Commercial air conditioner load dynamic influence factor prediction method and device

Abstract

The invention discloses a commercial air conditioner load dynamic influence factor prediction method and device. The method comprises the following steps: firstly, collecting historical data according to a power load four-level classification system, and constructing a historical impact factor time sequence; then decomposing the sequence into four components including a reference value, a medium and short term value, a short term value and a disturbance value through multi-scale decomposition; on the basis of the reference value sequence and the prediction day features, similar days are screened through grey correlation analysis, and a training data set is constructed; aiming at the characteristics of different time scale components, a plurality of machine learning models are adopted to train prediction sub-models respectively; and finally, fusing prediction results of the components by optimizing weights to generate a dynamic impact factor time sequence. According to the method, through a multi-scale decomposition and model cooperation mechanism, prediction precision and adaptability are effectively improved, and a reliable basis is provided for fine scheduling of a power system.

Description

Technical Field

[0001] This invention relates to the field of power system analysis and load forecasting, specifically to a method and apparatus for predicting dynamic influence factors of commercial air conditioning load. Background Technology

[0002] Electricity loads can be categorized into various types based on their purpose. Among these, commercial air conditioning loads, due to their high power consumption and adjustable potential, have a significant impact during peak summer electricity demand. With the advancement of electricity market reforms, the electricity spot market demands higher time accuracy for load forecasting, requiring precision down to the hour or even 15 minutes to achieve effective market clearing and dispatching decisions. Therefore, accurately quantifying the dynamic impact of commercial air conditioning loads on the total load has become crucial for improving load forecasting accuracy and supporting the refined operation of the system.

[0003] Currently, the main methods for quantifying the impact of a single load type include the static empirical coefficient method, regression analysis, load decomposition method, and machine learning prediction method. The static empirical coefficient method uses a fixed ratio and cannot reflect the dynamic changes within and between days; the regression analysis method focuses more on the prediction of air conditioning load itself and does not fully correlate its dynamic impact on the total load; the load decomposition method relies on high-frequency data and is difficult to apply to industrial and commercial scenarios; while machine learning methods perform well in total load prediction, they do not effectively distinguish the contributions of different load types.

[0004] In addition, existing methods generally suffer from the following shortcomings: First, they struggle to capture the rapid dynamic changes in influencing factors with seasonality, date type, and weather conditions; second, they ignore the coupling relationships between different load types, failing to reflect synergistic or superimposed effects; and third, they rely on a single time scale, lacking the ability to collaboratively model long-term trends, medium-term fluctuations, and short-term stochasticity. These limitations restrict the accurate prediction of dynamic influencing factors of commercial air conditioning loads, making it difficult to meet the actual needs of refined power system management. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for predicting the dynamic influence factors of commercial air conditioning load, which solves the problems in the prior art.

[0006] This invention is achieved through the following technical solution:

[0007] In a first aspect, embodiments of the present invention provide a method for predicting dynamic impact factors of commercial air conditioning load, including:

[0008] Based on the four-level classification system of electricity load, historical electricity load classification data were collected for a historical period. Influence factors were calculated and historical curves were constructed to obtain the historical influence factor time series corresponding to each historical day within the historical period. The four-level classification system of electricity load is as follows: the first level is the total load, the second level is the industrial and commercial electricity consumption and other industry electricity consumption under the total load, the third level is the air conditioning load and other electricity consumption types under the industrial and commercial electricity consumption, and the fourth level is the commercial air conditioning load and other air conditioning load under the air conditioning load. The influence factor is used to characterize the actual degree of influence of the commercial air conditioning load on the total load.

[0009] Based on the time series of all historical influencing factors within the historical period, multi-scale data decomposition is performed to obtain the number of modes and the multiple time scale components corresponding to the number of modes. The multiple time scale components include a baseline value sequence, a medium-short-term value sequence, a short-term value sequence, and a disturbance value sequence. The baseline value sequence represents the long-term seasonal trend, the medium-short-term value sequence represents the weekly cycle characteristics, the short-term value sequence represents intraday and adjacent-day random changes, and the disturbance value sequence represents high-frequency oscillation components.

[0010] Based on the baseline value sequence in the multiple time scale components, the meteorological forecast data and date characteristics of the day to be predicted, similar days are screened from the historical period, and the baseline value sequence corresponding to each similar day is extracted as a label. The time features and meteorological features are used as input feature vectors to obtain a training dataset for supervised learning.

[0011] Based on the training dataset, multi-model collaborative training is performed, in which different models are trained for different time scale components to obtain prediction sub-models for each time scale component.

[0012] Based on the output results of each prediction sub-model for the predicted day, multi-scale prediction results are fused to obtain the time series of dynamic influence factors of commercial air conditioning load for the predicted day.

[0013] Preferably, the step of collecting historical power load classification data according to the four-level power load classification system, calculating influencing factors and constructing historical curves to obtain the historical influencing factor time series corresponding to each historical day within the historical period includes:

[0014] Historical power load data is collected based on a predefined four-level classification system for power load.

[0015] The historical power load data is preprocessed to obtain a cleaned and verified historical load data sequence. The preprocessing includes identifying and removing abnormal data points, using interpolation methods to fill in missing values, and using smoothing algorithms to suppress random mutations in the data.

[0016] Based on the preprocessed historical load data sequence, the coupling correlation strength between loads at all levels is calculated using the curve similarity measurement method to obtain dynamic correction coefficients, wherein the dynamic correction coefficients reflect the nonlinear correlation characteristics between loads at different time periods and under different meteorological conditions.

[0017] Based on the commercial air conditioning load data, total load data, and dynamic correction coefficients in the four-level power load classification system, the original influencing factor sequence is calculated.

[0018] The original impact factor sequence is normalized to map the data range to a preset standard interval, eliminate the influence of units and enhance the comparability of time series, and generate the historical impact factor time series corresponding to each day in the historical period.

[0019] Preferably, the step of calculating the coupling correlation strength between load levels using a curve similarity measurement method based on the preprocessed historical load data sequence to obtain dynamic correction coefficients includes:

[0020] Based on the preprocessed historical load data sequence, normalized daily curves of load at each level are obtained through normalization processing;

[0021] Based on the normalized daily curve, the similarity between load curves at different levels is calculated using the dynamic time warping algorithm, and cluster analysis is performed based on the similarity matrix to obtain the clustering results.

[0022] Based on the clustering results and the characteristics of the day to be predicted, the characteristic distance between the day to be predicted and each historical day is calculated, and a weighted average influence factor is constructed based on the distance weight.

[0023] Based on the degree of meteorological extremes, date type, and historical baseline impact factors, the weighted average impact factors are modified in multiple dimensions to obtain the initial dynamic correction coefficients;

[0024] The initial dynamic correction coefficients are smoothed and constrained to obtain the dynamic correction coefficients.

[0025] Preferably, the step of performing multi-scale data decomposition based on the time series of all historical influencing factors corresponding to the historical period to obtain the number of modalities and the multiple time-scale components corresponding to the number of modalities includes:

[0026] Based on the time series of all historical influencing factors within the historical period, a preliminary decomposition is performed using the variational mode decomposition method to obtain a predetermined number of intrinsic mode functions.

[0027] Based on the frequency characteristics of each intrinsic mode function, a wavelet packet decomposition method is used for secondary decomposition to obtain multiple components with different time scale characteristics.

[0028] Based on the physical meaning and time scale characteristics of each component, each component is assigned to a baseline value sequence, a medium-short term value sequence, a short term value sequence, and a disturbance value sequence.

[0029] Preferably, the step involves filtering similar days from the historical time period based on the baseline value sequence in the multiple time scale components, the meteorological forecast data of the day to be predicted, and the date characteristics, extracting the baseline value sequence corresponding to each similar day as a label, and using the time features and meteorological features as input feature vectors to obtain a training dataset for supervised learning, including:

[0030] A reference sequence is constructed based on the meteorological forecast data and date characteristics of the day to be predicted, and a comparison sequence is constructed based on the corresponding meteorological data and date characteristics of each day in the historical period.

[0031] Based on the reference sequence and comparison sequence, the correlation degree between the day to be predicted and each historical day is calculated using the grey relational analysis method.

[0032] The historical days are sorted according to the correlation degree, and a preset number of historical days with the highest correlation degree are selected as the set of similar days;

[0033] Based on the set of similar days, supervised learning training samples are constructed. The baseline value sequence corresponding to each similar day is extracted as a label, and the time features and meteorological features are used as input feature vectors to obtain the training dataset.

[0034] Preferably, the step of constructing a reference sequence based on the meteorological forecast data and date characteristics of the day to be predicted, and constructing a comparison sequence based on the corresponding meteorological data and date characteristics of each day in the historical period, includes:

[0035] Based on the meteorological forecast data for the day to be predicted, temperature, humidity, and radiation intensity features are extracted and standardized to obtain meteorological feature vectors.

[0036] Based on the date information of the day to be predicted, extract the date type, weekday and month features and perform one-hot encoding to obtain the date feature vector;

[0037] A reference sequence is constructed by combining the meteorological feature vector and the date feature vector.

[0038] Based on the meteorological data and date information corresponding to each day in the historical period, features are extracted and encoded to construct a comparison sequence for each historical day.

[0039] Preferably, the step of calculating the correlation degree between the day to be predicted and each historical day using grey relational analysis based on the reference sequence and the comparison sequence includes:

[0040] Based on the reference sequence and the comparison sequence, calculate the absolute difference of each feature dimension;

[0041] Based on the absolute difference, determine the global minimum difference and the global maximum difference;

[0042] Based on the absolute difference, global minimum difference, and global maximum difference, the correlation coefficient for each feature dimension is calculated using the resolution coefficient, wherein the correlation coefficient satisfies:

[0043] ;

[0044] in, For the first On the 1st historical day Correlation coefficients across each feature dimension For the reference sequence in the 1st Values ​​in each feature dimension For the first The comparison sequence in the th ... Values ​​in each feature dimension The resolution coefficient;

[0045] Based on the correlation coefficients of each feature dimension, the correlation degree between the day to be predicted and each historical day is calculated.

[0046] Preferably, the step of performing multi-model collaborative training based on the training dataset, wherein different time scale components are trained using different models to obtain prediction sub-models for each time scale component, includes:

[0047] Based on the input feature vector and baseline value sequence in the training dataset, the XGBoost ensemble learning model is used for training to obtain the baseline value sequence prediction sub-model.

[0048] Based on historical short- and medium-term value sequence data and corresponding meteorological forecast characteristics and date characteristics, a support vector machine model is used for training to obtain a short- and medium-term value sequence prediction sub-model.

