A method and system for evaluating energy saving potential of a building park based on machine learning

Abstract

The present application relates to the technical field of building energy saving and energy management, in particular to a building park energy saving potential evaluation method and system based on machine learning, in use, the multi-source energy data acquisition module is subdivided to building, floor, equipment level to collect energy consumption data, and key factors such as meteorology, passenger flow and equipment working condition are fused, based on the 3-principle outlier elimination scheme and the standardized algorithm adapting to data distribution, the problem of single data dimension and insufficient continuity is effectively solved, high-quality data support is provided for evaluation, at the same time, cross features such as temperature and air conditioning load, passenger flow and equipment power consumption are extracted through feature engineering, combined with the hybrid model of the first sub-model capturing long and short-term dependence and the second sub-model correcting deviation, the problem that the traditional model cannot handle multi-factor interaction is solved, the prediction accuracy is improved, and reliable future energy consumption data is provided for energy saving potential evaluation.

Description

Technical Field

[0001] This invention relates to the field of building energy conservation and energy management technology, specifically to a method and system for assessing the energy-saving potential of building parks based on machine learning. Background Technology

[0002] With the large-scale development of building parks and the diversification of energy-consuming equipment, the proportion of energy consumption in the total social energy consumption has been increasing year by year. Energy-saving renovation of parks has become a key path to reduce total social energy consumption and carbon emissions, and accurate energy-saving potential assessment is the core prerequisite for ensuring the scientific and efficient implementation of renovation plans.

[0003] Most existing assessment methods only collect overall electricity consumption data for the park, without breaking it down to the building, floor, and equipment levels. They also fail to fully integrate key factors affecting energy consumption, such as weather, pedestrian flow, and equipment operating conditions, resulting in a single data dimension. Furthermore, the handling of missing data and outliers often involves simple filling or deletion, neglecting data continuity and trend consistency, which affects the accuracy of subsequent analysis. In addition, traditional prediction models struggle to capture the long-term and short-term dependencies of energy consumption data and the interaction of multiple factors, failing to provide reliable future energy consumption data support for energy-saving potential assessment.

[0004] In summary, existing energy-saving potential assessment schemes are insufficient to meet the needs of large-scale and refined energy-saving renovations in industrial parks. There is an urgent need for an assessment method and system that integrates multi-source data, precise data mining models, multi-dimensional assessments, and can be optimized in a closed loop. Summary of the Invention

[0005] This invention effectively solves the problems of insufficient data dimension and continuity, provides high-quality data support for evaluation, solves the problem that traditional models cannot handle the interactive influence of multiple factors, and improves prediction accuracy.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides a machine learning-based method for assessing the energy-saving potential of building parks, comprising the following steps:

[0008] S1. Collect energy characteristic data of the building park, preprocess the energy characteristic data, and output a standardized dataset;

[0009] S2. Based on standardized datasets, mine the time-of-use electricity patterns in building parks, construct a user behavior tagging system, and generate electricity user behavior profiles;

[0010] S3. Perform feature engineering on the standardized dataset, construct and train a multi-source fusion energy consumption prediction model, and output future time-of-use energy consumption prediction results;

[0011] S4. Combine historical renovation data with prediction results to calculate the overall energy-saving potential and energy-saving level of the building park;

[0012] S5. Generate energy-saving measures based on user behavior profiles and energy-saving assessment results, and link them with the virtual power plant dispatch platform;

[0013] S6. Select pilot parks for energy-saving renovation, collect actual data after the pilot renovation, and optimize the energy consumption prediction model and energy-saving potential assessment logic through pilot deviation verification.

[0014] Furthermore, the energy characteristic data includes basic energy consumption data, influencing factor data, and historical renovation data. The basic energy consumption data includes the time-of-use electricity consumption and electrical parameters of each building, floor, and equipment in the park. The sampling interval for the time-of-use electricity consumption is set as a first preset interval, determined based on the energy consumption fluctuation cycle of the building park's energy-consuming equipment. This is selected through wavelet analysis of 7 consecutive days of energy consumption data from the equipment, with each sample taken every 15 minutes. The electrical parameters include voltage, current, and power factor. The equipment includes air conditioners, lighting, charging piles, and office equipment. The influencing factor data includes meteorological data, pedestrian flow data, equipment operating condition data, and electricity price data. The sampling interval for the meteorological data is set as a second preset interval, determined based on the lag statistical impact of meteorological factors on energy consumption. This is selected by analyzing the correlation between meteorological data and energy consumption data. The historical renovation data includes the technical solutions for implemented energy-saving renovations, energy consumption comparison data for a preset period before and after the renovation, renovation investment costs, and energy-saving benefit data. The preset period is determined based on the stable observation duration of the energy-saving renovation effect, selected through energy consumption trend analysis for 3 to 24 months after the renovation, with energy consumption fluctuations of less than or equal to 5% for 6 consecutive months. Determined to be stable;

[0015] In step S1, the preprocessing includes missing value imputation, outlier removal, and data standardization. Missing value imputation is performed on time-series data with consecutive missing durations less than or equal to a first preset duration. The first preset duration is determined based on statistical analysis of historical missing data patterns in the park, selected by analyzing missing data records from the past three years. The first preset duration is 2 hours, and imputation is performed using time-series interpolation based on adjacent valid data. For time-series data with consecutive missing durations greater than the first preset duration, historical complete data segments with similar meteorological conditions, weekday and restday attributes, and energy consumption patterns to the missing period are selected for replacement, ensuring the continuity and trend consistency of the data sequence. Outlier removal is based on a preset deviation rule, marking data exceeding the normal fluctuation range as outliers and removing them. The preset deviation rule is used to identify abnormal fluctuation data in the data sequence, employing a 3... The principle is to identify and remove outliers. When implementing this principle, the first step is to screen data from the past 30 days showing normal operating conditions where the equipment load is between 30% and 80% of its rated capacity. This range is determined based on the equipment's design operating parameters and industry operating standards to ensure the data source is stable operating conditions. Based on the screened normal operating condition data, the mean and standard deviation of energy characteristic data for time-of-use electricity consumption, voltage, and current are calculated for this period. The mean μ is the arithmetic mean of the data sample, and the standard deviation... This quantifies the degree of dispersion of the data samples from the mean; it also quantifies the dispersion of energy characteristic data exceeding μ±3. Values ​​within a certain range are marked as outliers and removed. This range is determined based on the characteristics of a normal distribution and can cover 99.73% of normal data, effectively excluding extreme fluctuation data. The data standardization adopts a preset normalization algorithm to map non-electrical data to a preset numerical range, achieving scale uniformity for data of different dimensions. The preset normalization algorithm is used to eliminate the dimensional differences of different types of data. It is selected based on the data distribution characteristics. When the data is uniformly distributed, min-max normalization is used, and when it is normally distributed, z-score standardization is used. The preset numerical range is [0,1], which is suitable for the value range of the activation function of the subsequent energy prediction model.

