimensional hourly energy consumption data; Fitting the probability distribution of the state jump time difference set through kernel density estimation, calculating the corresponding bias mean and confidence interval; Shifting the multi-dimensional hourly energy consumption data along the time axis by the size of the bias mean, aligning the jump points of the multi-dimensional hourly energy consumption correction data with the jump points of the multi-dimensional hourly energy consumption data to obtain updated multi-dimensional hourly energy consumption data; Updating the parameters of the dynamic trend model based on the state jump time difference set and the confidence interval; Based on the parameters of the updated dynamic trend model, correcting the updated multi-dimensional hourly energy consumption data to obtain multi-dimensional hourly energy consumption fine correction data.

6. The method for fine division of energy consumption operating conditions of a subway station according to claim 3, characterized in that, Using the TICC algorithm to divide the multi-dimensional hourly energy consumption fine correction data to obtain the working condition division result, including: Setting a time window and dividing the multi-dimensional hourly energy consumption fine correction data into time window data; Setting TICC parameters; TICC parameters include the number of clusters, sparsity parameters, and the maximum number of iterations; Calculating the inverse covariance matrix of each hourly energy consumption window data; Using the TICC algorithm to iterate the inverse covariance matrix to obtain clustering results, with the objective of minimizing the inverse covariance matrix; Calculating the objective indicators of the clustering results; based on the objective indicators, evaluating the clustering results through the analytic hierarchy process and entropy weight method to obtain evaluation results; Based on the evaluation results, determining the working condition division result.

7. A subway station energy consumption working condition fine division system for implementing the subway station energy consumption working condition fine division method of any one of claims 1 to 6, characterized in that, ​ The multi-dimensional hourly energy consumption collection module is used to acquire multi-dimensional hourly energy consumption data of a certain subway station; the multi-dimensional hourly energy consumption data includes carbon dioxide sensing hourly data, equipment energy consumption hourly data, passenger flow hourly data and temperature hourly data; The data correction module is used to correct the multi-dimensional hourly energy consumption data by using a cross-correlation function and a fusion fuzzy time cognitive graph model to obtain multi-dimensional hourly energy consumption correction data; The data fine correction module is used to correct the multi-dimensional hourly energy consumption correction data in a horizontal error correction manner to obtain multi-dimensional hourly energy consumption fine correction data; The working condition division module is used to divide the multi-dimensional hourly energy consumption fine correction data by using a TICC algorithm to obtain a working condition division result.

Images