[0049] Based on the time series data of historical short-term value sequences, a long short-term memory neural network model is trained to obtain a short-term value sequence prediction sub-model;

[0050] Based on the high-frequency characteristics of the historical perturbation value sequence, wavelet analysis and statistical methods are used to process the data and obtain a constant numerical representation of the perturbation value sequence.

[0051] Preferably, the step of fusing multi-scale prediction results based on the output of each prediction sub-model for the predicted day to obtain the time series of dynamic influencing factors of commercial air conditioning load for the predicted day includes:

[0052] Based on the baseline value sequence prediction sub-model, the medium-short-term value sequence prediction sub-model, and the short-term value sequence prediction sub-model, the baseline value sequence prediction results, the medium-short-term value sequence prediction results, and the short-term value sequence prediction results for the date to be predicted are obtained respectively.

[0053] Based on the prediction performance of each prediction sub-model on the historical validation set, an optimization algorithm is used to solve for the optimal weight coefficients of each component. The objective function for solving the optimal weight coefficients is to minimize the root mean square error between the ensemble prediction results and the true values ​​on the validation set. The objective function satisfies the following:

[0054] ;

[0055] ;

[0056] ;

[0057] in, For the first The true value at any given moment. For the first The prediction results obtained at each time point , , These are the forecast results for the baseline value, the medium- and short-term value, and the short-term value, respectively. The value is a constant value in the perturbation value sequence;

[0058] Based on the optimal weighting coefficient, the prediction results of each sequence are weighted and fused with the constant value of the disturbance value sequence to obtain the final predicted value of the influence factor at each time of the day to be predicted.

[0059] Based on the final predicted values ​​of the impact factors at all times, a time series of dynamic impact factors of commercial air conditioning load is constructed.

[0060] Secondly, embodiments of the present invention provide a device for predicting the dynamic impact factors of commercial air conditioning load, comprising:

[0061] The curve construction module is used to collect historical power load classification data based on the four-level power load classification system, calculate the impact factor and construct the historical curve to obtain the historical impact factor time series corresponding to each historical day within the historical period. The four-level power load classification system is as follows: the first level is the total load, the second level is the industrial and commercial electricity consumption and other industry electricity consumption under the total load, the third level is the air conditioning load and other electricity consumption types under the industrial and commercial electricity consumption, and the fourth level is the commercial air conditioning load and other air conditioning load under the air conditioning load. The impact factor is used to characterize the actual impact of the commercial air conditioning load on the total load.

[0062] The decomposition module is used to perform multi-scale data decomposition based on the time series of all historical influencing factors corresponding to the historical period, to obtain the number of modes and the multiple time scale components corresponding to the number of modes. The multiple time scale components include a baseline value sequence, a medium-short-term value sequence, a short-term value sequence, and a disturbance value sequence. The baseline value sequence represents the long-term seasonal trend, the medium-short-term value sequence represents the weekly cycle characteristics, the short-term value sequence represents the intraday and adjacent day random changes, and the disturbance value sequence represents the high-frequency oscillation component.

[0063] The filtering module is used to filter similar days from the historical period based on the baseline value sequence in the multiple time scale components, the meteorological forecast data of the day to be predicted, and the date characteristics. It extracts the baseline value sequence corresponding to each similar day as a label, and uses the time characteristics and meteorological characteristics as input feature vectors to obtain a training dataset for supervised learning.

[0064] The training module is used to perform multi-model collaborative training based on the training dataset, wherein different models are trained for different time scale components to obtain prediction sub-models for each time scale component.

[0065] The prediction module is used to perform multi-scale prediction result fusion based on the output results of each prediction sub-model for the day to be predicted, and obtain the time series of dynamic influence factors of commercial air conditioning load for the day to be predicted.

[0066] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0067] This scheme constructs a dynamic impact factor quantification method under a four-level load classification system, combined with multi-scale decomposition and multi-model collaborative prediction technology, to achieve a refined time-varying characterization of the impact of commercial air conditioning load on total load. This effectively solves the problem that traditional static impact factor methods cannot capture intraday dynamic changes. The multi-scale decomposition strategy separates the complex non-stationary impact factor sequence into four components with clear physical meaning: baseline value, medium- and short-term value, short-term value, and disturbance value, enabling targeted modeling of characteristics at different time scales. The baseline value prediction based on similar days and the multi-model collaborative training mechanism fully leverage the advantages of various algorithms in different frequency bands, while the weighted fusion strategy ensures the integrity and stability of the final prediction results. This method can accurately reflect the dynamic changes of commercial air conditioning load impact factors at different times of the day, from the morning trough to the afternoon peak and then to the nighttime decline, providing a high-time-resolution decision-making basis for power system load decomposition, short-term load forecasting, and demand-side management, and supporting grid dispatching departments to achieve more refined power balance and resource optimization. Attached Figure Description

[0068] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0069] Figure 1 A flowchart illustrating the method for predicting dynamic influencing factors of commercial air conditioning load provided by this invention;

[0070] Figure 2 A schematic diagram of the structure of the commercial air conditioning load dynamic influence factor prediction device provided by the present invention. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0072] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0073] It should be noted that all actions involving the acquisition of signals, information, or data in this invention are carried out in compliance with the relevant data protection laws and regulations of the locality and with authorization from the owner of the relevant device.

[0074] Example 1

[0075] Please see Figure 1 This invention provides a method for predicting dynamic impact factors of commercial air conditioning load, including:

[0076] S1. Collect historical power load classification data according to the four-level power load classification system, calculate the impact factor and construct the historical curve to obtain the historical impact factor time series corresponding to each day in the historical period. The first level of the four-level power load classification system is the total load, the second level is the industrial and commercial electricity consumption and other industry electricity consumption under the total load, the third level is the air conditioning load and other electricity consumption types under the industrial and commercial electricity consumption, and the fourth level is the commercial air conditioning load and other air conditioning load under the air conditioning load. The impact factor is used to characterize the actual impact of the commercial air conditioning load on the total load.

[0077] Specifically, historical data was first collected based on a four-level classification system for electricity load. This system uses a hierarchical tree structure to systematically divide load components: the first level is the total load as the root node; the second level is divided by industry into industrial and commercial electricity consumption and other industry branches; the third level, within industrial and commercial electricity consumption, is divided by purpose into air conditioning load and other electricity consumption types; and the fourth level further refines air conditioning load into commercial air conditioning load and other air conditioning load categories. This hierarchical classification accurately identifies the contribution of commercial air conditioning load to the total load. The impact factor calculation uses a dynamic ratio algorithm, combining the ratio of the fourth-level commercial air conditioning load value to the first-level total load value with dynamic correction parameters such as weather and date to generate a quantitative indicator reflecting the actual impact of commercial air conditioning. This calculation process is executed daily, forming a time series of equally spaced samples, thereby constructing a complete historical impact factor curve, providing a standardized data foundation for subsequent multi-scale analysis.

[0078] S2. Based on the time series of all historical influencing factors corresponding to the historical period, perform multi-scale data decomposition to obtain the number of modes and the multiple time scale components corresponding to the number of modes. The multiple time scale components include a baseline value sequence, a medium-short-term value sequence, a short-term value sequence, and a disturbance value sequence. The baseline value sequence represents the long-term seasonal trend, the medium-short-term value sequence represents the weekly cycle characteristics, the short-term value sequence represents the intraday and adjacent day random changes, and the disturbance value sequence represents the high-frequency oscillation component.

[0079] Specifically, a multi-scale decomposition was performed on the historical influencing factor time series, employing a signal decomposition algorithm to separate the non-stationary series into subsequences with different frequency characteristics. The baseline value series was obtained by extracting low-frequency modal components, reflecting slow trends on a monthly or quarterly basis; the medium- and short-term value series captured medium-frequency fluctuation components, exhibiting regular fluctuations on a weekly basis; the short-term value series contained higher-frequency components, reflecting rapid fluctuations within and between adjacent days; and the perturbation value series focused on residual high-frequency oscillations, characterizing random noise. This strategy of decomposing based on physical characteristics transforms the complex series into multiple relatively stationary subsequences, preserving the original series' variation patterns at different time scales while reducing the prediction complexity of each subsequence, thus creating conditions for subsequent differentiated modeling.

[0080] S3. Based on the baseline value sequence in the multiple time scale components, the meteorological forecast data and date characteristics of the day to be predicted, similar days are filtered from the historical period, and the baseline value sequence corresponding to each similar day is extracted as a label. The time features and meteorological features are used as input feature vectors to obtain a training dataset for supervised learning.

[0081] Specifically, a multi-dimensional reference vector is constructed based on meteorological forecast data and date features for the day to be predicted. This vector contains standardized values ​​of meteorological elements such as temperature and humidity, as well as encoded features of the date type. A similarity calculation algorithm is used to select the dates with the closest feature vectors from historical data as a set of similar days. Then, the baseline value sequence data corresponding to these similar days is extracted as label values ​​for supervised learning. Simultaneously, the temporal and meteorological features of the corresponding times are combined into an input feature vector. This process improves the representativeness of the training data and the model's generalization ability by ensuring the similarity of key features between the training samples and the prediction target, providing high-quality supervised learning samples for baseline value sequence prediction.

[0082] S4. Based on the training dataset, perform multi-model collaborative training, wherein different models are used to train different time scale components to obtain prediction sub-models for each time scale component.

[0083] Specifically, for benchmark value sequences reflecting long-term trends, an ensemble learning model is used to handle the complex nonlinear relationships between them and features such as weather and date; for short- to medium-term value sequences exhibiting periodic fluctuations, a kernel method model is used to capture their regular change patterns; for short-term value sequences with high-frequency changes, a time-series neural network model is used to learn their short-term dependencies; and for perturbation value sequences with strong randomness, statistical methods are used to extract their distribution characteristics. This method of customizing models according to component characteristics fully leverages the advantages of each algorithm on different tasks, and improves the adaptability and robustness of the overall prediction framework through division of labor and collaboration.

[0084] S5. Based on the output results of the predictions of each prediction sub-model for the day to be predicted, perform multi-scale prediction result fusion to obtain the time series of the dynamic influence factors of commercial air conditioning load for the day to be predicted.

[0085] Specifically, the output results of each prediction sub-model for the predicted date are first obtained, including the predicted sequences of the baseline value, short- and medium-term values, and short-term values, as well as the statistical characteristics of the perturbation values. Then, based on the prediction performance of each component on the historical validation set, an optimization algorithm is used to solve for the weight allocation scheme that minimizes the overall error. Finally, a weighted fusion algorithm is used to integrate the prediction results of each component into a unified impact factor time series. This fusion process considers both the differences in contribution of each component at different time scales and reduces the prediction bias of a single model by optimizing the weights, thereby generating a more complete and reliable impact factor prediction curve.