[0016] Further, in step S2, the time-of-use electricity consumption pattern is mined using a clustering algorithm. The clustering algorithm adopts K-means clustering or DBSCAN clustering. The number of clusters in the clustering algorithm is determined by a preset silhouette coefficient rule. The mined time-of-use electricity consumption patterns include high-load continuous type, intermittent fluctuation type, and low-load stable type. The association rule algorithm is combined to mine the correlation between influencing factor data and time-of-use electricity consumption patterns. Key influencing factors with a correlation degree greater than a first preset correlation degree threshold are screened. The first preset correlation degree threshold is selected based on the contribution of the influencing factors to the classification accuracy of the electricity consumption pattern. The selection number is 0.7, and the classification accuracy is greater than or equal to 85%. The association rule algorithm is used to mine the potential correlation between influencing factors and electricity consumption patterns.

[0017] In step S2, the user behavior tagging system is a three-dimensional tagging system, which includes energy intensity tags, load fluctuation tags, and adjustable potential tags. The energy intensity tags are divided into three levels: high, medium, and low, based on electricity consumption thresholds. These thresholds include a first electricity consumption threshold and a second electricity consumption threshold, selected based on the quantiles of the park's historical energy consumption data over the past year. The energy consumption data sample size is greater than or equal to 365 groups. The first electricity consumption threshold is 120 kWh / day, and the second electricity consumption threshold is 60 kWh / day. The load fluctuation tags are divided into three levels: stable, medium, and fluctuating, based on daily load fluctuation coefficients. These fluctuation coefficient thresholds include a first fluctuation coefficient threshold and a second fluctuation coefficient threshold. The values ​​are selected based on the statistical distribution of the daily load fluctuation coefficient over a year. The first fluctuation coefficient threshold is 0.4, and the second fluctuation coefficient threshold is 0.2. The adjustable potential label is divided into three levels: high, medium, and low, according to the load reduction ratio threshold. The ratio threshold includes a first ratio threshold and a second ratio threshold. The first and second ratio thresholds are selected based on the equipment load regulation capacity test data. Typically, the first ratio threshold is 30%, and the second ratio threshold is 15%. By inputting key influencing factors and time-of-use electricity consumption feature vectors into the classification model, a power user behavior profile is generated. The classification model is used to generate user behavior labels based on key features. The high load persistence label corresponds to the energy intensity label with applied electricity exceeding the first electricity consumption threshold, and the adjustable potential label corresponds to the load reduction ratio exceeding the first ratio threshold.

[0018] Furthermore, in step S3, the feature engineering includes extracting time-series features and cross features from the standardized dataset, filtering key features through mutual information entropy, and eliminating redundant features with a correlation degree less than a second preset correlation degree threshold. The mutual information entropy is used to measure the degree of correlation between features and electricity consumption. The time-series features include the average electricity consumption over the previous third preset period and the electricity consumption during the same period of the previous preset period. The third preset period is the same as the preset period before and after the historical transformation data in step S1, which is twelve months, to ensure the uniformity of the data period.

[0019] In step S3, the construction and training of the energy prediction model includes the following sub-steps:

[0020] S31. Construct a multi-source fusion hybrid prediction model, which includes a first prediction sub-model and a second prediction sub-model. The first prediction sub-model is used to capture the long-term and short-term dependencies of energy consumption data, and the second prediction sub-model is used to correct time-series prediction bias.

[0021] S32. The selected key features are divided into training set and test set. The division ratio is executed according to the preset data division ratio. The preset data division ratio is determined based on the sample size of the dataset. When the sample size is greater than 10,000, the ratio is 7:3. When the sample size is less than or equal to 10,000, the ratio is 8:2. The first prediction sub-model inputs time series features. The number of neurons in the hidden layer of the first prediction sub-model is set to the first preset number of neurons range, which is 64-128. The second prediction sub-model inputs cross features. The cross features include temperature and air conditioning load interaction features and human flow and equipment power consumption combination features. The tree depth is set to the first preset tree depth range, which is 5-8 layers.

[0022] S33. Using the mean absolute percentage error as the loss function, a preset early stopping rule is adopted to prevent model overfitting. The preset early stopping rule is to stop training if the validation set error does not decrease for 5 consecutive rounds. This is used to avoid the model being overtrained and causing a decrease in generalization ability. The model hyperparameters are adjusted through a grid search algorithm. The parameter optimization algorithm is used to find the optimal combination of hyperparameters for the model. After training, the model's prediction mean absolute percentage error is less than or equal to a preset error threshold. The preset error threshold is less than or equal to 8%. The time-sharing energy consumption prediction results for the fourth preset duration of 24 hours, the fifth preset duration of 7 days, and the sixth preset duration of 30 days are output.

[0023] Furthermore, in step S4, the comprehensive energy saving of the building park is constructed using a multi-dimensional energy-saving potential assessment index system. The energy-saving potential assessment index system includes a technical dimension, an economic dimension, and a scheduling dimension. Under the technical dimension, two indicators are set: equipment energy-saving potential and system loss reduction potential. Under the economic dimension, two indicators are set: energy-saving benefits and investment payback period. Under the scheduling dimension, two indicators are set: adjustable load and response speed. The response speed is within a preset response speed threshold, which is less than or equal to 5 minutes.

[0024] In step S4, the weights of the evaluation indicators are determined by a combined weighting method, which combines the analytic hierarchy process (AHP) and the entropy weighting method. In the subjective weights determined by the AHP, the weights of technological potential, economic potential, and scheduling potential are set as a first preset weight percentage, a second preset weight percentage, and a third preset weight percentage, respectively, representing 45%, 35%, and 20%. The entropy weighting method calculates objective weights based on the dispersion of each indicator in historical transformation data. This dispersion directly reflects the degree of difference between different indicator data, ensuring that the calculation of objective weights aligns with the actual transformation scenario.

