Flexible load grading method based on multi-dimensional feature analysis

Abstract

The invention discloses a flexible load grading method based on multi-dimensional feature analysis, and the method comprises the following steps: data acquisition and preprocessing: collecting historical and real-time power utilization data of various loads in a building, and carrying out the preprocessing of the historical and real-time power utilization data; constructing a multi-dimensional feature index; load preliminary classification and adjustable load classification: establishing a multi-dimensional classification model according to the multi-dimensional characteristic indexes in the adjustable load classification, and performing classification on the preliminarily classified loads through the multi-dimensional classification model to obtain preliminary classification and classification results; the step of grading label output and regulation and control suggestion generation comprises the sub-item load proportion statistical report and the time sequence distribution chart; and outputting the graded various load lists and corresponding regulation and control suggestions. Compared with the prior art, the flexible load grading model based on the characteristic parameters is constructed through multi-dimensional characteristic analysis, grading optimization and result output of the building load, and the flexible load grading model has the advantages of being high in flexibility, fine in regulation and control capacity and high in response speed.

Description

Technical Field

[0001] The present invention relates to the field of load regulation in a power supply and distribution system, and particularly to a flexible load classification method based on multi-dimensional feature analysis. Background Art

[0002] With the rapid development of social economy and the continuous growth of power demand, the problem of peak-valley difference in electricity consumption has become increasingly prominent, and building energy consumption has become one of the main sources of urban energy consumption. During peak electricity consumption periods such as summer peak load periods, traditional demand-side management often adopts simple and extensive load reduction measures, such as forcibly cutting off loads such as air conditioners, lighting, or charging piles to relieve the operation pressure of the power grid. However, this direct load cutting method not only reduces the operation efficiency of the building, but also significantly affects the comfort and experience of users. Especially for office buildings with strong functionality and complex load distribution, extensive measures may seriously interfere with normal office activities and cannot meet the requirements of refined management for modern buildings.

[0003] In the context of the construction of a smart energy system, virtual power plant technology provides a new technical path for the optimal regulation of flexible loads. A virtual power plant aggregates dispersed load resources to achieve power grid peak shaving, frequency modulation trading, and balance regulation. However, the existing regulation schemes for flexible loads in office buildings in virtual power plants still have the following main deficiencies:

[0004] (1) Extensiveness of load classification and response: The existing technology fails to fully analyze and utilize the characteristics of different types of loads in buildings and lacks refined classification management of various loads. There are many types of loads in office buildings, and the response speeds, reduction potentials, and importance to building functions of different loads vary significantly. If loads cannot be scientifically classified, regulation strategies often tend to be "one-size-fits-all", resulting in low building operation efficiency.

[0005] (2) Absence and misjudgment of baseline load: An accurate baseline load is a prerequisite for flexible load regulation. However, when constructing a baseline load curve, the existing methods fail to fully consider the dynamic characteristics of electricity consumption in office buildings and the load fluctuation law, resulting in baseline misjudgment and further affecting the accuracy and effect of flexible regulation.

[0006] (3) Lack of load classification and priority strategies: There are significant differences in the importance and regulation potential of different loads to building functions. For example, the load of the air conditioning system has a large regulation potential, but the regulation range is restricted by thermal comfort; the lighting system can be partially reduced by lowering the brightness, but excessive reduction will directly affect the suitability of the working environment; the charging pile load needs to seek a balance between user needs and power supply restrictions. The existing schemes lack classification strategies for these loads and fail to combine the response potential with the operation priority.

[0007] (4)Insufficient dynamic response ability: Most of the existing regulation schemes are statically designed and are difficult to respond to the peak shaving instructions of the power grid and market price signals in real time. In particular, the flexible loads in office buildings often have complex dynamic characteristics, and traditional methods are difficult to meet the requirements of fast, accurate, and flexible regulation. Summary of the Invention

[0008] The purpose of the present invention is to provide a flexible load classification method based on multi-dimensional feature analysis. The technical problem to be solved is to scientifically and refinedly classify loads, providing support for the demand response of virtual power plants and peak shaving and valley filling of the power grid.