[0086] In some implementations, S1 involves collecting historical power load classification data based on a four-level power load classification system, calculating influencing factors and constructing historical curves to obtain the historical influencing factor time series for each day within the historical period, including:

[0087] S11. Collect historical power load data according to a predefined four-level power load classification system;

[0088] Specifically, historical power load data is systematically collected based on a pre-constructed four-level classification system for power load. This classification system adopts a tree-like hierarchical structure from total load to specific load types, subdividing it step by step down to commercial air conditioning load, providing a structured data foundation for subsequent influencing factor calculations.

[0089] S12. The data quality of the historical power load data is preprocessed to obtain a cleaned and verified historical load data sequence. The preprocessing includes identifying and removing abnormal data points, using interpolation methods to fill in missing values, and using smoothing algorithms to suppress random mutations in the data.

[0090] Specifically, the raw data collected is processed to improve its quality. An outlier detection algorithm is used to identify and remove load data points that deviate significantly from the normal range. An interpolation algorithm is used to fill in the data gaps caused by missing data collection. A smoothing filter algorithm is applied to suppress measurement noise and random fluctuations, ultimately forming a complete and reliable historical load data sequence.

[0091] S13. Based on the preprocessed historical load data sequence, the coupling correlation strength between loads at all levels is calculated using the curve similarity measurement method to obtain the dynamic correction coefficient, wherein the dynamic correction coefficient reflects the nonlinear correlation characteristics between loads at different time periods and under different meteorological conditions.

[0092] Specifically, based on the preprocessed load data, a curve similarity measurement algorithm is used to quantify the degree of morphological correlation between load curves at different levels. By analyzing the matching patterns and consistency of changes of curves in time series, a correction coefficient reflecting the dynamic coupling relationship between loads is constructed. This coefficient can capture the nonlinear characteristics of load correlation as weather conditions and time periods change.

[0093] S14. Based on the commercial air conditioning load data, total load data and dynamic correction coefficient in the four-level power load classification system, the original influencing factor sequence is calculated.

[0094] Specifically, commercial air conditioning load data and total load data from the four-level classification system are used in combination with dynamic correction coefficients for composite calculation. By multiplying the ratio of commercial air conditioning load to total load by the correction coefficient, the original influencing factor sequence is obtained, thus initially quantifying the dynamic proportion of commercial air conditioning load in the total load.

[0095] S15. Normalize the original impact factor sequence to map the data range to a preset standard interval, eliminate the influence of units and enhance the time series comparability, and generate the historical impact factor time series corresponding to each day in the historical period.

[0096] Specifically, the original impact factor sequence was normalized by mapping the data to a standardized interval of zero to one using a linear transformation algorithm. This process eliminated the absolute value differences caused by seasonal variations, preserved the relative variation characteristics of the load impact, and ultimately generated a standardized historical impact factor time series.

[0097] In some implementations, S13, based on the preprocessed historical load data sequence, a curve similarity measurement method is used to calculate the coupling correlation strength between loads at different levels, and a dynamic correction coefficient is obtained, including:

[0098] S131. Based on the preprocessed historical load data sequence, normalized daily curves of load at each level are obtained through normalization processing;

[0099] Specifically, normalization refers to the mathematical transformation process of mapping the original values ​​of loads at all levels to a unified dimensional range. By performing min-maximum or Z-score normalization on historical data of four levels—total load, industrial and commercial electricity load, air conditioning load, and commercial air conditioning load—the dimensional differences and amplitude imbalances between different load levels are eliminated. This allows the morphological characteristics of load curves at all levels to be compared and analyzed on the same scale. The normalized daily curve represents the temporal distribution characteristics of loads at all levels over 24 hours after removing the influence of total scale, providing standardized basic data for subsequent calculations of the morphological similarity between different load levels. This process effectively preserves the temporal morphological characteristics of the load curves while suppressing the interference of absolute numerical differences on similarity calculations, making the analysis of coupling relationships between different load levels more objective and reliable.

[0100] S132. Based on the normalized daily curve, the similarity between load curves at different levels is calculated using the dynamic time warping algorithm, and cluster analysis is performed based on the similarity matrix to obtain the clustering results.

[0101] Specifically, the dynamic time warping algorithm is a sequence similarity measurement method that can handle time axis scaling deformation. By constructing a nonlinear alignment path to minimize the cumulative distance between two sequences, it overcomes the limitation of traditional Euclidean distance being sensitive to time offset. It calculates the time warping distance between the normalized load curves of adjacent levels, forming a similarity matrix that characterizes the degree of similarity in load morphology at each level. This matrix reflects the morphological correlation strength between load curves on different dates. Subsequently, spectral clustering or hierarchical clustering methods are used to reduce the dimensionality of the similarity matrix and group typical day sets with similar load coupling patterns. The clustering process can automatically discover the load correlation patterns hidden in historical data and classify dates with similar morphological characteristics into the same category. This provides data support for constructing differentiated correction coefficients for different cluster categories, enabling the dynamic correction coefficients to adaptively reflect the load correlation characteristics under different operating modes.

[0102] S133. Based on the clustering results and the characteristics of the day to be predicted, calculate the characteristic distance between the day to be predicted and each historical day, and construct a weighted average influence factor based on the distance weight.

[0103] Specifically, the feature distance calculation comprehensively considers date type attributes and meteorological condition parameters. It measures the comprehensive difference between the predicted day and historical days using a weighted Euclidean distance formula. The weight allocation is based on the sensitivity analysis of meteorological elements to the load, with temperature parameters given higher weights, followed by humidity and radiation parameters. At the same time, a cluster category matching factor is introduced to adjust the distance calculation results. Historical days belonging to the same cluster category as the predicted day are assigned a distance decay coefficient to enhance their similarity score. Based on the calculated feature distance, an exponential decay function is used to convert the distance into a weight coefficient. Historical days with smaller distances receive higher weights. The weighted average influence factor is a weighted average of the original influence factor values ​​of historical days according to the distance weight. This process allows historical days with similar meteorological conditions and date types to the predicted day to contribute more to the calculation of the correction coefficient, effectively improving the matching degree between the correction coefficient and the actual working conditions of the predicted day, and avoiding the pattern confusion problem caused by simple arithmetic averaging.

[0104] S134. Based on the degree of meteorological extremes, date type, and historical baseline influence factors, the weighted average influence factors are modified in multiple dimensions to obtain the initial dynamic correction coefficients.

[0105] Specifically, the degree of meteorological extremes is quantified by the degree of deviation of standardized temperature and humidity from historical averages. A meteorological anomaly index is constructed as a correction factor. Date type correction is based on the statistical differences of historical impact factors for weekdays, weekends, and holidays, with adjustment coefficients set. The historical benchmark impact factor reflects the average level of impact factors under long-term trends. Through multiple linear combination or nonlinear function mapping, the correction factors of the above three dimensions are integrated with the weighted average impact factor to form an initial dynamic correction coefficient. This multi-dimensional correction process enables the correction coefficient to respond simultaneously to meteorological anomalies, special date operating patterns, and long-term trend changes, enhancing the adaptability of the correction coefficient to different operating environments. It effectively compensates for extreme conditions and special date effects that cannot be captured by simply using a weighted average of similar days, ensuring that the correction coefficient maintains a reasonable range of values ​​under different operating conditions.

[0106] S135. The initial dynamic correction coefficients are smoothed and constrained to obtain dynamic correction coefficients.

[0107] Specifically, the smoothing process employs moving average or low-pass filtering methods to continuously optimize the initial dynamic correction coefficients over time, eliminating abrupt jumps between adjacent time periods and ensuring a smooth transition across the 96 time points within a day. The constraint processing sets upper and lower thresholds for the correction coefficients, determining a reasonable range based on historical statistical patterns. When the smoothed correction coefficients exceed the threshold range, a nonlinear compression function maps them back to the effective range, while retaining reasonable deviations under extreme conditions. This process ensures the continuity of the dynamic correction coefficients over time and the reasonableness of their numerical range, avoiding interference from abnormal fluctuations in the correction coefficients on the final impact factor calculation. This allows the correction coefficients to reflect changes in the actual load coupling relationship while remaining within an acceptable stable range for engineering applications, providing high-quality input data for subsequent multi-scale decomposition.

[0108] For example:

[0109] Curve similarity measurement and cluster analysis: The Dynamic Time Warping (DTW) algorithm is used to calculate the similarity between the commercial air conditioning load curve and the upper-level load curve, and to identify their association patterns.

[0110] Step 1: Calculate DTW distance

[0111] The calculation of DTW distance follows a recursive relationship:

[0112] ;

[0113] in, For sequence The former Points and sequences The former The minimum cumulative alignment distance between points This is the point-to-point Euclidean distance.

[0114] Step 2: Similarity Conversion

[0115] Convert DTW distance into similarity (range 0-1):

[0116] in, Let be the attenuation coefficient, and take . =0.5. When When =0, Sim=1 (completely similar). The larger the value, the closer Sim is to 0.

[0117] Step 3: Cluster Analysis

[0118] K-means clustering (K=5) was performed on the DTW distance matrix for all historical 365 days, resulting in five clusters representing different load association patterns: Cluster C1: High-temperature weekdays (temperature >32℃, Monday to Friday), highest air conditioning contribution; Cluster C2: Normal-temperature weekdays (temperature 15-25℃, Monday to Friday), medium air conditioning contribution; Cluster C3: Low-temperature weekdays (temperature <10℃, Monday to Friday), low air conditioning contribution; Cluster C4: Weekends and quasi-holidays (Saturday and Sunday), varying pedestrian traffic, different air conditioning patterns; Cluster C5: Public holidays (statutory holidays), special load patterns; for forecast days... To determine if it belongs to a cluster .

[0119] (3) Layer-by-layer influence calculation model

[0120] Based on the clustering results, a multi-level influencing factor calculation model of commercial air conditioning load on total load is constructed.

[0121] First layer: Factors affecting the total air conditioning load of commercial air conditioning: ;

[0122] in, For a moment Commercial air conditioning load value, For a moment air conditioning load value, For a moment The dynamic correction coefficient.

[0123] Second layer: Factors affecting air conditioning load on industrial and commercial electricity consumption: ;

[0124] in, For a moment The value of industrial and commercial electricity consumption.

[0125] Third layer: Factors affecting the total load of industrial and commercial electricity consumption: ;

[0126] in, For a moment Total load value.

[0127] The final dynamic impact factor of commercial air conditioning on the total load is:

[0128] .

[0129] (4) Dynamic correction coefficient Construction method: The dynamic correction coefficient is used to adjust the static nature of the simple ratio so that it can reflect the actual impact of different periods, different meteorological conditions and different working modes.