[0025] The combined weights are calculated according to a weighting formula, the expression of which is:

[0026] ,

[0027] In the formula, The final combined weight of a certain evaluation index. The subjective weights of this evaluation indicator, determined using the analytic hierarchy process, represent the subjective weights. The objective weights for this evaluation indicator are determined using the entropy weight method. The first preset weighting coefficient, with a value of 0.6, is used to reflect subjective weighting. The second preset weighting coefficient, with a value of 0.4, is used to balance subjective weighting, and =1, to ensure the rationality and normalization of weight calculation, and to avoid the combined weight deviating from the actual needs due to coefficient imbalance;

[0028] The overall energy-saving potential of the building complex is expressed by the formula: Calculate, where, The combined weight of the i-th evaluation index is calculated using the above symbolic weighting formula. Let i be the quantified value of the i-th evaluation indicator. To evaluate the total number of indicators, For each of the six corresponding equipment, the potential for energy saving, the potential for reducing system losses, the energy saving benefits, the investment payback period, the adjustable load, and the response speed;

[0029] The energy-saving level is determined according to a preset level classification rule, which is the comprehensive energy-saving potential. When the potential is greater than the first potential threshold, it is judged as a level one energy-saving potential. When the potential value is between the second potential threshold and the first potential threshold, it is determined to be a level two energy-saving potential. If the potential is less than the second potential threshold, it is judged as a level three energy-saving potential. The first potential threshold and the second potential threshold are 80 and 60, respectively.

[0030] Furthermore, in step S5, the energy-saving measure generation logic is determined based on the matching relationship between user behavior profile tags and energy-saving levels. When the user behavior profile tag is high load continuous type and the energy-saving level of the building park is determined to be Level 1, high-yield energy-saving solutions are selected and matched from historical renovation data. The high-yield energy-saving solutions are designed for high load scenarios and can achieve high-efficiency energy saving. When the user behavior profile tag is high adjustable potential type and the energy-saving level is determined to be Level 1, the load transfer strategy is calculated based on the peak-valley electricity price difference and adjustable load, and a virtual power plant peak-valley arbitrage strategy is generated.

[0031] In step S5, the linkage with the virtual power plant dispatching platform involves first uploading the adjustable load characteristic data recorded in the user behavior profile and the adjustable load data calculated in step S4 to the virtual power plant dispatching platform. The data upload frequency is consistent with the first preset interval. After receiving the load adjustment instruction issued by the platform, the parameters of the park's power equipment are dynamically adjusted according to the preset equipment adjustment rules. The preset equipment adjustment rules specify the specific adjustment methods for the operating parameters of different types of equipment to ensure that the normal power consumption experience of the park is guaranteed while responding to the load dispatching needs of the virtual power plant.

[0032] Furthermore, in step S6, the energy-saving renovation pilot parks include office parks, commercial parks and science and education parks, with a building area difference of more than or equal to 5,000 square meters for each type of park. The energy-saving levels of the selected parks include Level 1 and Level 2. Each type of park includes one Level 1 park and one Level 2 park. The selection of different types of parks is to ensure the scenario applicability of the pilot results.

[0033] In step S6, the actual data collected after the pilot transformation is implemented according to the energy-saving measures generated in step S5. The actual data collected for twelve months after the transformation is carried out continuously, including actual time-of-use energy consumption data, actual energy-saving benefit data and actual response data of virtual power plant linkage. The collection frequency is matched with the monthly electricity bill settlement cycle.

[0034] Furthermore, in step S6, the pilot deviation verification performs a consistency check on the collected actual data, using 3... Outliers are removed in principle. The actual time-of-use energy consumption data is reprocessed according to standardized rules to ensure consistency with the units and statistical dimensions of the predicted data, avoiding misjudgments caused by differences in statistical caliber. The preprocessed actual data is compared one by one with the time-of-use energy consumption prediction data and the comprehensive energy-saving potential assessment data to calculate the energy-saving potential deviation and response speed deviation. Based on the park type, the park type coefficient is set to 1.0 for office parks, 1.2 for commercial parks, and 1.1 for science and education parks. The judgment threshold is calculated according to the dynamic threshold = basic threshold × type coefficient. If both the potential deviation and the response speed deviation are less than or equal to the corresponding dynamic threshold, the judgment is verified. If the potential deviation or the response speed deviation is greater than the corresponding dynamic threshold, the model and indicator optimization process is triggered.

[0035] Furthermore, the model and index optimization adjusts the number of hidden layer neurons by ±16 within a first preset range for the first prediction sub-model, and adjusts the tree depth by ±1 within a first preset range for the second prediction sub-model. After adjustment, the model is retrained. Based on the actual data after the pilot park renovation, the actual threshold of user behavior tags is calculated. Two to three new parks that did not participate in the pilot are selected. The energy-saving levels of the new parks cover Level 1 and Level 2 to ensure that the optimization effect is applicable to different potential scenarios. Steps S1-S4 are re-executed to calculate the energy-saving potential deviation and response speed deviation of the new parks. If both deviations are less than or equal to the dynamic threshold, the optimization is deemed effective. If the potential deviation or response speed deviation is greater than the corresponding dynamic threshold, the above parameter iteration and threshold correction steps are repeated until both deviations are less than or equal to the dynamic threshold.

[0036] On the other hand, the present invention also provides a machine learning-based building park energy-saving potential assessment system, which includes: an energy characteristic data acquisition module, a data preprocessing module, a generation module, an energy consumption prediction model module, an energy-saving measures and scheduling linkage module, and a pilot verification and model optimization module.

[0037] The energy characteristic data acquisition module collects multi-dimensional energy characteristic data such as time-of-use electricity consumption, electrical parameters, meteorological data, and population flow data through a hardware interface unit, an environmental data access unit, and a historical data storage unit, and uses a time-series database to store historical renovation-related data.

[0038] The data preprocessing module completes the cleaning and standardization of energy characteristic data through missing value processing, outlier removal, and data standardization units.

[0039] The generation module uses pattern clustering, correlation feature filtering, and three-dimensional label generation unit to mine time-of-use electricity patterns, construct and output electricity user behavior profiles;

[0040] The energy consumption prediction model module outputs time-of-use energy consumption prediction results for different future durations through feature engineering, hybrid prediction model construction, and training optimization units.