[0009] To solve the above problems, the present invention is implemented by adopting the following technical solutions: A flexible load classification method based on multi-dimensional feature analysis, including the following steps:

[0010] Step S101, data acquisition and preprocessing, collect historical and real-time electricity consumption data of various loads in the building and preprocess them;

[0011] Step S102, construction of multi-dimensional feature indicators, extract load characteristics based on the time dimension, space dimension, and response characteristic dimension. The multi-dimensional feature indicators include: time dimension characteristics, space dimension characteristics, regulation ability dimension characteristics, functional necessity characteristics, operation stability characteristics, and environmental dependence characteristics;

[0012] Step S103, preliminary load classification and adjustable load classification. Among them, the preliminary load classification initially divides the loads into adjustable loads, non-adjustable loads, and special loads. The adjustable load classification establishes a multi-dimensional classification model according to the multi-dimensional feature indicators, and classifies the initially classified loads through the multi-dimensional classification model to obtain the initially classified and classified results;

[0013] Step S105, output of classification labels and generation of regulation suggestions, including: generating a statistical report on the proportion of sub-item loads and a time series distribution chart; outputting a list of various classified loads and corresponding regulation suggestions.

[0014] Further, between step S103 and step S105, there is also step S104, optimization of classification results. After obtaining the preliminary classification results, optimize the preliminary classification results. After optimizing the classification results, perform deep learning on the historical loads and the optimized classification results through a neural network model, and finally obtain the optimized classification labels.

[0015] Further, it also includes step S106, effectiveness evaluation and optimization: After the response ends on the response day, conduct a comprehensive evaluation according to the actual adjustment amount, adjustment flexibility, economic subsidy income, and internal feedback of the building, and iteratively optimize the classification results according to the evaluation results.

[0016] Further, the preprocessing in step S101 includes abnormal data rejection, data interpolation, and baseline load fitting.

[0017] Further, based on step S101, more historical load data is added, or weighted processing is performed on certain special days, which is implemented by the following formula:

[0018]

[0019] where Q 基线(t) is the baseline load at time t, Q i (t) is the load data on the i-th day, and ε is the correction factor.

[0020] Further, in step S103, threshold method or clustering algorithm is used for preliminary load classification, where:

[0021] When the mean and standard deviation of the load fluctuation range satisfy the following relationship, the partition threshold method is adopted:

[0022]

[0023] where σx is the standard deviation of the load fluctuation range, and μx is the mean of the load fluctuation range;

[0024] When the mean and standard deviation of the load fluctuation range satisfy the following relationship, the hierarchical clustering algorithm is adopted:

[0025]

[0026] Further, the establishment of the multi-dimensional hierarchical model in step S103 is specifically to construct a high-dimensional original feature vector X i = [response delay, load operation cycle, load fluctuation frequency...], and after dimensionality reduction processing, the dynamic evolution process of flexible load is modeled through a discrete-time finite-state Markov chain model to obtain a multi-dimensional hierarchical model.

[0027] Further, the modeling of the dynamic evolution process of flexible load by the discrete-time finite-state Markov chain model is specifically as follows: Define the state set S = {S1, S2, S3}, corresponding to level-I, level-II, and level-III flexible loads respectively. By annotating the multi-dimensional characteristic indexes of the target load in the historical operation data, its historical state sequence is constructed, and a state transition probability matrix P = [P ij is established based on the state transition statistical frequency, where P ij represents the probability that the load transfers from state i to state j;

[0028] Through the Markov chain steady-state analysis method, solve the condition πP = π, ∑ i π iThe long-term stationary distribution vector π = [π1, π2, π3]. According to the principle of the maximum component,

[0029] If π1 = max(π), it is determined as a Class I load;

[0030] If π2 = max(π), it is determined as a Class II load;

[0031] If π3 = max(π), it is determined as a Class III load.

[0032] Furthermore, in the step S104, to optimize the preliminary classification result, the fuzzy logic rule is used to optimize and correct the classification result, which is implemented by the following formula:

[0033] F final = max(μ adjustable , μ non-adjustabl , μ special )

[0034] where μ is the membership function;

[0035] If μ adjustable is the maximum, then this load is officially marked as "adjustable";

[0036] If μ non-adjustabl is the maximum, then this load is officially marked as "non-adjustable";

[0037] If μ special is the maximum, it is listed as a "special load" and does not enter the conventional regulation pool;

[0038] After optimizing and correcting the classification result, the optimized classification label is obtained.

[0039] Furthermore, in the step S105, the regulation suggestions include: for adjustable loads, a multi-dimensional scoring method is used for quantitative evaluation. When there are multiple adjustable loads with the possibility of regulation at the same time, the one with a higher Score value is preferentially selected for regulation; the Score value is calculated as follows:

[0040] Score = ω1·F 必要性 + ω2·F 灵活性 + ω3·F 响应性

[0041] ω1 + ω2 + ω3 = 1

[0042] where ω1, ω2, ω3 are weight coefficients. Optionally, they are set according to experience as: ω1 = 0.4, ω2 = 0.3, ω3 = 0.3.

[0043] Compared with the prior art, the present invention constructs a flexible load classification model based on characteristic parameters through multi-dimensional characteristic analysis, hierarchical optimization and result output of building loads, which has the effects of high flexibility, strong fine-tuning ability and fast response speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of the present invention.