[0130] Step 1: Constructing the baseline impact factor

[0131] Calculate the average impact factor for each month over a historical 365-day period as the seasonal baseline:

[0132] ;

[0133] in, For months The average value of the impact factor For months The number of days included.

[0134] Step 2: Calculate the similarity between the predicted date and historical dates within the cluster.

[0135] For the predicted date First, determine if it belongs to a cluster. Calculate the characteristic distance between the predicted date and each historical date within the cluster:

[0136] ;

[0137] in, For the predicted date, For the historical date within the cluster, The average daily temperature The average daily humidity The total daily radiation intensity is represented by z-score normalized values ​​(mean 0, standard deviation 1).

[0138] Similarity weight: ;

[0139] in, =0.8 is the attenuation coefficient, which is determined in the following way:

[0140] In the 3D normalized feature space, the maximum value of the historical distance distribution is approximately 3.5 (the 99th quantile extreme case: dist≈3.5). To ensure that the weights decay reasonably with distance, the following rules are set: when dist=0 (completely similar), the weight w=1.0; when dist=1.0 (moderately similar), the weight w≈0.45; when dist=2.0 (moderately dissimilar), the weight w≈0.20; and when dist=3.0 (dissimilar), the weight w≈0.05. The solution is obtained by reverse engineering. When w(2.0) = 0.20, exp(- ·2.0)=0.20; =-ln(0.20) / 2.0=1.609 / 2.0=0.8045≈0.8.

[0141] Step 3: Calculate the intra-cluster weighted average impact factor

[0142] Calculate the weighted average of all historical days within a cluster: ;

[0143] ;

[0144] This represents the average impact factor of commercial air conditioning on the historical day most similar to the predicted date.

[0145] Step 5: Correction for extreme weather conditions: ;

[0146] For sustained high temperatures (3 consecutive days) (>32℃), multiply by a cooling factor of 0.92 (because the air conditioner is already running at maximum power).

[0147] Rainfall correction: When rainfall is ≥5mm, multiply by a factor of 0.88 (cooling, air conditioning load decrease).

[0148] Step 6: Date Special Circumstances Correction: ;

[0149] Step 7: Comprehensive dynamic correction coefficient: ;

[0150] Step 8: Smoothing and Constraints

[0151] To avoid sudden jumps in the correction factor, a 3-day moving average is used: ;

[0152] And apply upper and lower bound constraints: ;

[0153] Ensure that the correction factor is within a reasonable range of [0.50, 1.50].

[0154] Step 9: Final Influence Factor Calculation. Incorporating correction factors, the dynamic influence factor for commercial air conditioning is: .

[0155] 2. Data preprocessing and historical curve generation

[0156] In this embodiment, historical data from January 2024 to December 2024 were collected from a city in Southwest China, including: total load data: 96 points / day (15-minute intervals); industry-specific load data: also 96 points / day; meteorological data: temperature, humidity, radiation intensity, etc.

[0157] Data preprocessing steps:

[0158] (1) Error data removal: Identify and delete obviously erroneous data points, such as negative values, values ​​that exceed physical limits, etc.

[0159] (2) Missing value completion: For missing data, Lagrange interpolation is used for completion.

[0160] ;

[0161] in, For a moment The load estimate is obtained through interpolation. The index of the known data point currently being processed. This serves as an auxiliary index used when constructing basis functions to traverse other known data points.

[0162] (3) Smoothing of abnormal fluctuations: For abnormal fluctuations caused by special events (such as major holidays, extreme weather, etc.), a moving average filter is used for smoothing.

[0163] (4) Generation of impact factor curve: Using the algorithm defined in S101, the impact factor curve of commercial air conditioning for historical days (365 days in total) was calculated: 0-6 am: impact factor is about 3%-5%, commercial buildings are closed and air conditioning load is very low; 7-9 am: impact factor rises rapidly to 10%-15%, commercial buildings are turned on and air conditioning is pre-cooled; 12-3 pm: impact factor reaches peak value of 25%-32%, air conditioning is running at full load during high temperature period; 5-7 pm: impact factor drops slightly to 18%-22%, some commercial buildings are closed; 8-12 am: impact factor drops to 8%-12%, only air conditioning in large shopping malls and other places continues to operate.

[0164] (5) Normalization: Normalize the impact factor curves to eliminate seasonal absolute value differences and retain relative change characteristics.

[0165] In some implementations, S2 involves performing multi-scale data decomposition based on the time series of all historical influencing factors corresponding to the historical period, to obtain the number of modalities and multiple time-scale components corresponding to the number of modalities, including:

[0166] S21. Based on the time series of all historical influencing factors within the historical period, a preliminary decomposition is performed using the variational mode decomposition method to obtain a predetermined number of intrinsic mode functions.

[0167] Specifically, variational mode decomposition (VMD) is an adaptive signal processing technique that decomposes the original signal into a predetermined number K intrinsic mode functions (EMFs) with different center frequencies by constructing a variational optimization problem. This method solves the constrained optimization problem using an alternating direction multiplier algorithm, ensuring that each EMF is compact in the frequency domain and that the center frequencies of each mode are separated. The decomposition process ensures that the decomposition results retain the characteristics of the original signal while avoiding mode aliasing by setting appropriate penalty parameters and convergence thresholds. The predetermined number K is determined based on the signal's spectral characteristics through empirical analysis or optimization criteria. The decomposed EMFs are arranged from low to high frequency, representing different frequency band components in the original influence factor sequence, from long-term trends to high-frequency disturbances. This decomposition method can effectively separate different frequency components in the signal, providing a foundation for subsequent multi-scale analysis. It overcomes the limitation of traditional Fourier transform's insufficient adaptability to non-stationary signals, decomposing complex influence factor sequences into multiple sub-components with distinct spectral characteristics.

[0168] S22. Based on the frequency characteristics of each intrinsic mode function, a wavelet packet decomposition method is used for secondary decomposition to obtain multiple components with different time scale characteristics.

[0169] Specifically, the wavelet packet decomposition method is an extension of the traditional wavelet transform, enabling fine decomposition in both high and low frequency regions. By selecting appropriate mother wavelet basis functions and decomposition levels, the specific intrinsic mode functions obtained from VMD decomposition are further decomposed. This method recursively applies low-pass and high-pass filters to decompose each mode component into finer-grained sub-bands, forming a complete binary tree structure decomposition result. Frequency feature analysis includes calculating the power spectral density, center frequency, and bandwidth parameters of each intrinsic mode function to determine which modes require further decomposition and the optimal decomposition depth. The secondary decomposition process mainly targets mid-to-high frequency intrinsic mode functions, while low-frequency modes are usually kept unchanged. The multiple sub-components obtained from the decomposition have different resolution characteristics in the time-frequency plane, enabling more accurate capture of multi-scale features in the influencing factor sequence, including seasonal variations, weekly cycles, intraday fluctuations, and random disturbances at different time scales. This two-level decomposition strategy realizes multi-level time-frequency analysis of complex non-stationary signals, providing a richer feature basis for subsequent component reconstruction.

[0170] S23. Based on the physical meaning and time scale characteristics of each component, each component is assigned to a baseline value sequence, a medium-short term value sequence, a short term value sequence, and a disturbance value sequence.

[0171] Specifically, the physical meaning refers to the actual load change characteristics reflected by each component in the operation of the power system. The baseline value sequence characterizes the long-term seasonal trend of the commercial air conditioning load influencing factors, mainly affected by seasonal climate changes and long-term commercial activity planning, with slow changes and high predictability; the medium- and short-term value sequences characterize medium-term fluctuation characteristics, including weekly cycles, the transition between working days and non-working days, and load adjustments caused by weather system changes, with cycles typically ranging from several days to several weeks; the short-term value sequence characterizes intraday load change patterns and random fluctuations between adjacent days, reflecting the time-specific characteristics of commercial activities and the impact of short-term weather changes; the disturbance value sequence characterizes unpredictable high-frequency noise and transient disturbances. The dynamics include measurement errors, sudden events, and uncontrollable random factors. The time-scale characteristics are quantitatively evaluated by analyzing the autocorrelation function, spectral distribution, and periodicity index of each component. The allocation process adopts a combination of frequency band mapping and statistical verification. Multiple sub-components obtained from VMD and wavelet packet decomposition are sequentially assigned to four predefined categories according to frequency from low to high. At the same time, cross-validation is used to ensure the statistical consistency within each sequence. This component allocation method gives each reconstructed sequence clear physical interpretation and predictive characteristics, laying the foundation for subsequent use of differentiated prediction models for different time-scale characteristics, while maintaining the integrity and reconstructability of the original influence factor sequence.

[0172] For example, a two-level decomposition strategy combining variational mode decomposition (VMD) and wavelet packet decomposition (WPD) is adopted.

[0173] First layer: Variational Mode Decomposition (VMD)

[0174] The VMD method decomposes the original impact factor sequence f(t) into K intrinsic mode functions (IMFs):

[0175] ;

[0176] Constraints: ;

[0177] in, For the k-th modal component, For its center frequency, K=6 is set in this embodiment, and 6 IMF components are obtained by solving the alternating direction multiplier method (ADMM).

[0178] Second layer: Wavelet packet decomposition (WPD)

[0179] The six IMF components obtained from VMD were analyzed to identify their frequency characteristics and classified according to their physical meaning: IMF1-IMF2 (low-frequency components): reflect long-term trends and are classified into the baseline value sequence D1; IMF3-IMF4 (mid-frequency components): reflect medium-term fluctuations and are classified into the short-to-medium-term value sequence D2; IMF5 (higher-frequency components): reflect short-term randomness and are classified into the short-term value sequence D3; IMF6 (high-frequency components): serve as the disturbance value sequence D4.

[0180] To further refine the decomposition, wavelet packet decomposition was applied to IMF3-IMF5, using the db4 wavelet basis and a decomposition level of 3. The recursive formula for wavelet packet decomposition is as follows:

[0181] ;

[0182] in, and These are the coefficients of the low-pass and high-pass filters, respectively; j is the number of decomposition layers; and n is the node number.

[0183] After the two-level decomposition, the four subsequences obtained are: baseline value sequence D1: the amplitude changes steadily, the periodicity is obvious, and it reflects the seasonal trend; medium and short-term value sequence D2: there is a periodic fluctuation of about 7 days, reflecting the weekly cycle characteristics; short-term value sequence D3: the fluctuation is frequent, reflecting the random changes within the day and between adjacent days; perturbation value sequence D4: high-frequency oscillation, small amplitude, close to zero mean.