[0041] The energy-saving measures and scheduling linkage module achieves coordinated linkage between energy-saving measures generation and the scheduling platform through the measures matching unit and the virtual power plant linkage unit;

[0042] The pilot verification and model optimization module completes pilot data collection, deviation verification, and model and label threshold optimization through pilot park management, actual data collection, deviation verification, and model index optimization units.

[0043] Compared with existing technologies, the advantages and positive effects of this invention are that it collects energy consumption data at the building, floor, and equipment levels through an energy characteristic data acquisition module, and integrates key factors such as weather, pedestrian flow, and equipment operating conditions, based on 3D... The principle-based outlier removal scheme and standardized algorithm adapted to data distribution effectively solve the problems of single data dimension and insufficient continuity, providing high-quality data support for assessment. At the same time, through feature engineering, cross-features such as temperature and air conditioning load, and pedestrian flow and equipment power consumption are extracted. Combined with the first sub-model to capture long-term and short-term dependencies and the second sub-model to correct biases, a hybrid prediction model is formed, which solves the problem that traditional models cannot handle the interaction of multiple factors, improves prediction accuracy, and provides reliable future energy consumption data for energy-saving potential assessment. Attached Figure Description

[0044] Figure 1 This is a flowchart of a machine learning-based method for assessing the energy-saving potential of building parks according to the present invention;

[0045] Figure 2 This is a model diagram of a machine learning-based method and system for assessing the energy-saving potential of building parks according to the present invention. Detailed Implementation

[0046] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0047] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0048] The present invention will now be described in further detail with reference to the accompanying drawings:

[0049] Example 1:

[0050] like Figures 1-2 As shown, this invention provides a machine learning-based method for assessing the energy-saving potential of building parks, comprising the following steps:

[0051] S1. Collect energy characteristic data of the building park, preprocess the energy characteristic data, and output a standardized dataset.

[0052] Furthermore, the energy characteristic data includes basic energy consumption data, influencing factor data, and historical renovation data. Basic energy consumption data includes the time-of-use electricity consumption and electrical parameters of each building, floor, and equipment in the park. The sampling interval for time-of-use electricity consumption is set as the first preset interval, which is determined based on the energy consumption fluctuation cycle of the energy-consuming equipment in the building park. It is selected through wavelet analysis of the energy consumption data of the equipment for 7 consecutive days, with each sample taken every 15 minutes. The electrical parameters include voltage, current, and power factor. The equipment includes air conditioners, lighting, charging piles, and office equipment. Influencing factor data includes meteorological data, pedestrian flow data, equipment operating condition data, and electricity price data. The sampling interval for meteorological data is set as the second preset interval, which is determined based on the lag statistical analysis of the impact of meteorological factors on energy consumption. It is selected by analyzing the correlation between meteorological data and energy consumption data. Historical renovation data includes the technical solutions for energy-saving renovations that have been implemented, the energy consumption comparison data before and after the renovation for the preset period, the renovation investment cost, and the energy-saving benefit data. The preset period is determined based on the stable observation period of the energy-saving renovation effect. It is selected through the analysis of energy consumption trends for 3 to 24 months after the renovation. Energy consumption fluctuations of less than or equal to 5% for 6 consecutive months are considered stable.

[0053] Further preprocessing includes missing value imputation, outlier removal, and data standardization. Missing value imputation is performed on time-series data with consecutive missing durations less than or equal to a first preset duration. The first preset duration is determined based on statistical analysis of historical missing data patterns in the park, selected through analysis of missing data records from the past three years. The first preset duration is 2 hours, and imputation is performed using time-series interpolation based on adjacent valid data. For time-series data with consecutive missing durations greater than the first preset duration, historical complete data segments with similar meteorological conditions, weekday and restday attributes, and energy consumption patterns to the missing period are selected for replacement, ensuring the continuity and trend consistency of the data sequence. Outlier removal is based on preset deviation rules, marking data exceeding the normal fluctuation range as outliers and removing them. The preset deviation rules are used to identify abnormal fluctuation data in the data sequence, employing a 3... The principle is to identify and remove outliers. When implementing this principle, the first step is to screen data from the past 30 days showing normal operating conditions where the equipment load is between 30% and 80% of its rated capacity. This range is determined based on the equipment's design operating parameters and industry operating standards to ensure the data source is stable operating conditions. Based on the screened normal operating condition data, the mean and standard deviation of energy characteristic data for time-of-use electricity consumption, voltage, and current are calculated for this period. The mean μ is the arithmetic mean of the data sample, and the standard deviation... This quantifies the degree of dispersion of the data samples from the mean; it also quantifies the dispersion of energy characteristic data exceeding μ±3. Values ​​within a certain range are marked as outliers and removed. This range is determined based on the characteristics of a normal distribution and can cover 99.73% of normal data, effectively excluding extreme fluctuation data. Data standardization adopts a preset normalization algorithm to map non-electrical data to a preset numerical range, achieving scale uniformity for data of different dimensions. The preset normalization algorithm is used to eliminate the dimensional differences of different types of data. It is selected based on the data distribution characteristics. When the data is uniformly distributed, min-max normalization is used, and when it is normally distributed, z-score standardization is used. The preset numerical range is [0,1], which is suitable for the value range of the activation function of the subsequent energy consumption prediction model.

[0054] S2. Based on standardized datasets, mine the time-of-use electricity patterns in building parks, construct a user behavior tagging system, and generate electricity user behavior profiles.

[0055] Furthermore, time-of-use (TOU) electricity patterns are mined using clustering algorithms, specifically K-means or DBSCAN. The number of clusters is determined by a preset silhouette coefficient rule. The mined TOU electricity patterns include high-load continuous, intermittent fluctuating, and low-load stable patterns. Association rule algorithms are then used to mine the correlation between influencing factor data and TOU electricity patterns. Key influencing factors with a correlation degree greater than a first preset correlation degree threshold are selected. This first preset correlation degree threshold is chosen based on the contribution of the influencing factors to the classification accuracy of the electricity patterns, with a selection factor of 0.7. The classification accuracy is greater than or equal to 85%. The association rule algorithm is used to mine the potential correlation between influencing factors and electricity patterns.