[0045] Figure 2 is the daily sub-item baseline load in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] The present invention will be further described in detail below with reference to the drawings and embodiments.

[0047] As Figure 1 shown, the present invention discloses a flexible load classification method based on multi-dimensional characteristic analysis, including the following steps:

[0048] Step S101, data acquisition and preprocessing.

[0049] Collect historical and real-time electricity consumption data of various loads in the building and preprocess them.

[0050] Specifically, obtain historical load data of several normal working days before the target response day (the target response day includes the response period). After obtaining the historical load data, preprocess the data to obtain the preprocessed historical load data. The preprocessing includes removing abnormal data, data interpolation and baseline load fitting.

[0051] As Figure 2 shown, it is the baseline fitting curves of various loads in an office building before a specified response day in one embodiment of the present invention. The figure shows the proportion of the baseline load of key flexible loads such as air conditioners, lighting, and charging piles during the response period from 9:00 to 18:00, providing a data basis for subsequent classification and control strategy formulation.

[0052] The load data is the electricity consumption load (electricity consumption data) of the electrical equipment. The electricity consumption load refers to the load of each sub-item or loop of the classification and control object. That is to say, it can be an air conditioner loop, a lighting loop, or some charging pile loops.

[0053] The load data can be obtained through the following methods:

[0054] 1. Through the building automation system (BAS): Real-time collect the electricity consumption load data of the sub-items of electrical equipment such as air conditioners, lighting, power, sockets, and charging piles;

[0055] 2. The power consumption load of electrical equipment on its line can be collected through an electricity meter or a sub-meter collection system. Smart electricity meters can be deployed on each main load branch to record the sub-item power curve (typical sampling interval: 15 minutes or 1 hour).

[0056] 3. Record the power consumption load of special loads (such as large computer room equipment) that have not been included in the automated monitoring by on-site recording.

[0057] The target response day is the specific date when the power grid or the virtual power plant operator sends a demand response call to the building side.

[0058] The abnormal data can be eliminated by adopting the 3σ principle or an outlier detection algorithm based on time series.

[0059] Lagrange interpolation method or time series prediction model is used for data interpolation to ensure data continuity and integrity.

[0060] For the baseline load fitting, one or more of the moving average method, weighted least squares method, polynomial fitting, and Markov chain prediction can be adopted. Optionally, the moving average method or the weighted least squares method can be used. The formula for the weighted least squares method is as follows:

[0061] Q baseline (t) = ɑ0 + α1t + ∈

[0062] Where Q baseline (t) is the baseline load, t is the time, ɑ0 and ɑ1 are regression coefficients, and ∈ is the random error.

[0063] By adopting one or more data fitting algorithms, the baseline load curve for the response period is obtained.

[0064] In the present invention, the algorithm adopted for the baseline load fitting can be selected according to the characteristics of the electrical equipment. Among them, the moving average method and the weighted least squares method are suitable for the situation where the data is stable and the fluctuation is small, and the baseline curve can be quickly obtained. The load standard deviation σ can be calculated. If σ does not exceed 10% of the mean value in the selected sample period (such as 5 working days), it is considered that the load data is stable and the fluctuation is small.

[0065] Polynomial fitting is suitable for data with a certain periodicity or trend, and can capture non-linear changes. When the first-order or second-order polynomial regression is performed on the load curve, if the fitting determination coefficient R 2 ≥0.85, the residual is <5%, and the slope term is significant, it is determined that there is a trend, and polynomial fitting can be adopted.

[0066] Markov chain prediction is suitable for considering the state transition of the load in a short period of time, if there are significant discrete states in the building load (such as different modes during the day / night). When the autocorrelation cycle peak ACF(T) ≥ 0.6 or the spectral energy is concentrated in the main frequency, and the cycle T is stable (such as the regular change every weekday), it is determined as a periodic load, and cycle modeling or Markov prediction can be used.

[0067] On the basis of step S101, to ensure the accuracy and representativeness of the data, more historical load data can be added, or weighted processing can be carried out on some special days (such as holidays or special climate days). Specifically, it can be achieved by the following formula:

[0068]

[0069] Among them, Q 基线(t) is the baseline load at time t, Q i (t) is the load data on the i-th day, and ε is the correction factor.

[0070] Step S102, construction of multi-dimensional characteristic indexes.

[0071] Extract load characteristics based on the time dimension, space dimension and response characteristic dimension.