[0184] In some implementations, S3 involves filtering similar days from the historical time period based on the baseline value sequence in the multiple time scale components, the meteorological forecast data for the day to be predicted, and the date characteristics. The baseline value sequence corresponding to each similar day is extracted as a label, and the time characteristics and meteorological characteristics are used as input feature vectors to obtain a training dataset for supervised learning, including:

[0185] S31. Construct a reference sequence based on the meteorological forecast data and date characteristics of the day to be predicted, and construct a comparison sequence based on the corresponding meteorological data and date characteristics of each day in the historical period;

[0186] Specifically, the reference sequence is a multi-dimensional feature vector composed of forecast values ​​of various meteorological elements and date attributes for the day to be predicted, including meteorological parameters such as temperature, humidity, and solar radiation intensity, as well as calendar features such as week type, month, and whether it is a holiday. The comparison sequence is a feature set constructed from the same type of measured meteorological data and date attributes corresponding to each day in the historical period. All features are standardized before input to eliminate dimensional differences. The meteorological data uses the daily average values ​​provided by regional meteorological stations or numerical weather prediction models. The date features are converted into binary representation through one-hot encoding. This construction process converts the raw data into a feature sequence in a unified format, so that the day to be predicted and historical days can be similarly measured in the same feature space, providing structured input data for subsequent grey relational analysis. This sequence construction method can comprehensively capture the key external factors affecting air conditioning load, while retaining time-related periodic features, making the similarity day selection process more consistent with the actual load change pattern.

[0187] S32. Based on the reference sequence and comparison sequence, the grey relational analysis method is used to calculate the correlation degree between the day to be predicted and each historical day;

[0188] Specifically, grey relational analysis is a system analysis technique for handling systems where some information is known and some is unknown. It determines the correlation degree by calculating the geometric similarity between a reference sequence and a comparison sequence across various feature dimensions. The calculation process first performs dimensionless processing on the standardized feature sequences, then calculates the absolute difference between each historical day and the day to be predicted across each feature dimension, thereby obtaining the correlation coefficient. Finally, a weighted average is used to obtain the comprehensive correlation degree. The weight allocation is determined based on the sensitivity of each meteorological element to the air conditioning load. Temperature features are usually given higher weights, followed by humidity and radiation features. Date type features are assigned appropriate weights based on historical statistical patterns. This method can effectively handle similarity measurement problems in small-sample, uncertain systems, and overcomes the shortcomings of traditional distance measurement methods that are sensitive to extreme values. It enables the correlation degree calculation results to more robustly reflect the similarity of load patterns during different days, and is particularly suitable for identifying similar days under complex and variable meteorological conditions.

[0189] S33. Sort the historical days according to the correlation degree, and select the historical days with the highest correlation degree as a set of similar days;

[0190] Specifically, the sorting process uses a quicksort algorithm to sort all historical days in descending order of correlation value, forming an ordered candidate list of historical days. The preset number is determined based on a balance between training sample requirements and computational complexity, typically within the range of 20 to 50 days. The selection process not only considers the correlation threshold but also introduces constraints on the diversity of date distribution to ensure that the selected similar days cover different working modes (such as weekdays, weekends, and holidays) and different meteorological conditions, avoiding excessive concentration of samples in a specific type. At the same time, a time decay factor is set to appropriately increase the priority of recent historical days to reflect the slow evolution trend of load characteristics. This screening strategy maintains the diversity and representativeness of training data while ensuring that the samples are highly similar to the predicted days, enabling the subsequent training prediction model to adapt to different working conditions, avoiding overfitting to a single mode, and improving the model's generalization ability in practical applications. The resulting set of similar days serves as the basis for high-quality training samples and directly affects the performance and reliability of subsequent prediction models.

[0191] S34. Based on the set of similar days, construct supervised learning training samples, extract the baseline value sequence of each similar day as a label, and use time features and meteorological features as input feature vectors to obtain the training dataset.

[0192] Specifically, the construction process first extracts the baseline value sequence data corresponding to 96 15-minute interval points for each similar day as the target label for supervised learning. At the same time, it constructs a multi-dimensional input feature vector, including time features (such as hour, week type, and month unique thermal encoding), standardized meteorological features (such as the lag term and rate of change of temperature, humidity, and radiation), and statistical features of historical baseline value sequences (such as the value of the same period of the previous day and the moving average). The feature engineering process uses standardization and normalization to eliminate the influence of dimensions, and cyclically encodes periodic features to maintain temporal continuity. Finally, a structured input-output pair is formed. Each training sample contains a feature vector of a time point and its corresponding baseline value label. All samples are combined to form a complete training dataset. This construction method fully integrates multi-source information from the time domain, frequency domain, and meteorological domain, enabling the training data to comprehensively represent the changing patterns of the baseline value sequence. It provides information-rich and structurally reasonable training samples for subsequent machine learning models, preserving the temporal dependence of the original data while enhancing the ability to express complex nonlinear relationships through multi-dimensional features.

[0193] In some implementations, S31 involves constructing a reference sequence based on meteorological forecast data and date characteristics for the day to be predicted, and constructing a comparison sequence based on meteorological data and date characteristics for each day in historical time periods, including:

[0194] S311. Based on the meteorological forecast data for the day to be predicted, extract the temperature, humidity, and radiation intensity features and perform standardization processing to obtain the meteorological feature vector;

[0195] Specifically, the meteorological forecast data comes from the 24-hour forecast results output by the regional meteorological forecast center or numerical weather prediction model. Temperature features include daily average temperature, maximum temperature, and minimum temperature; humidity features use relative humidity percentage; and radiation intensity features represent solar radiation energy flux density. Standardization processing employs feature-specific transformation strategies. Temperature features are transformed into a distribution with a mean of 0 and a standard deviation of 1 through Z-score standardization. Humidity and radiation intensity features are mapped to the 0-1 interval using Min-Max normalization. This processing eliminates the dimensional differences and numerical range imbalances between different meteorological elements, making each feature comparable in similarity calculations while preserving the distribution characteristics of the original data. The standardized meteorological feature vectors form a multi-dimensional numerical representation, providing a consistent meteorological dimension input for subsequent grey relational analysis. This standardization method can adapt to the dynamic range of meteorological parameters in different seasons and maintain the stability of the feature vectors throughout the year.

[0196] S312. Based on the date information of the day to be predicted, extract the date type, weekday and month features and perform one-hot encoding to obtain the date feature vector;

[0197] Specifically, the date information comes from the Gregorian calendar system, and the date types are divided into three categories: weekdays, weekends, and statutory holidays. The weekday feature represents the seven states from Monday to Sunday, and the month feature represents the twelve months of the year. One-hot encoding converts each categorical feature into a binary vector, where the activation bit is 1 and the rest are 0. For example, Monday is encoded as [1,0,0,0,0,0,0], and January is encoded as [1,0,0,0,0,0,0,0,0,0,0]. The date type is also represented by a three-dimensional weight vector. This encoding method converts categorical features into a numerical form that can be processed by machine learning algorithms, while avoiding misleading relationships caused by numerical magnitude. The date feature vector after one-hot encoding maintains the discreteness and category independence of date attributes, introduces a time periodicity pattern for similar day screening, and enables the model to identify the potential correlation between historical days of the same weekday type or the same month and the day to be predicted. This encoding structure effectively captures the periodic variation pattern of commercial air conditioning load.

[0198] S313. Construct a reference sequence by combining the meteorological feature vector and the date feature vector;

[0199] Specifically, the reference sequence is a multi-dimensional vector representing the comprehensive characteristics of the day to be predicted. The combination process adopts a feature splicing strategy, which connects the standardized meteorological feature vector with the date feature vector after unique thermal encoding in a predefined order to form a single feature vector. Before splicing, the dimensionality consistency of the two types of features is checked to ensure the integrity of the vector structure. The combined reference sequence contains all the key external factors affecting air conditioning load, forming a complete feature representation of the predicted day. This sequence serves as a benchmark for grey relational analysis and is used to compare similarity with historical day features. The construction of the reference sequence follows the principle of feature completeness, including both continuous meteorological variables and discrete calendar features, making the feature description of the day to be predicted comprehensive and structured. This combination method maintains the data characteristics of different feature types, provides a unified input format for subsequent similarity calculations, and preserves the original semantic information of each feature, making the correlation analysis results have clear physical meaning and interpretability.

[0200] S314. Based on the meteorological data and date information corresponding to each day in the historical period, perform feature extraction and encoding to construct a comparison sequence for each historical day.

[0201] Specifically, historical meteorological data for each period is sourced from the historical record database of meteorological monitoring stations, and date information is extracted from historical calendars. The feature extraction process employs a completely consistent processing procedure for each historical day as for the day to be predicted, including the selection of the same meteorological elements, the same standardized parameters, and the same unique thermal coding rules. This ensures that the comparison sequence and the reference sequence are comparable in the feature space. During the historical data processing, the temporal integrity of the original meteorological observations is preserved, missing data is filled using time series interpolation, and abnormal data is identified and corrected using statistical testing methods. Each historical day ultimately forms a feature vector that is completely consistent with the dimension and structure of the reference sequence. The set of feature vectors for all historical days constitutes a comparison sequence library. This library reflects the correlation between historical load patterns and meteorological and date features, providing a comprehensive comparison benchmark for grey relational analysis. This consistent processing ensures the fairness and objectivity of the selection of similar days, avoids systematic bias caused by differences in feature processing, and establishes the similarity assessment between historical days and predicted days on the same feature measurement standard.

[0202] In some implementations, S32, based on the reference sequence and comparison sequence, uses grey relational analysis to calculate the correlation degree between the day to be predicted and each historical day, including:

[0203] S321. Calculate the absolute difference of each feature dimension based on the reference sequence and the comparison sequence;

[0204] Specifically, the absolute difference of each feature dimension is calculated based on the reference sequence and the comparison sequence. The reference sequence consists of meteorological forecast data and date features for the day to be predicted, while the comparison sequence consists of corresponding feature data for each historical day. By calculating the absolute value of the difference between the two values ​​dimension by dimension, an absolute difference sequence reflecting the local differences of each feature is obtained.

[0205] S322. Determine the global minimum difference and the global maximum difference based on the absolute difference;

[0206] Specifically, the global minimum difference and global maximum difference are determined based on the absolute differences across all feature dimensions. The global minimum difference is the smallest absolute difference across all feature dimensions for all historical days, representing the degree of difference between the most similar features; the global maximum difference is the largest absolute difference across all dimensions, representing the range of difference between the least similar features. These two extreme values ​​provide a benchmark scale for subsequent correlation coefficient calculations.