[0056] Furthermore, the user behavior tagging system is a three-dimensional tagging system, including energy intensity tags, load fluctuation tags, and adjustability potential tags. Energy intensity tags are divided into three levels: high, medium, and low, based on electricity consumption thresholds. These thresholds include a first electricity consumption threshold and a second electricity consumption threshold, selected based on the quantiles of the park's historical energy consumption data over the past year. The energy consumption data sample size is greater than or equal to 365 groups, with the first threshold at 120 kWh / day and the second at 60 kWh / day. Load fluctuation tags are divided into three levels: stable, medium, and fluctuating, based on daily load fluctuation coefficients. These thresholds include a first fluctuation coefficient threshold and a second fluctuation coefficient threshold, based on... Based on the statistical distribution of daily load fluctuation coefficients over a year, the first fluctuation coefficient threshold is 0.4 and the second fluctuation coefficient threshold is 0.2. Adjustable potential labels are divided into three levels: high, medium, and low, according to the threshold of the proportion of load that can be reduced. The proportion thresholds include the first proportion threshold and the second proportion threshold. The first and second proportion thresholds are selected based on the test data of equipment load regulation capacity, with the first proportion threshold being 30% and the second proportion threshold being 15%. By inputting key influencing factors and time-of-use electricity consumption feature vectors into the classification model, a profile of electricity user behavior is generated. The classification model is used to generate user behavior labels based on key features. The high load persistence label corresponds to the energy intensity label with applied electricity exceeding the first electricity consumption threshold, and the adjustable potential label corresponds to the adjustable potential label with the proportion of load that can be reduced exceeding the first proportion threshold.

[0057] S3. Perform feature engineering on the standardized dataset, construct and train a multi-source fusion energy consumption prediction model, and output future time-of-use energy consumption prediction results.

[0058] Furthermore, feature engineering includes extracting temporal and cross features from the standardized dataset, filtering key features through mutual information entropy, and eliminating redundant features with a correlation degree less than the second preset correlation degree threshold. Mutual information entropy is used to measure the degree of correlation between features and electricity consumption. Temporal features include the average electricity consumption over the previous third preset period and the electricity consumption during the same period of the previous preset period. The third preset period is the same as the preset period before and after the historical transformation data in step S1, which is twelve months, to ensure the uniformity of the data period.

[0059] Furthermore, the construction and training of the predictive model includes the following steps:

[0060] S31. Construct a multi-source fusion hybrid prediction model. The hybrid prediction model includes a first prediction sub-model and a second prediction sub-model. The first prediction sub-model is used to capture the long-term and short-term dependencies of energy consumption data, and the second prediction sub-model is used to correct time series prediction bias.

[0061] S32. Divide the selected key features into training set and test set. The division ratio is executed according to the preset data division ratio. The preset data division ratio is determined based on the sample size of the dataset. When the sample size is greater than 10,000, the ratio is 7:3. When the sample size is less than or equal to 10,000, the ratio is 8:2. The first prediction sub-model inputs time series features. The number of neurons in the hidden layer of the first prediction sub-model is set to the first preset range of neurons, which is 64-128. The second prediction sub-model inputs cross features. Cross features include temperature and air conditioning load interaction features and human flow and equipment power consumption combination features. The tree depth is set to the first preset range of tree depth, which is 5-8 layers.

[0062] S33. Using the mean absolute percentage error as the loss function, a preset early stopping rule is adopted to prevent model overfitting. The preset early stopping rule is to stop training if the validation set error does not decrease for 5 consecutive rounds. This is used to avoid the model being overtrained and causing a decrease in generalization ability. The model hyperparameters are adjusted through a grid search algorithm, and the parameter optimization algorithm is used to find the optimal combination of hyperparameters. After training, the model's prediction mean absolute percentage error is less than or equal to a preset error threshold. The preset error threshold is less than or equal to 8%. The time-sharing energy consumption prediction results for the fourth preset duration of 24 hours, the fifth preset duration of 7 days, and the sixth preset duration of 30 days are output.

[0063] S4. Combine historical renovation data with forecast results to calculate the overall energy-saving potential and energy-saving level of the building park.

[0064] Furthermore, a multi-dimensional energy-saving potential assessment index system is constructed for the comprehensive energy conservation of the building park. The energy-saving potential assessment index system includes technical, economic and scheduling dimensions. Under the technical dimension, two indicators are set: equipment energy-saving potential and system loss reduction potential. Under the economic dimension, two indicators are set: energy-saving benefits and investment payback period. Under the scheduling dimension, two indicators are set: adjustable load and response speed. The response speed is within a preset response speed threshold, which is less than or equal to 5 minutes.

[0065] The weights of the evaluation indicators are determined using a combined weighting method, which combines the analytic hierarchy process (AHP) and the entropy weighting method. Within the subjective weight system determined by the AHP, the weight percentages corresponding to technical potential, economic potential, and dispatch potential are set as the first, second, and third preset weight percentages, respectively, at 45%, 35%, and 20%. Specifically, the subjective weight values ​​are: equipment energy saving potential 25%, system loss reduction potential 20%, energy saving benefit 20%, investment payback period 15%, adjustable load 10%, and response speed 10%. These values ​​are based on Appendix B.0.2 of the "General Specification for Building Energy Conservation and Renewable Energy Utilization" GB55015-2021, which states "Technical category ≥ 40%, Economic category ≥ 30%, Dispatch category ≤ 25%", and the "Technical Guidelines for Energy Conservation Assessment". According to Article 5.2.2 of JGJ / T223-2010, "the weight of technical categories is higher than that of the other two categories." The total weight of technical dimensions is 45%, the total weight of economic dimensions is 35%, and the total weight of scheduling dimensions is 20%, satisfying the requirement of a total weight of 100% and the standard constraint. The entropy weight method calculates the objective weight based on the dispersion of each indicator in historical transformation data. The dispersion can intuitively reflect the degree of difference between different indicator data, ensuring that the calculation basis of objective weight is close to the actual transformation scenario. Specifically, the objective weight values ​​are: equipment energy saving potential 0.200, system loss reduction potential 0.175, energy saving benefit 0.150, investment payback period 0.125, adjustable load 0.100, and response speed 0.250. These values ​​are calculated according to the dispersion method of Article 6.3.2 of the "Technical Guidelines for Energy Saving Assessment" GB / T5271.34-2022. After dispersion normalization, the values ​​are directly taken, with the total dispersion fixed at 4.0 and the total weight sum = 1.0.