[0072] The construction of multi-dimensional characteristic indexes is used to extract and model the characteristics of various adjustable or potentially adjustable electricity loads (such as air conditioners, lighting, charging piles, etc.) in office buildings. Specifically, for each sub-item electricity load or the electricity load of each circuit (such as the main air conditioner unit, a regional lighting circuit, a charging pile circuit, etc.), real-time power or energy consumption data is collected respectively to form a "load time series". Subsequently, these load time series are analyzed and mined in six dimensions such as the time dimension, space dimension, and regulation ability dimension, so that the hierarchical model can correctly identify and classify different loads.

[0073] The multi-dimensional characteristic indexes specifically include:

[0074] 1. Time dimension characteristics, used to describe the operation behavior and response characteristics of the load, specifically including 4 items:

[0075] Response delay (the time required for the load to actually complete the response from receiving the regulation instruction, unit: min); load operation cycle (whether there is a typical peak distribution within a day / week); load fluctuation frequency (obtaining the main frequency component through frequency domain analysis); load duration ratio (the proportion of the time period when the load is higher than the set threshold).

[0076] 2. Space dimension characteristics, used to describe the functional area distribution and relative importance of the load inside the building,

[0077] Specifically including 3 items:

[0078] Area attribution (such as office area, meeting room, garage); Area load weight (the proportion of area load in the total load); Safety level label (whether it is a load area for security).

[0079] 3. Regulation ability dimension features, used to describe the adjustable and response efficiency of the load, specifically including 3 items:

[0081] Adjustment potential (the ratio of the load that can be reduced to the baseline load); Regulation mode type (such as power-off control, dimming control, V1G dynamic power regulation, etc.); Response success rate and historical lag time characteristics.

[0082] 4. Functional necessity features, reflecting the impact degree of the load on the core operation of the building and personnel activities,

[0083] Specifically including 3 items:

[0084] The importance level of the used area (core office / auxiliary space); The impact intensity on comfort and task continuity; Interruptibility or interruption tolerance time.

[0085] 5. Operation stability features, used to judge the fluctuation degree and predictability of the load curve, specifically including 3 items:

[0086] Standard deviation σ and mean μ; Coefficient of variation σ / μ; Periodic peak of autocorrelation function ACF, etc.

[0087] 6. Environmental dependence features, used to describe the coupling intensity between the load and external environmental variables, specifically including 3 items:

[0088] The correlation between air-conditioning power consumption and outdoor temperature; The dependence relationship between lighting load and natural light intensity; The correlation between charging pile load and vehicle battery state or usage rate.

[0089] Step S103, preliminary classification of loads and grading of adjustable loads.

[0090] For the preliminary classification of loads, use the threshold method or clustering algorithm to preliminarily divide the loads into adjustable loads, non-adjustable loads and special loads, where:

[0091] When the mean and standard deviation of the load fluctuation range satisfy the following relationship, the partition threshold method can be selected:

[0092]

[0093] Among them, σx is the standard deviation of the load fluctuation range, and μx is the mean of the load fluctuation range. That is, when the load fluctuation is relatively stable, the partition threshold method can be directly used.

[0094] Through the partition threshold method, the loads in the building are divided into the following three categories:

[0095] i. Adjustable load: large fluctuation range and the load with a response time less than 10 minutes, including air-conditioning load, lighting load in leased areas (such as office areas, meeting rooms and other office or commercial places), and charging pile load;

[0096] ii. Non-adjustable load: small fluctuation range or rigid load, specifically office equipment sockets and special production equipment, and such loads cannot be cut arbitrarily;

[0097] iii. Special load: including electrical equipment related to safety and key functions (fire protection system, placement system and special production equipment), and such loads cannot be interrupted.

[0098] When the mean and standard deviation of the load fluctuation range satisfy the following relationship, the hierarchical clustering algorithm is adopted:

[0099]

[0100] That is, when the load data fluctuates greatly and the classification rules are complex, the hierarchical clustering algorithm based on similarity calculation is preferably selected. A load similarity matrix is constructed, and the calculation formula is as follows:

[0101]

[0102] where X i 、X j are the characteristic vectors of two loads, and σ is the clustering radius parameter, which can generally be set to 0.5 - 1.0.

[0103] Initially, each load is an independent class; for any two loads i and j, when S(i,j)≥0.85, they are considered to have highly similar characteristics and are merged into the same class; repeated iteration is performed until the similarity between classes is lower than the threshold (0.85), and the clustering is completed. After the clustering is completed, the overall characteristics of each class are labeled as follows:

[0104] Adjustable load: response time ≤ 10min.

[0105] Non-adjustable load: response time > 30min.

[0106] Special load: the area where it is located is the core safety area, fire protection system or related to key tasks; does not participate in flexible regulation; cannot be interrupted or cut.