[0207] S323. Based on the absolute difference, the global minimum difference, and the global maximum difference, calculate the correlation coefficient for each feature dimension using the resolution coefficient, wherein the correlation coefficient satisfies:

[0208] ;

[0209] in, For the first On the 1st historical day Correlation coefficients across each feature dimension For the reference sequence in the 1st Values ​​in each feature dimension For the first The comparison sequence in the th ... Values ​​in each feature dimension The resolution coefficient;

[0210] Specifically, the correlation coefficient for each feature dimension is calculated using the resolution coefficient based on the absolute difference, the global minimum difference, and the global maximum difference. In the calculation formula, the numerator is formed by adding the global minimum difference to the product of the resolution coefficient and the global maximum difference to form a benchmark term; the denominator is formed by adding the absolute difference of the current dimension to the same product term. The final correlation coefficient reflects the degree of local similarity on each feature dimension, and the resolution coefficient is used to adjust the sensitivity of the correlation coefficient to the range of differences.

[0211] S324. Calculate the correlation degree between the day to be predicted and each historical day based on the correlation coefficient of each feature dimension.

[0212] Specifically, based on the correlation coefficients of each feature dimension, the comprehensive correlation degree between the day to be predicted and each historical day is calculated. By taking the arithmetic mean of the correlation coefficients of all feature dimensions for the same historical day, a comprehensive correlation degree index reflecting the overall similarity is obtained. This index comprehensively considers the matching degree of multiple features such as weather and date, providing a quantitative basis for subsequent screening of similar days.

[0213] For example, multi-model collaborative prediction is obtained in the following way:

[0214] 1. Prediction of the baseline value sequence D1

[0215] The baseline value series reflects long-term trends, and its changes are relatively slow and highly regular. This embodiment uses the following process for prediction:

[0216] (1) Similar day screening and training sample preparation: Grey relational analysis was used to screen historical day data similar to the predicted day as training samples for the XGBoost model. Let the predicted day be day t. The correlation between each historical day and the predicted day was calculated on the following features. The reference sequence y0 (feature vector) includes: date type: weekday / weekend / holiday (one-hot encoding, 3-dimensional);

[0217] Seasonal features: month, temperature range (unique thermal encoding, 12 dimensions); temperature range: (standardized range [-2, +2], 1 dimension); mean of the previous day's influencing factors (normalized range [0, 1], 1 dimension); total: 17-dimensional feature vectors.

[0218] Comparison Sequences : The corresponding feature vectors of each historical day i, totaling 365 dates;

[0219] Grey relational degree calculation formula: ;

[0220] in, As a reference sequence (predicted daily characteristics), To compare sequences (historical daily characteristics), The resolution coefficient is usually set to 0.5.

[0221] correlation The calculation is as follows: ;

[0222] The top 30 historical days with the highest correlation were selected as training samples.

[0223] Where n is the feature dimension (17 dimensions).

[0224] Training sample selection: The top 30 historical days with the highest correlation were selected as the training samples for the XGBoost model.

[0225] Input and label descriptions for training samples:

[0226]

[0227] In some implementations, S4 involves performing multi-model collaborative training based on the training dataset, wherein different time scale components are trained using different models to obtain prediction sub-models for each time scale component, including:

[0228] S41. Based on the input feature vector and baseline value sequence in the training dataset, the XGBoost ensemble learning model is used for training to obtain the baseline value sequence prediction sub-model.

[0229] Specifically, the XGBoost ensemble learning model is a decision tree ensemble algorithm based on a gradient boosting framework. It iteratively constructs multiple decision trees and combines their prediction results in a weighted manner. The training process uses a regularized objective function to control model complexity, including L1 and L2 regularization terms to prevent overfitting. During training, hyperparameters such as the maximum tree depth, learning rate, and subsample rate are set, and the parameter configuration is optimized using a k-fold cross-validation method. The baseline value sequence is used as the target variable to establish a mapping relationship with the input feature vector. The model automatically learns the importance of features and constructs a nonlinear decision boundary. After training, the baseline value sequence prediction sub-model can output the baseline value prediction results corresponding to 96 time periods based on the feature input of the day to be predicted. This model has a strong fitting ability for structured feature data and can effectively capture the complex relationship between meteorological factors and date features and long-term load trends, providing a stable long-term trend prediction component for multi-scale prediction frameworks.

[0230] S42. Based on historical short- and medium-term value sequence data and corresponding meteorological forecast characteristics and date characteristics, a support vector machine model is used for training to obtain a short- and medium-term value sequence prediction sub-model.

[0231] Specifically, the Support Vector Machine (SVM) model is a classification and regression method based on statistical learning theory. It finds the optimal separating hyperplane by mapping input data to a high-dimensional feature space. For short-to-medium-term value sequence prediction tasks, a variant of SVM is used, and a radial basis function kernel function is selected to handle nonlinear relationships. The model training process includes feature standardization, kernel function parameter optimization, and penalty coefficient adjustment. The optimal hyperparameter combination is determined through grid search and cross-validation. Short-to-medium-term value sequence data reflects the weekly cycle characteristics of load and medium-term fluctuations caused by weather system changes. A multi-dimensional input space is constructed by combining meteorological forecast features and date features. The trained short-to-medium-term value sequence prediction sub-model can effectively learn load fluctuation patterns on medium-term time scales. This model has good generalization ability for small sample data, and the kernel trick enables it to handle complex nonlinear interactions between features, making it suitable for capturing the correlation between short-to-medium-term weather changes and load fluctuations.

[0232] S43. Based on the time series data of historical short-term value sequences, a long short-term memory neural network model is trained to obtain a short-term value sequence prediction sub-model.

[0233] Specifically, the Long Short-Term Memory (LSTM) neural network model is a special type of recurrent neural network structure that addresses the long-term dependency problem of traditional RNNs by introducing memory units and gating mechanisms. This model includes three key components: an input gate, a forget gate, and an output gate. It can selectively retain or discard historical information. The training process uses a sliding window of historical short-term value sequences to construct samples. Each window contains observations from several past time periods as input, and the value of the next time period as the prediction target. The network structure design includes the number of hidden layer units, layer configuration, and time step selection. The network weights are adjusted through backpropagation and an adaptive learning rate optimizer. Short-term value sequences represent intraday load change patterns and random fluctuations between adjacent days, exhibiting strong temporal dependence. The LSTM neural network can effectively capture the dynamic change patterns within these sequences. After training, the resulting short-term value sequence prediction sub-model can predict short-term component changes in future periods based on historical short-term fluctuation patterns, providing a fine characterization of intraday fluctuations for multi-scale prediction.

[0234] S44. Based on the high-frequency characteristics of the historical disturbance value sequence, wavelet analysis and statistical methods are used to process the data to obtain a constant value representation of the disturbance value sequence.

[0235] Specifically, the disturbance value sequence contains high-frequency noise and unpredictable random disturbance components from the influencing factor data. Wavelet analysis is used to perform spectral decomposition on historical disturbance components to identify their energy distribution characteristics and statistical regularities. Discrete wavelet transform is then used to decompose the sequence into sub-components of different frequency bands, analyzing the energy proportion and statistical characteristics of each sub-component. Statistical methods are combined to calculate the root mean square value, standard deviation, and quantile statistics of the historical disturbance value sequence. Seasonal grouping analysis is used to determine the differences in disturbance levels under different meteorological conditions. Finally, a weighted average method is used to synthesize various statistical indicators to obtain a disturbance value constant applicable to the meteorological conditions of the day to be predicted. This constant value characterizes the typical amplitude level of the disturbance component. This processing strategy is based on the assumption of the unpredictability of the disturbance component, simplifying it into a constant term. This avoids overfitting to random noise while retaining its contribution to the total influencing factor, providing a stable high-frequency component estimate for multi-scale fusion, and ensuring that the final prediction result includes a reasonable range of uncertainty.

[0236] In some implementations, S5 involves fusing multi-scale prediction results based on the output of each prediction sub-model for the day to be predicted, to obtain a time series of dynamic influencing factors of commercial air conditioning load for the day to be predicted, including:

[0237] S51. Based on the baseline value sequence prediction sub-model, the medium-short-term value sequence prediction sub-model, and the short-term value sequence prediction sub-model, obtain the baseline value sequence prediction results, the medium-short-term value sequence prediction results, and the short-term value sequence prediction results for the date to be predicted, respectively.

[0238] Specifically, the three prediction sub-models each receive feature inputs corresponding to the day to be predicted. The baseline value sequence prediction sub-model processes structured meteorological and date features based on the XGBoost algorithm, outputting 96 predicted values ​​at various time points reflecting long-term trends. The medium- and short-term value sequence prediction sub-model uses the support vector machine algorithm to process similar feature inputs but focuses on mid-frequency fluctuation characteristics, generating prediction sequences that characterize the impact of weekly cycles and weather system changes. The short-term value sequence prediction sub-model processes temporal dependencies through a long short-term memory neural network, using the short-term component values ​​of the previous moment to predict the next moment, forming prediction results of intraday fluctuation characteristics. Each model runs independently but shares the same input data for the prediction day, generating prediction sequences with different time scale characteristics. These sequences maintain the original time resolution and together constitute the complete basis for predicting influencing factors in subsequent fusion steps. The prediction of components at different time scales is completed by specially optimized models, enabling each model to focus on feature extraction and prediction of specific frequency bands, improving the professionalism and adaptability of the overall prediction system.

[0239] S52. Based on the prediction performance of each prediction sub-model on the historical validation set, an optimization algorithm is used to solve for the optimal weight coefficients of each component. The objective function for solving the optimal weight coefficients is to minimize the root mean square error between the ensemble prediction results and the true values ​​on the validation set. The objective function satisfies:

[0240] ;

[0241] ;

[0242] ;

[0243] in, For the first The true value at any given moment. For the first The prediction results obtained at each time point , , These are the forecast results for the baseline value, the medium- and short-term value, and the short-term value, respectively. The value is a constant value in the perturbation value sequence;

[0244] Specifically, the optimization process is conducted on a historical validation set, which contains the actual influencing factor values ​​for multiple historical days and their corresponding predicted values ​​for each component. A constrained optimization problem is solved using sequential quadratic programming or differential evolution algorithms. The objective function quantifies the deviation between the integrated prediction and the actual values. Constraints ensure that the sum of the weight coefficients is 1 and that each coefficient is within the range of 0 to 1. The optimization variables include the weight coefficients of four components: baseline value, short-term value, short-term value, and disturbance value. The optimization algorithm iteratively adjusts the weight combination and evaluates the overall performance of the integrated prediction under different weight configurations. This process considers the performance differences of each prediction sub-model under different meteorological conditions and date types, enabling the weight coefficients to adaptively reflect the contribution of each component in a specific prediction scenario. By minimizing the prediction error on the validation set, the weight allocation ensures that the advantages of each sub-model are maximized, while suppressing the negative impact of poorly performing components on the final result. The resulting optimal weight coefficient set reflects the relative importance of each time-scale component in the integrated prediction.