[0066] Furthermore, the combined weights are calculated using the weight formula, which is expressed as follows: In the formula, The final combined weight of a certain evaluation index. The subjective weights of this evaluation indicator, determined using the analytic hierarchy process, represent the subjective weights. The objective weights for this evaluation indicator are determined using the entropy weight method. The first preset weighting coefficient, with a value of 0.6, is used to reflect subjective weighting. The second preset weighting coefficient, with a value of 0.4, is used to balance subjective weighting, and =1, ensuring the rationality and normalization of weight calculation, and avoiding deviation of combined weights from actual needs due to coefficient imbalance. and The value is determined based on Article 12 of the "Measures for Energy Conservation Review of Fixed Asset Investment Projects" (2023 edition), which states that the subjective weight is ≥50% and the known standard for weighting engineering combinations. It is the fixed combination with the smallest error rate.

[0067] Furthermore, the overall energy-saving potential of the building complex can be expressed by the formula: Calculate, where, The combined weight of the i-th evaluation index is calculated using the above symbolic weighting formula. Let i be the quantified value of the i-th evaluation indicator. To evaluate the total number of indicators, The corresponding equipment energy-saving potential, system loss reduction potential, energy-saving benefits, investment payback period, adjustable load, and response speed are 6.

[0068] Furthermore, the energy efficiency rating is determined according to a preset rating classification rule, which is based on comprehensive energy efficiency potential. When the potential is greater than the first potential threshold, it is judged as a level one energy-saving potential. When the potential value is between the second potential threshold and the first potential threshold, it is determined to be a level two energy-saving potential. If the potential is less than the second potential threshold, it is judged as a level three energy-saving potential. The first potential threshold and the second potential threshold are 80 and 60, respectively.

[0069] S5. Based on user behavior profiles and energy-saving assessment results, generate energy-saving measures and link them with the virtual power plant dispatch platform.

[0070] Furthermore, the logic for generating energy-saving measures is determined based on the matching relationship between user behavior profile tags and energy-saving levels. When the user behavior profile tag is high load continuous type and the energy-saving level of the building park is determined to be Level 1, high-yield energy-saving solutions are selected and matched from historical renovation data. High-yield energy-saving solutions are designed for high load scenarios and can achieve high-efficiency energy saving. When the user behavior profile tag is high adjustable potential type and the energy-saving level is determined to be Level 1, the load transfer strategy is calculated based on the peak-valley electricity price difference and adjustable load, and a virtual power plant peak-valley arbitrage strategy is generated.

[0071] Furthermore, in conjunction with the virtual power plant dispatch platform, the adjustable load characteristic data recorded in the user behavior profile and the adjustable load data calculated in step S4 are first uploaded to the virtual power plant dispatch platform. The data upload frequency is consistent with the first preset interval. After receiving the load adjustment instruction issued by the platform, the parameters of the park's power equipment are dynamically adjusted according to the preset equipment adjustment rules. The preset equipment adjustment rules specify the specific adjustment methods for the operating parameters of different types of equipment to ensure that the normal power consumption experience of the park is guaranteed while responding to the load dispatch needs of the virtual power plant.

[0072] S6. Select pilot parks for energy-saving renovation, collect actual data after the pilot renovation, and optimize the energy consumption prediction model and energy-saving potential assessment logic through pilot deviation verification.

[0073] Furthermore, the pilot parks for energy-saving renovation include office parks, commercial parks, and science and education parks. The building area of ​​each type of park is greater than or equal to 5,000 square meters. The selected parks have energy-saving levels of Level 1 and Level 2. Each type of park includes one Level 1 park and one Level 2 park. The selection of different types of parks is to ensure the applicability of the pilot results to the scenarios.

[0074] Furthermore, the pilot deviation verification performs consistency checks on the collected actual data, using 3... Outliers are removed in principle. Actual time-of-use energy consumption data is reprocessed according to standardized rules to ensure consistency with the units and statistical dimensions of the predicted data, avoiding misjudgments due to differences in statistical standards. The preprocessed actual data is compared one-to-one with the time-of-use energy consumption prediction data and the comprehensive energy-saving potential assessment data. Energy-saving potential deviation and response speed deviation are calculated. Based on park type, park type coefficients are set: office park coefficient 1.0, commercial park coefficient 1.2, and science and education park coefficient 1.1. A judgment threshold is calculated using the formula: Dynamic Threshold = Basic Threshold × Type Coefficient. If both potential deviation and response speed deviation are less than or equal to the corresponding dynamic threshold, the verification is passed. If either potential deviation or response speed deviation is greater than the corresponding dynamic threshold, the model and indicator optimization process is triggered.

[0075] Furthermore, for the first prediction sub-model, the number of hidden layer neurons is adjusted by ±16 within the first preset number of neurons. For the second prediction sub-model, the tree depth is adjusted by ±1 layer within the first preset tree depth. After adjustment, the model is retrained. Based on the actual data after the pilot park renovation, the actual threshold of user behavior tags is calculated. Two to three new parks that did not participate in the pilot are selected. The energy-saving level of the new parks covers level one and level two to ensure that the optimization effect is applicable to different potential scenarios. Steps S1-S4 are executed again to calculate the energy-saving potential deviation and response speed deviation of the new parks. If both deviations are less than or equal to the dynamic threshold, the optimization is deemed effective. If the potential deviation or response speed deviation is greater than the corresponding dynamic threshold, the above parameter iteration and threshold correction steps are repeated until both deviations are less than or equal to the dynamic threshold.

[0076] Example 2:

[0077] like Figures 1-2 As shown in Example 2, a machine learning-based system for assessing the energy-saving potential of building parks is provided. The system includes: an energy characteristic data acquisition module, a data preprocessing module, a generation module, an energy consumption prediction model module, an energy-saving measures and scheduling linkage module, and a pilot verification and model optimization module.

[0078] Furthermore, the energy characteristic data acquisition module collects multi-dimensional energy characteristic data such as time-of-use electricity consumption, electrical parameters, meteorological data, and population flow data through hardware interface units, environmental data access units, and historical data storage units, and uses a time-series database to store historical transformation-related data.