[0107] Through this process, the internal loads of the building can be classified and determined according to the fluctuation characteristics of the loads, laying a foundation for the subsequent grading work.

[0108] For the classification of adjustable loads:

[0109] Based on the multi-dimensional characteristic indexes of S102, a multi-dimensional classification model is established. Specifically, establishing the multi-dimensional classification model is to construct a high-dimensional original feature vector X i = [response delay, load operation cycle, load fluctuation frequency...], and optionally, the principal component analysis (PCA) method is used to perform dimensionality reduction processing on it. By extracting the first several principal components through PCA (the cumulative explained variance exceeds 90%), the main distinguishing information can be retained and the modeling calculation complexity can be reduced.

[0110] On this basis, a discrete-time finite-state Markov chain model is further introduced to model the dynamic evolution process of flexible loads.

[0111] First, define the state set S = {S1, S2, S3}, corresponding to level-I, level-II, and level-III flexible loads respectively. By annotating the multi-dimensional characteristic indexes of the target load in the historical operation data, its historical state sequence is constructed, and a state transition probability matrix P = [P ij is established based on the state transition statistical frequency, where P ij represents the probability that the load transfers from state i to state j.

[0112] Furthermore, the Markov chain steady-state analysis method is used to solve the long-term stationary distribution vector π = [π1, π2, π3] that satisfies the conditions πP = π, ∑ i π i , which is used to reflect the long-term attribution probability of the load in each classification state. According to the maximum component principle, if π1 > π2 and π1 > π3, it is determined that the load finally belongs to level-I load; if π3 is the largest, it is a level-III load. Specifically expressed as:

[0113] If π1 = max(π), it is determined as a level-I load (high response, strong regulation ability);

[0114] If π2 = max(π), it is determined as a level-II load (moderate response ability);

[0115] If π3 = max(π), it is determined as a level-III load (slow response, low regulation potential, only intervenes in emergencies).

[0116] The multi-dimensional classification model is used to systematically evaluate the reduction potential, response speed, and importance of each electrical load, providing a scientific basis for subsequent demand response strategies; before the execution of the regulation task, through this model, it can quickly locate which loads are the most valuable for regulation.

[0117] Finally, the preliminary classification and grading results are obtained.

[0118] Step S104, optimization of classification results.

[0119] After obtaining the preliminary classification results, it is necessary to optimize the preliminary classification results and use fuzzy logic rules to optimize and correct the classification results. The specific formula is as follows:

[0120] F final = max(μ adjustable , μ non-adjustable , μ special )

[0121] where μ is the membership function.

[0122] If μ adjustable is the maximum, then mark this load as "adjustable" officially;

[0123] If μ non-adjustab is the maximum, then mark this load as "non - adjustable" officially;

[0124] If μ special is the maximum, then classify it as "special load" and do not put it into the conventional regulation pool;

[0125] If the difference between the maximum value and the second - largest value in the output result of the membership function of a certain load is less than the set threshold δ (preferably δ = 0.15), then it is determined that there is classification ambiguity for this load. At this time, further judgment is combined with a recurrent neural network. Specifically:

[0126]

[0127] where (a, b, c, d) are 4 adjustable thresholds, which can be initially set to (0, 0.2, 0.8, 1.0) respectively;

[0128]

[0129] where (e, f) are 2 adjustable thresholds, which can be initially set to (-0.2, 0.1) respectively;

[0130]

[0131] Iterate in the recurrent neural network until the difference between the maximum value and the second - largest value of μ adjistable , μ non-adjustable , μ special is greater than or equal to the set threshold δ, and obtain the optimized classification labels, namely adjustable load, non - adjustable load, and special load.

[0132] Verify the classification results through a neural network model (such as LightGBM or RNN) and optimize the classification rules. During optimization, obtain the time characteristics of the load (including the time series of the load, etc.), spatial characteristics (including the distribution of the load in different areas of the building, etc.), load characteristics (average load value, maximum load value, minimum load value, fluctuation range), and environmental constraint characteristics (temperature, humidity, luminance, current charging pile utilization rate, vehicle battery status, etc.) through the system and input them into the neural network. Output the classification label of the load (such as adjustable load, non-adjustable load, special load), and predict the future load change trend according to the real-time characteristics of the load (such as load reduction potential, reduction risk, etc.).

[0133] When the load characteristics exist in a static or low-frequency discrete form, that is, the change frequency is lower than 1 time per hour, the LightGBM model is preferably used at this time; when the load characteristics include continuous time series changes, that is, the change frequency is higher than 1 time per hour, and it is necessary to identify the sequence pattern or predict the trend, the RNN model is preferably used.