[0245] S53. Based on the optimal weighting coefficient, the prediction results of each sequence are weighted and fused with the constant value of the disturbance value sequence to obtain the final predicted value of the influence factor at each time of the day to be predicted.

[0246] Specifically, the weighted fusion process is executed independently for each prediction time t. The prediction results of the baseline value sequence, the medium- and short-term value sequence, and the short-term value sequence are multiplied by their respective optimal weight coefficients. The perturbation value sequence is multiplied by its weight coefficient using a pre-determined constant value. The four weighted results are summed to obtain the final predicted value of the influencing factor at that time. The fusion formula follows the principle of linear superposition, preserving the physical meaning and contribution ratio of each component. The fusion process is performed point-by-point in the time dimension, ensuring that there is a corresponding predicted value at each 15-minute interval. The weighted fusion not only preserves the prediction information of each time scale component, but also balances the reliability of different components through optimal weights. The constant value of the perturbation component reflects the statistical characteristics of historical high-frequency perturbations, providing a reasonable range of uncertainty for the prediction results. This fusion mechanism enables the final predicted value to integrate multi-scale characteristics such as long-term trends, medium-term fluctuations, short-term changes, and random perturbations, forming a complete and balanced prediction result. At the same time, the weight constraints ensure the numerical stability and physical rationality of the predicted value.

[0247] S54. Based on the predicted values ​​of the final impact factors at all times, construct a time series of dynamic impact factors for commercial air conditioning load.

[0248] Specifically, the time series construction process arranges the final predicted values ​​of the influencing factors at 96 15-minute intervals in chronological order, forming a continuous sequence covering the entire 24 hours. Each value in the sequence represents the proportion of commercial air conditioning load to the total load during the corresponding time period, and the actual degree of influence after dynamic correction. This time series maintains the temporal resolution and structural characteristics of the original data, and can intuitively show the changing patterns of the influencing factors at different times of the day. The smooth connection of the beginning and end of the sequence ensures the continuity of the daytime transition. The sequence output format meets the data interface requirements of power system dispatching and planning software and can be directly used for subsequent load decomposition and forecasting applications. The constructed dynamic influencing factor time series not only includes point predicted values, but also generates confidence intervals by combining historical error distributions, providing an assessment of forecast uncertainty. As the final output of the method, this sequence provides the power grid dispatching department with fine-grained predictions of the proportion of commercial air conditioning load, supporting power system operation decisions and demand response strategy formulation. At the same time, the sequence data can be stored in a historical database for continuous optimization of the performance of the forecasting model.

[0249] For example,

[0250] (1) XGBoost model training: Based on the data of the above 30 similar historical days, an XGBoost model is constructed for each time t.

[0251] Training sample size: 30 historical days, 96 time points per day = 2880 samples;

[0252] Input feature dimension: The dimension consisting of the above feature vectors (time feature + meteorological feature + historical feature).

[0253] Label: The actual D1 value for each sample.

[0254] The objective function of XGBoost is:

[0255]

[0256] in, The loss function is the mean squared error. The output value of the XGBoost model. For tags, This is a regularization term used to prevent overfitting. Let k be the decision tree.

[0257] Model parameter settings: learning rate 0.1, maximum depth 6, subsampling rate 0.8, feature sampling rate 0.8, number of iterations 200.

[0258] Influence factor prediction for the predicted date: Prediction input data: For the predicted date t (96 time points), the model input is:

[0259]

[0260] Predicted output:

[0261] The model predicts the baseline value D1 for 96 future time points and obtains... .

[0262] (2) Generation of probability prediction intervals: Establish a prediction error library and collect prediction errors on the validation set. Estimate the kernel density of the error distribution:

[0263] ;

[0264] in, For the prediction error, K is the kernel function, using a Gaussian kernel, and h is the bandwidth, determined using the Silverman criterion. For the first One historical error data point.

[0265] Based on the error distribution, a prediction interval with a 95% confidence level is generated:

[0266]

[0267] in, and These are the 2.5% and 97.5% quantiles of the error distribution, respectively.

[0268] 2. Prediction of the short-to-medium-term value series D2

[0269] Short- and medium-term value series reflect the influence of medium-term factors such as weather changes and weekly cycles, and have certain non-linear characteristics. This embodiment uses support vector machine (SVM) for prediction.

[0270] SVM maps the input to a high-dimensional space using kernel functions, and then searches for the optimal hyperplane in that high-dimensional space.

[0271] ;

[0272] in, For support vectors With new samples The inner product in a certain high-dimensional feature space. For the first The Lagrange multipliers corresponding to each support vector For the first The true label or output value of each support vector in the training set. For bias terms;

[0273] This embodiment uses radial basis functions (RBF):

[0274]

[0275] Model parameters: Penalty coefficient C=100, kernel parameters =0.01.

[0276] Input features include: weather forecast data: temperature, humidity, and precipitation probability for the next 24 hours; date features: whether it is a weekday or a holiday; historical D2 series: D2 values ​​for the previous 3 days.

[0277] Training samples: All historical data, totaling 365 sample points.

[0278] 3. Prediction of short-term value series D3

[0279] Short-term value sequences reflect short-term random fluctuations within and between days, exhibiting a clear time dependence. This embodiment uses a Long Short-Term Memory (LSTM) neural network for prediction.

[0280] Network structure: 3-layer LSTM, 128 neurons per layer, dropout rate 0.2, activation function tanh, and the output layer is a fully connected layer.

[0281] Input sequence: 96 D3 values ​​(15-minute intervals) over the past 7 days, forming a time series of shape (7×96, 1).

[0282] Output: D3 prediction value for 96 points in the next day.

[0283] Training parameters: batch size 32, learning rate 0.001, optimizer Adam, 200 iterations.

[0284] Processing of S304 and perturbation value sequence D4

[0285] The perturbation sequence D4 mainly consists of high-frequency noise and unpredictable random perturbations. Directly predicting this sequence is of little significance and has extremely low accuracy. This embodiment uses a feature extraction method to transform it into usable statistical features.

[0286] Specific steps:

[0287] (1) Perform db4 wavelet 3-level decomposition on the historical D4 sequence and extract the low-frequency approximation coefficients. ;

[0288] (2) Calculate the integral average of cA3 over the entire historical time window:

[0289]

[0290] in, For historical data, the time span is 12 months. This represents the number of sampling points.

[0291] (3) In this embodiment, c=0.0021 is calculated and this value is added as a constant term in the final integration during prediction.

[0292] Result integration and verification based on optimal weighting

[0293] 1. Validation set construction and sub-model prediction

[0294] The historical data from January to December 2024 were divided into training set, validation set and backtesting set in chronological order: Training set: January 2024 - June 2024 (181 days); Validation set: July 2024 - September 2024 (92 days); Backtesting set: October 2024 - December 2024 (92 days).

[0295] After training each sub-model using the training set, predictions are made on the validation set: XGBoost predictions are obtained. The validation set RMSE = 0.0162; SVM predictions were obtained. Validation set RMSE=0.0241; LSTM predictions yielded... The validation set RMSE = 0.0318.

[0296] 2. Solving for optimal weights

[0297] The optimal weights are solved using the Particle Swarm Optimization (PSO) algorithm.

[0298] (1) Objective function: Minimize the root mean square error between the ensemble result and the true value on the validation set: ;

[0299] in:

[0300] (2) Constraints: ;

[0301] (3) PSO algorithm parameters: number of particles: 50; maximum number of iterations: 100; inertia weight: 0.7; learning factor: c1=c2=1.5.

[0302] (4) Optimization results: After 91 iterations, the algorithm converged and the optimal weights were obtained: w1=0.44 (baseline value weight); w2=0.32 (short-term value weight); w3=0.21 (short-term value weight); w4=0.03 (perturbation value weight).

[0303] The weight distribution is reasonable: the benchmark value contributes the most, reflecting the dominant role of the long-term trend; the disturbance value has the least weight, which is consistent with its statistical characteristics.

[0304] 3. Generation of the final prediction curve

[0305] The trained sub-models were used to make predictions on the backtesting set (October-December 2024), resulting in... , , By combining the constant c and the optimal weights, the final prediction curve is generated:

[0306]

[0307] Example 2

[0308] Please see Figure 2 This invention provides a device for predicting the dynamic impact factors of commercial air conditioning load, comprising:

[0309] The curve construction module 201 is used to collect power load classification data within a historical period based on the four-level power load classification system, calculate the impact factor and construct the historical curve to obtain the historical impact factor time series corresponding to each historical day within the historical period. The first level of the four-level power load classification system is the total load, the second level is the industrial and commercial electricity consumption and other industry electricity consumption under the total load, the third level is the air conditioning load and other electricity consumption types under the industrial and commercial electricity consumption, and the fourth level is the commercial air conditioning load and other air conditioning load under the air conditioning load. The impact factor is used to characterize the actual impact of the commercial air conditioning load on the total load.

[0310] The decomposition module 202 is used to perform multi-scale data decomposition based on all historical influencing factor time series corresponding to the historical period, to obtain the number of modes and multiple time scale components corresponding to the number of modes. The multiple time scale components include a baseline value sequence, a medium-short-term value sequence, a short-term value sequence, and a disturbance value sequence. The baseline value sequence represents long-term seasonal trends, the medium-short-term value sequence represents weekly cyclical characteristics, the short-term value sequence represents intraday and adjacent-day random changes, and the disturbance value sequence represents high-frequency oscillation components.

[0311] The filtering module 203 is used to filter similar days from the historical period based on the baseline value sequence in the multiple time scale components, the meteorological forecast data of the day to be predicted, and the date characteristics. It extracts the baseline value sequence corresponding to each similar day as a label, and uses the time characteristics and meteorological characteristics as input feature vectors to obtain a training dataset for supervised learning.

[0312] The training module 204 is used to perform multi-model collaborative training based on the training dataset, wherein different time scale components are trained using different models to be trained, so as to obtain prediction sub-models for each time scale component.

[0313] The prediction module 205 is used to perform multi-scale prediction result fusion based on the output results of each prediction sub-model for the predicted day, and obtain the time series of the dynamic influence factor of commercial air conditioning load for the predicted day.