[0079] Specifically, the energy characteristic data acquisition module is used to achieve standardized access and real-time acquisition of multi-dimensional energy characteristic data in the building park. The energy characteristic data acquisition module includes a hardware interface unit, an environmental data access unit, and a historical data storage unit. It drives the hardware interface unit to be configured with an RS485 / Ethernet dual-mode acquisition gateway, supports industrial protocols such as Modbus-RTU and DL / T645-2007, and is compatible with air conditioning controllers, smart meters, and charging pile management system equipment. It collects the time-of-use electricity consumption and electrical parameters of each building, floor, and equipment at a first preset interval. The environmental data access unit connects to the National Meteorological Information Center and the park's pedestrian flow statistics system via API, and collects meteorological data, pedestrian flow data, and equipment operating condition data at a second preset interval. The historical data storage unit uses a time-series database to store the technical solutions for energy-saving renovations that have been implemented, energy consumption comparison data for 3-24 months before and after the renovation, and renovation investment costs and energy-saving benefits data.

[0080] Furthermore, the data preprocessing module completes the cleaning and standardization of energy characteristic data through missing value processing, outlier removal, and data standardization units.

[0081] Specifically, the data preprocessing module is used to clean and standardize the energy characteristic data of the building park. The data preprocessing module includes a missing value processing unit, an outlier removal unit, and a data standardization unit. The missing value processing unit has a built-in time-series interpolation algorithm and a similar data segment matching engine. It performs interpolation completion on data with consecutive missing durations less than or equal to a first preset duration. For data with consecutive missing durations greater than the first preset duration, it automatically retrieves and replaces complete data segments from the same historical period that have similar meteorological conditions, weekday, and weekend attributes. The outlier removal unit is based on 3D... The principle is to construct an anomaly detection model. First, screen normal operating data where the equipment load is within 30%-80% of the rated capacity in the past 30 days. Then, calculate the mean and standard deviation of time-of-use electricity consumption, voltage, and current, and automatically mark and remove data exceeding ±3% of the mean. For outliers within a range, the data standardization unit is equipped with a distribution identification algorithm. It performs min-max normalization on uniformly distributed data and z-score standardization on normally distributed data, thus mapping the data uniformly to a preset numerical range.

[0082] Furthermore, the generation module uses pattern clustering, associated feature filtering, and a 3D label generation unit to mine time-of-use electricity patterns and construct and output a profile of electricity user behavior.

[0083] Specifically, the generation module is used to mine time-of-use electricity consumption patterns in building parks and construct electricity user behavior profiles. The generation module includes a pattern clustering unit, a correlation feature filtering unit, and a 3D label generation unit. The pattern clustering unit automatically determines the number of clusters through contour coefficient optimization, mining three types of time-of-use electricity consumption patterns: high-load continuous, intermittent fluctuating, and low-load stable. The correlation feature filtering unit mines the correlation between influencing factors and electricity consumption patterns based on correlation rule algorithms, automatically removes redundant features with correlation degrees less than a first preset correlation degree threshold, and outputs key cross features. The 3D label generation unit generates energy intensity labels according to electricity consumption thresholds, load fluctuation labels according to daily load fluctuation coefficient thresholds, and adjustable potential labels according to the proportion of load that can be reduced. Then, the key influencing factors are correlated with time-of-use electricity consumption feature vectors through a classification model to generate a complete electricity user behavior profile. Furthermore, the generation module mines time-of-use electricity consumption patterns and constructs and outputs electricity user behavior profiles through pattern clustering, correlation feature filtering, and 3D label generation units.

[0084] Specifically, the energy consumption prediction model module is used to construct and train a multi-source fusion prediction model, outputting time-of-use energy consumption prediction results for the building park. The energy consumption prediction model module includes a feature engineering unit, a hybrid prediction model construction unit, and a model training and optimization unit. The feature engineering unit extracts time-series and cross-features from the standardized dataset, calculates the correlation between features and electricity consumption through mutual information entropy, and removes redundant features with a correlation less than the second preset correlation threshold. The hybrid prediction model construction unit includes a first prediction sub-model and a second prediction sub-model. The first prediction sub-model is used to capture the long-term and short-term dependencies of energy consumption data, and the second prediction sub-model is used to correct time-series prediction biases, supporting the output of time-of-use energy consumption prediction results for the fourth, fifth, and sixth preset durations. The model training and optimization unit automatically adjusts the ratio of training and test sets according to the dataset sample size, uses a grid search algorithm to optimize model hyperparameters, and combines preset early stop rules to prevent model overfitting.

[0085] Furthermore, the energy-saving measures and dispatch linkage module achieves coordinated linkage between energy-saving measures generation and the dispatch platform through the measures matching unit and the virtual power plant linkage unit.

[0086] Specifically, the energy-saving measures and scheduling linkage module is used to generate energy-saving solutions for building parks and achieve collaborative linkage with the virtual power plant scheduling platform. The energy-saving measures and scheduling linkage module includes a measure matching unit and a virtual power plant linkage unit. The measure matching unit establishes a matching relationship database between user behavior profile tags and energy-saving levels. When the user behavior profile tag is high load continuity and the energy-saving level is level one, it automatically recommends high-yield energy-saving solutions. When the user behavior profile tag is high adjustability potential and the energy-saving level is level one, it calculates load transfer strategies based on peak-valley electricity price differences and adjustable load volume, and generates virtual power plant peak-valley arbitrage strategies. The virtual power plant linkage unit uploads adjustable load characteristic data and adjustable load volume data to the virtual power plant scheduling platform at a first preset interval. After receiving load adjustment instructions issued by the platform, it sends control instructions to the park equipment controller according to preset equipment adjustment rules.

[0087] Furthermore, the pilot verification and model optimization module completes pilot data collection, deviation verification, and model and label threshold optimization through pilot park management, actual data collection, deviation verification, and model index optimization units.

[0088] Specifically, the pilot verification and model optimization module is used to complete data collection, deviation verification, and model optimization for energy-saving retrofit pilot parks. This module includes a pilot park management unit, a unit for collecting actual data after pilot retrofits, a deviation verification unit, and a model and indicator optimization unit. The pilot park management unit supports selecting office parks, commercial parks, and science and education parks as pilot parks; each type of park includes first-level and second-level energy-saving potential parks. The unit for collecting actual data after pilot retrofits collects actual time-of-use energy consumption data, actual energy-saving benefit data, and virtual power plant linkage actual response data for twelve consecutive months after the energy-saving measures are implemented. The deviation verification unit performs consistency verification on the collected actual data, using 3D modeling and analysis. Outliers are removed in principle. Then, the preprocessed actual data and predicted data are compared one by one to calculate the energy-saving potential deviation and response speed deviation. The verification is then determined. The model and index optimization unit adjusts the number of hidden layer neurons in the first prediction sub-model and the tree depth in the second prediction sub-model. The user behavior label threshold is updated based on the actual pilot data. The evaluation process is re-executed in new parks that have not participated in the pilot to verify the optimization effect.