[0134] Step S105, Output classification labels and generate regulation suggestions.

[0135] Outputting classification labels and generating regulation suggestions includes:[[]]

[0136] Generate a statistical report on the proportion of sub-loads and a time series distribution chart;

[0137] Output the list of various loads after classification and the corresponding regulation suggestions.

[0138] The regulation suggestions include:[[]]

[0139] For adjustable loads, a multi-dimensional scoring method is used for quantitative evaluation. When multiple adjustable loads have the possibility of regulation at the same time, the one with a higher Score value is preferably selected for regulation; for example, in the same building, the Score of the air-conditioning load is 4.7 and the lighting is 3.9, and the air-conditioning load is preferably regulated first. The calculation formula for the Score value is as follows:

[0140] Score = ω1·F 必要性 + ω2·F 灵活性 + ω3·F 响应性

[0141] ω1 + ω2 + ω3 = 1

[0142] Among them, ω1, ω2, and ω3 are weight coefficients. Optionally, they are set according to experience as: ω1 = 0.4, ω2 = 0.3, ω3 = 0.3. Of course, they can also be determined based on multiple rounds of regulation experiments or machine learning optimization algorithms. F 必要性 、F 灵活性 、F 响应性They respectively represent the scores of the load in terms of necessity, flexibility, and responsiveness.

[0143] For F 必要性 , (value range: 1 - 5) is used to reflect the importance of this load to the operation of the building's core functions, and is defined as follows: It can be determined in combination with the building zoning management system and the division of space usage functions. If this load involves the core office area or high-value production area, the score is relatively high, otherwise it is relatively low; for F 灵活性 , (value range: 1 - 5) is used to reflect the importance of this load to the operation of the building's core functions, and is defined as follows: The above can be determined in combination with the building zoning management system and the division of space usage functions. For example, a high score can be given according to the load adjustment ratio (such as > 40%), otherwise a medium / low score is given; for F 响应性 , (value range: 1 - 5) is assigned according to the time delay from receiving the instruction to completing the response of this load, as follows: The response time delay is obtained from the device control logic or previous response records. For example, a high score (flexible) can be given for a short response time (≤ 5 min), and a low score for a long response time (> 30 min).

[0144] To enhance the unity and enforceability of the quantitative scoring, the value definitions for 1 - 5 points in the scoring criteria are now as follows:

[0145] 1 point: Extremely low level (e.g., adjustment ratio < 10%, involving critical safety loads, response time ≥ 30 min, etc.).

[0146] 2 points: Relatively low level (e.g., 10% ≤ adjustment ratio < 20%, relatively high regional importance, 20 min ≤ response time < 30 min).

[0147] 3 points: Medium level (e.g., 20% ≤ adjustment ratio < 30%, 10 min ≤ response time < 20 min).

[0148] 4 points: Relatively high level (e.g., non-critical area, 30% ≤ adjustment ratio < 40%, 5 min ≤ response time < 10 min).

[0149] 5 points: Extremely high level (e.g., non-critical area, adjustment ratio > 50%, response time < 5 min).

[0150] Score can also be used to generate the recommended adjustment range of the load. For example:

[0151] When Score ∈ [4, 5], the allowed adjustment range ≥ 50%; when Score ∈ [3, 4), the adjustment range is [30%, 50%]; when Score ∈ [2, 3), the adjustment range is [10%, 30%); when Score < 2, try not to adjust or only enable it during extreme peak shaving.

[0152] Step S106, Effectiveness Evaluation and Optimization:

[0153] After the system response ends on the response day, a comprehensive evaluation is conducted based on the actual adjustment amount, adjustment flexibility, economic subsidy benefits, and internal building feedback (such as personnel comfort and satisfaction with actual needs). The classification results are iteratively optimized according to the evaluation results to continuously improve the adjustment performance and economic benefits in subsequent regulation events. The iterative optimization is specifically achieved by regularly adjusting the weight parameters, sub - load ratios, and regulation thresholds to continuously adapt to the dynamic changes in electricity prices and energy consumption patterns. The formula is as follows:

[0154] ΔScore = α·ΔQ 调节 +β·Δ 灵活性 -γ·Δ 用户影响

[0155] where α, β, and γ are optimization weight parameters.

[0156] ΔQ 调节 refers to the actually reduced load (or how much more / less than expected), which can be obtained by comparing the baseline load and the measured load curve, i.e., ΔQ 调节 =Q target -Q actual ; Δ 灵活性 refers to whether the adjustment potential has increased / decreased after the actual load regulation compared to the originally estimated adjustment potential.