[0314] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting dynamic influencing factors of commercial air conditioning load, characterized in that, include: Based on the four-level classification system of electricity load, historical electricity load classification data were collected for a historical period. Influence factors were calculated and historical curves were constructed to obtain the historical influence factor time series corresponding to each historical day within the historical period. The four-level classification system of electricity load is as follows: the first level is the total load, the second level is the industrial and commercial electricity consumption and other industry electricity consumption under the total load, the third level is the air conditioning load and other electricity consumption types under the industrial and commercial electricity consumption, and the fourth level is the commercial air conditioning load and other air conditioning load under the air conditioning load. The influence factor is used to characterize the actual degree of influence of the commercial air conditioning load on the total load. Based on the time series of all historical influencing factors within the historical period, multi-scale data decomposition is performed to obtain the number of modes and the multiple time scale components corresponding to the number of modes. The multiple time scale components include a baseline value sequence, a medium-short-term value sequence, a short-term value sequence, and a disturbance value sequence. The baseline value sequence represents the long-term seasonal trend, the medium-short-term value sequence represents the weekly cycle characteristics, the short-term value sequence represents intraday and adjacent-day random changes, and the disturbance value sequence represents high-frequency oscillation components. Based on the baseline value sequence in the multiple time scale components, the meteorological forecast data and date characteristics of the day to be predicted, similar days are screened from the historical period, and the baseline value sequence corresponding to each similar day is extracted as a label. The time features and meteorological features are used as input feature vectors to obtain a training dataset for supervised learning. Based on the training dataset, multi-model collaborative training is performed, in which different models are trained for different time scale components to obtain prediction sub-models for each time scale component. Based on the output results of each prediction sub-model for the predicted day, multi-scale prediction results are fused to obtain the time series of dynamic influence factors of commercial air conditioning load for the predicted day.

2. The method according to claim 1, characterized in that, The process involves collecting historical power load classification data based on a four-level power load classification system, calculating influencing factors, and constructing historical curves to obtain the historical influencing factor time series for each day within the historical period, including: Historical power load data is collected based on a predefined four-level classification system for power load. The historical power load data is preprocessed to obtain a cleaned and verified historical load data sequence. The preprocessing includes identifying and removing abnormal data points, using interpolation methods to fill in missing values, and using smoothing algorithms to suppress random mutations in the data. Based on the preprocessed historical load data sequence, the coupling correlation strength between loads at all levels is calculated using the curve similarity measurement method to obtain dynamic correction coefficients, wherein the dynamic correction coefficients reflect the nonlinear correlation characteristics between loads at different time periods and under different meteorological conditions. Based on the commercial air conditioning load data, total load data, and dynamic correction coefficients in the four-level power load classification system, the original influencing factor sequence is calculated. The original impact factor sequence is normalized to map the data range to a preset standard interval, eliminate the influence of units and enhance the comparability of time series, and generate the historical impact factor time series corresponding to each day in the historical period.

3. The method according to claim 2, characterized in that, The process of calculating the coupling correlation strength between load levels based on the preprocessed historical load data sequence using a curve similarity measurement method to obtain dynamic correction coefficients includes: Based on the preprocessed historical load data sequence, normalized daily curves of load at each level are obtained through normalization processing; Based on the normalized daily curve, the similarity between load curves at different levels is calculated using the dynamic time warping algorithm, and cluster analysis is performed based on the similarity matrix to obtain the clustering results. Based on the clustering results and the characteristics of the day to be predicted, the characteristic distance between the day to be predicted and each historical day is calculated, and a weighted average influence factor is constructed based on the distance weight. Based on the degree of meteorological extremes, date type, and historical baseline impact factors, the weighted average impact factors are modified in multiple dimensions to obtain the initial dynamic correction coefficients; The initial dynamic correction coefficients are smoothed and constrained to obtain the dynamic correction coefficients.

4. The method according to claim 1, characterized in that, The process involves multi-scale data decomposition based on the time series of all historical influencing factors within the historical period to obtain the number of modalities and the multiple time-scale components corresponding to the number of modalities, including: Based on the time series of all historical influencing factors within the historical period, a preliminary decomposition is performed using the variational mode decomposition method to obtain a predetermined number of intrinsic mode functions. Based on the frequency characteristics of each intrinsic mode function, a wavelet packet decomposition method is used for secondary decomposition to obtain multiple components with different time scale characteristics. Based on the physical meaning and time scale characteristics of each component, each component is assigned to a baseline value sequence, a medium-short term value sequence, a short term value sequence, and a disturbance value sequence.

5. The method according to claim 1, characterized in that, The process involves filtering similar days from the historical time period based on the baseline value sequence in the multiple time scale components, the meteorological forecast data for the day to be predicted, and the date characteristics. The baseline value sequence corresponding to each similar day is extracted as a label, and the time characteristics and meteorological characteristics are used as input feature vectors to obtain a training dataset for supervised learning, including: A reference sequence is constructed based on the meteorological forecast data and date characteristics of the day to be predicted, and a comparison sequence is constructed based on the corresponding meteorological data and date characteristics of each day in the historical period. Based on the reference sequence and comparison sequence, the correlation degree between the day to be predicted and each historical day is calculated using the grey relational analysis method. The historical days are sorted according to the correlation degree, and a preset number of historical days with the highest correlation degree are selected as the set of similar days; Based on the set of similar days, supervised learning training samples are constructed. The baseline value sequence corresponding to each similar day is extracted as a label, and the time features and meteorological features are used as input feature vectors to obtain the training dataset.

6. The method according to claim 5, characterized in that, The process of constructing a reference sequence based on meteorological forecast data and date characteristics for the day to be predicted, and constructing a comparison sequence based on the corresponding meteorological data and date characteristics for each day in historical time periods, includes: Based on the meteorological forecast data for the day to be predicted, temperature, humidity, and radiation intensity features are extracted and standardized to obtain meteorological feature vectors. Based on the date information of the day to be predicted, extract the date type, weekday and month features and perform one-hot encoding to obtain the date feature vector; A reference sequence is constructed by combining the meteorological feature vector and the date feature vector. Based on the meteorological data and date information corresponding to each day in the historical period, features are extracted and encoded to construct a comparison sequence for each historical day.

7. The method according to claim 5, characterized in that, The step of calculating the correlation degree between the day to be predicted and each historical day using grey relational analysis based on the reference sequence and comparison sequence includes: Based on the reference sequence and the comparison sequence, calculate the absolute difference of each feature dimension; Based on the absolute difference, determine the global minimum difference and the global maximum difference; Based on the absolute difference, global minimum difference, and global maximum difference, the correlation coefficient for each feature dimension is calculated using the resolution coefficient, wherein the correlation coefficient satisfies: ； in, For the first On the 1st historical day Correlation coefficients across each feature dimension For the reference sequence in the 1st Values ​​in each feature dimension For the first The comparison sequence in the th ... Values ​​in each feature dimension The resolution coefficient; Based on the correlation coefficients of each feature dimension, the correlation degree between the day to be predicted and each historical day is calculated.

8. The method according to claim 1, characterized in that, The step involves multi-model collaborative training based on the training dataset, wherein different time scale components are trained using different models to obtain prediction sub-models for each time scale component, including: Based on the input feature vector and baseline value sequence in the training dataset, the XGBoost ensemble learning model is used for training to obtain the baseline value sequence prediction sub-model. Based on historical short- and medium-term value sequence data and corresponding meteorological forecast characteristics and date characteristics, a support vector machine model is used for training to obtain a short- and medium-term value sequence prediction sub-model. Based on the time series data of historical short-term value sequences, a long short-term memory neural network model is trained to obtain a short-term value sequence prediction sub-model; Based on the high-frequency characteristics of the historical perturbation value sequence, wavelet analysis and statistical methods are used to process the data and obtain a constant numerical representation of the perturbation value sequence.

9. The method according to claim 1, characterized in that, The process involves fusing the multi-scale prediction results from each prediction sub-model for the predicted day to obtain a time series of dynamic influencing factors of commercial air conditioning load for the predicted day, including: Based on the baseline value sequence prediction sub-model, the medium-short-term value sequence prediction sub-model, and the short-term value sequence prediction sub-model, the baseline value sequence prediction results, the medium-short-term value sequence prediction results, and the short-term value sequence prediction results for the date to be predicted are obtained respectively. Based on the prediction performance of each prediction sub-model on the historical validation set, an optimization algorithm is used to solve for the optimal weight coefficients of each component. The objective function for solving the optimal weight coefficients is to minimize the root mean square error between the ensemble prediction results and the true values ​​on the validation set. The objective function satisfies the following: ； ； ； in, For the first The true value at any given moment. For the first The prediction results obtained at each time point , , These are the forecast results for the baseline value, the medium- and short-term value, and the short-term value, respectively. The value is a constant value in the perturbation value sequence; Based on the optimal weighting coefficient, the prediction results of each sequence are weighted and fused with the constant value of the disturbance value sequence to obtain the final predicted value of the influence factor at each time of the day to be predicted. Based on the final predicted values ​​of the impact factors at all times, a time series of dynamic impact factors of commercial air conditioning load is constructed.

10. A device for predicting dynamic influencing factors of commercial air conditioning load, characterized in that, include: The curve construction module is used to collect historical power load classification data based on the four-level power load classification system, calculate the impact factor and construct the historical curve to obtain the historical impact factor time series corresponding to each historical day within the historical period. The four-level power load classification system is as follows: the first level is the total load, the second level is the industrial and commercial electricity consumption and other industry electricity consumption under the total load, the third level is the air conditioning load and other electricity consumption types under the industrial and commercial electricity consumption, and the fourth level is the commercial air conditioning load and other air conditioning load under the air conditioning load. The impact factor is used to characterize the actual impact of the commercial air conditioning load on the total load. The decomposition module is used to perform multi-scale data decomposition based on the time series of all historical influencing factors corresponding to the historical period, to obtain the number of modes and the multiple time scale components corresponding to the number of modes. The multiple time scale components include a baseline value sequence, a medium-short-term value sequence, a short-term value sequence, and a disturbance value sequence. The baseline value sequence represents the long-term seasonal trend, the medium-short-term value sequence represents the weekly cycle characteristics, the short-term value sequence represents the intraday and adjacent day random changes, and the disturbance value sequence represents the high-frequency oscillation component. The filtering module is used to filter similar days from the historical period based on the baseline value sequence in the multiple time scale components, the meteorological forecast data of the day to be predicted, and the date characteristics. It extracts the baseline value sequence corresponding to each similar day as a label, and uses the time characteristics and meteorological characteristics as input feature vectors to obtain a training dataset for supervised learning. The training module is used to perform multi-model collaborative training based on the training dataset, wherein different models are trained for different time scale components to obtain prediction sub-models for each time scale component. The prediction module is used to perform multi-scale prediction result fusion based on the output results of each prediction sub-model for the day to be predicted, and obtain the time series of dynamic influence factors of commercial air conditioning load for the day to be predicted.

Images