[0157] It can be defined based on the change in the "adjustable potential" score before and after regulation: Δ 灵活性 =F 灵活性(after) -F 灵活性(befor) ; Δ 用户影响 refers to the change in user satisfaction / comfort.

[0158] It can come from: indoor thermal comfort (such as changes in PMV value), lighting complaint rate, vehicle charging waiting time; or subjective questionnaires or operation and maintenance records (number of complaints, survey scores, etc.).

[0159] If ΔScore>0, it means that the benefits of the reduction amount (ΔQ 调节 ) and flexibility (Δ 灵活性 ) are greater than the user dissatisfaction (Δ 用户影响 ), then the current classification and regulation suggestions are retained; if ΔScore<0, it means that this strategy has too much impact on users, or the reduction amount is insufficient, resulting in a decrease in the score. At this time, it is necessary to return to S104 or the next cycle to update the parameters (such as reducing the adjustment amplitude) for iteration.

[0160] If ΔScore>0 after iteration, the classification is exited; if ΔScore<0 after three iterations, this regulation is terminated.

[0161] Through the above steps, the present invention realizes the refined and efficient regulation of flexible loads in the power supply and distribution system of office buildings, while meeting the peak shaving requirements of the power grid, minimizing the impact on the normal operation of the building and the comfort of personnel to the greatest extent. Compared with the traditional method of forcibly cutting off loads, the present invention not only has higher flexibility, can optionally achieve hierarchical and quantitative regulation, but also preferably has the ability of rapid response, better meeting the requirements of virtual power plants and intelligent energy systems, and creating greater value for office buildings in energy management and power market transactions.

[0162] The present invention can be used in a variety of scenarios in office buildings, including but not limited to: typical demand response scenarios such as air-conditioning load reduction, lighting brightness regulation, and charging pile load optimization. For the load management of high-energy-consuming office buildings, through flexible regulation strategies and load classification methods, effective demand response resources are provided for virtual power plants, improving the tense situation of power supply and enhancing energy utilization efficiency.

[0163] Through a clear classification method, refined regulation strategies and dynamic optimization mechanisms, the present invention realizes the efficient management of flexible loads in office buildings and the improvement of response capabilities, providing a reliable guarantee for peak shaving and valley filling of the power system. Compared with the prior art, the present invention has the following beneficial effects:

[0164] (1) Enhanced flexibility: By designing classification and refinement strategies for adjustable loads such as air conditioners, lighting, and charging piles, the present invention no longer simply relies on the crude means of directly cutting off loads, thus significantly improving the flexibility and accuracy of the response on the building side.

[0165] (2) Ensuring the normal operation of the building: By preferably appropriately reducing the indoor thermal comfort level from level I to level II, reducing the illuminance standard value by one level, and reasonably arranging the regulation of charging piles, the present invention minimizes the impact on the indoor environment, working lighting, and the use of electric vehicles in the building while ensuring the peak shaving requirements of the power grid, significantly improving the user experience and spatial comfort.

[0166] (3) Rapid response and precise control: The present invention can optionally be connected to the virtual power plant management cloud platform to achieve a second-level response based on price signals and regulation instructions. Using intelligent control systems, Internet of Things lighting technologies, and intelligent charging facilities, the load reduction intensity and duration can be flexibly adjusted to achieve precise control of different types of loads.

[0167] (4) Significant economic and environmental benefits: Through effective flexible load regulation strategies, the present invention can not only reduce energy consumption during peak power periods, relieve the pressure on the power grid, reduce the cost of power purchase and peak electricity price expenditure, but also obtain demand response subsidies and energy-saving benefits in long-term operation, promoting the improvement of building energy utilization efficiency.

[0168] (5) Dynamic optimization and continuous improvement: By recording and feeding back the actual response effect in real time, the present invention can continuously iterate and optimize the strategy according to the adjustment result and economic benefit, ensuring high adjustment performance and economic return in the long-term operation, and improving the energy management level of buildings in the context of smart power grids and virtual power plants.

Claims

1. A flexible load classification method based on multi-dimensional feature analysis, characterized in that: It includes the following steps: Step S101, data acquisition and preprocessing: Collect historical and real-time electricity consumption data of various loads in the building and preprocess it. Step S102, construction of multi-dimensional feature indicators: Extract load characteristics based on the time dimension, space dimension, and response characteristic dimension. The multi-dimensional feature indicators include: time dimension characteristics, space dimension characteristics, regulation ability dimension characteristics, functional necessity characteristics, operation stability characteristics, and environmental dependence characteristics. Step S103, preliminary load classification and adjustable load grading: Among them, the preliminary load classification initially divides the loads into adjustable loads, non-adjustable loads, and special loads. The adjustable load grading establishes a multi-dimensional grading model based on the multi-dimensional feature indicators and grades the preliminarily classified loads through the multi-dimensional grading model to obtain the preliminary classification and grading results. Step S105, output of grading labels and generation of regulation suggestions: It includes generating a statistical report on the proportion of sub-item loads and a time series distribution chart; outputting a list of various graded loads and corresponding regulation suggestions.

2. The flexible load grading method based on multi-dimensional feature analysis according to claim 1, wherein: Between step S103 and step S105, there is also step S104, optimization of classification results: After obtaining the preliminary classification results, optimize the preliminary classification results. After optimizing the classification results, perform deep learning on the historical loads and the optimized classification results through a neural network model to finally obtain the optimized classification labels.

3. The flexible load classification method based on multi-dimensional feature analysis according to claim 1 or 2, characterized in that: It also includes step S106, effectiveness evaluation and optimization: After the response ends on the response day, conduct a comprehensive evaluation based on the actual adjustment amount, adjustment flexibility, economic subsidy income, and internal feedback of the building, and iteratively optimize the grading results according to the evaluation results.

4. The flexible load classification method based on multi-dimensional feature analysis according to claim 1, characterized in that: The preprocessing in step S101 includes removing abnormal data, data interpolation, and baseline load fitting.

5. The flexible load grading method based on multi-dimensional feature analysis according to claim 1, characterized in that: On the basis of step S101, add more historical load data or perform weighted processing on certain special days, which is achieved by the following formula: Among them, Q 基线(t) is the baseline load at time t, and Q i (t) is the load data for the i-th day, and ε is the correction factor.

6. The flexible load classification method based on multi-dimensional feature analysis according to claim 1, wherein: In step S103, the preliminary load classification uses the threshold method or clustering algorithm, where: When the mean and standard deviation of the load fluctuation range satisfy the following relationship, the partition threshold method is adopted: where σx is the standard deviation of the load fluctuation range, and μx is the mean of the load fluctuation range; When the mean and standard deviation of the load fluctuation range satisfy the following relationship, the hierarchical clustering algorithm is adopted:

7. The flexible load grading method based on multi-dimensional feature analysis according to claim 1, characterized in that: Specifically, in step S103, the multi-dimensional hierarchical model is established by constructing a high-dimensional original feature vector X i = [response delay, load operation cycle, load fluctuation frequency…], and after dimensionality reduction processing, the dynamic evolution process of the flexible load is modeled through a discrete-time finite-state Markov chain model to obtain a multi-dimensional hierarchical model.

8. The flexible load grading method based on multi-dimensional feature analysis according to claim 7, characterized in that: The discrete-time finite-state Markov chain model is used to model the dynamic evolution process of flexible loads as follows: Define the state set S = {S1, S2, S3}, corresponding to level-I, level-II, and level-III flexible loads respectively. By annotating the multi-dimensional characteristic indexes of the target load in the historical operation data, its historical state sequence is constructed, and based on the state transition statistical frequency, a state transition probability matrix P = [P ij , where P ij represents the probability that the load transfers from state i to state j; By using the steady-state analysis method of Markov chain, solve the long-term steady-state distribution vector π = [π1, π2, π3] that satisfies the condition πP = π, ∑ i π i According to the principle of the largest component, If π1 = max(π), it is determined as a grade I load; If π2 = max(π), it is determined as a grade II load; If π3 = max(π), it is determined as a grade III load.

9. The flexible load classification method based on multi-dimensional feature analysis according to claim 2, characterized in that: In step S104, the optimization of the preliminary classification results uses fuzzy logic rules to optimize and correct the classification results, which is achieved by the following formula: F final = max(μ adjustable , μ non-adjustable , μ special ) where μ is the membership function; If μ adjustable is the largest, then the load is officially marked as "adjustable"; If μ non-adjustabl is the largest, then the load is officially marked as "non-regulated"; If μ special is the largest, it is classified as a "special load" and does not enter the conventional regulation pool; After optimizing and correcting the classification results, the optimized classification labels are obtained.

10. The flexible load grading method based on multi-dimensional feature analysis according to claim 1, wherein: In step S105, the regulation suggestions include: For adjustable loads, use a multi-dimensional scoring method for quantitative evaluation. When multiple adjustable loads have the possibility of regulation at the same time, preferentially select the one with a higher Score value to execute the regulation; The Score value is calculated as follows: Score = ω1·F 必要性 + ω2·F 灵活性 + ω3·F 响应性 ω1 + ω2 + ω3 = 1 Among them, ω1, ω2, and ω3 are weight coefficients. Optionally, they are set according to experience as: ω1 = 0.4, ω2 = 0.3, ω3 = 0.3.

Images