[0089] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for assessing the energy-saving potential of building parks based on machine learning, characterized in that, Includes the following steps: S1. Collect energy characteristic data of the building park, preprocess the energy characteristic data, and output a standardized dataset; S2. Based on standardized datasets, mine the time-of-use electricity patterns in building parks, construct a user behavior tagging system, and generate electricity user behavior profiles; S3. Perform feature engineering on the standardized dataset, build and train a multi-source fusion energy consumption prediction model, and output future time-of-use energy consumption prediction results; S4. Combine historical renovation data with prediction results to calculate the overall energy-saving potential and energy-saving level of the building park; S5. Generate energy-saving measures based on user behavior profiles and energy-saving assessment results, and link them with the virtual power plant dispatch platform; S6. Select pilot parks for energy-saving renovation, collect actual data after the pilot renovation, and optimize the energy consumption prediction model and energy-saving potential assessment logic through pilot deviation verification. The energy characteristic data includes basic energy consumption data, influencing factor data, and historical renovation data. The basic energy consumption data includes the time-of-use electricity consumption and electrical parameters of each building, floor, and equipment in the park. The sampling interval for the time-of-use electricity consumption is set to a first preset interval. The electrical parameters include voltage, current, and power factor. The equipment includes air conditioners, lighting, charging piles, and office equipment. The influencing factor data includes meteorological data, pedestrian flow data, equipment operating condition data, and electricity price data. The sampling interval for the meteorological data is set to a second preset interval. The historical renovation data includes the technical solutions for energy-saving renovations that have been implemented, energy consumption comparison data before and after the renovation for a preset period, renovation investment costs, and energy-saving benefits data. In step S2, the time-of-use electricity consumption pattern is mined using a clustering algorithm. The mined time-of-use electricity consumption patterns include high-load continuous type, intermittent fluctuation type, and low-load stable type. The association rule algorithm is combined with the data of influencing factors to mine the correlation between the data of influencing factors and the time-of-use electricity consumption pattern, and the key influencing factors with a correlation degree greater than the first preset correlation degree threshold are screened. The association rule algorithm is used to mine the potential correlation between influencing factors and electricity consumption patterns. In step S2, the user behavior tagging system is a three-dimensional tagging system, which includes energy intensity tags, load fluctuation tags, and adjustable potential tags. The energy intensity tags are divided into three levels: high, medium, and low, based on electricity consumption thresholds. The electricity consumption thresholds include a first electricity consumption threshold and a second electricity consumption threshold. The load fluctuation tags are divided into three levels: stable, medium, and fluctuating, based on daily load fluctuation coefficients. The fluctuation coefficient thresholds include a first fluctuation coefficient threshold and a second fluctuation coefficient threshold. The adjustable potential tags are divided into three levels: high, medium, and low, based on the proportion of load that can be reduced. The proportion thresholds include a first proportion threshold and a second proportion threshold. By inputting key influencing factors and time-of-use electricity consumption feature vectors into a classification model, a power user behavior profile is generated. The classification model is used to generate user behavior tags based on key features. The high load continuity tag corresponds to the energy intensity tag with applied electricity exceeding the first electricity consumption threshold, and the high adjustable potential tag corresponds to the adjustable potential tag with the proportion of load that can be reduced exceeding the first proportion threshold. In step S5, the energy-saving measure generation logic is determined based on the matching relationship between user behavior profile tags and energy-saving levels. When the user behavior profile tag is high load continuous type and the energy-saving level of the building park is determined to be Level 1, high-yield energy-saving solutions are selected and matched from historical renovation data. The high-yield energy-saving solutions are designed for high load scenarios and can achieve high-efficiency energy saving. When the user behavior profile tag is high adjustable potential type and the energy-saving level is determined to be Level 1, the load transfer strategy is calculated based on the peak-valley electricity price difference and adjustable load, and a virtual power plant peak-valley arbitrage strategy is generated.

2. The method for assessing the energy-saving potential of building parks based on machine learning according to claim 1, characterized in that, In step S1, the preprocessing includes missing value imputation, outlier removal, and data standardization. For time-series data with consecutive missing values ​​less than or equal to a first preset duration, missing value imputation uses temporal interpolation based on adjacent valid data. For time-series data with consecutive missing values ​​greater than the first preset duration, historical complete data segments with similar meteorological conditions, weekday and weekend attributes, and energy consumption patterns to the missing period are selected for replacement. Outlier removal is based on preset deviation rules, marking data exceeding the normal fluctuation range as outliers and removing them. These preset deviation rules are used to identify abnormal fluctuation data in the data sequence, employing a three-principle approach for outlier identification and removal. Data standardization uses a preset normalization algorithm to map non-electrical data to a preset numerical range, achieving scale uniformity for data of different dimensions. This preset normalization algorithm is used to eliminate dimensional differences between different types of data.

3. The method for assessing the energy-saving potential of building parks based on machine learning according to claim 2, characterized in that, In step S3, the feature engineering includes extracting time-series features and cross features from the standardized dataset, filtering key features through mutual information entropy, and removing redundant features with a correlation degree less than a second preset correlation degree threshold. The mutual information entropy is used to measure the degree of correlation between features and electricity consumption. The time-series features include the average electricity consumption over the previous third preset duration and the electricity consumption during the same period of the preset period. The third preset duration is the same as the preset period before and after the historical transformation data in step S1, ensuring the uniformity of the data period. In step S3, the construction and training of the energy prediction model includes the following sub-steps: S31. Construct a multi-source fusion hybrid prediction model, which includes a first prediction sub-model and a second prediction sub-model. The first prediction sub-model is used to capture the long-term and short-term dependencies of energy consumption data, and the second prediction sub-model is used to correct time-series prediction bias. S32. The selected key features are divided into training set and test set. The division ratio is executed according to the preset data division ratio. The preset data division ratio is determined based on the sample size of the dataset. The first prediction sub-model inputs time series features. The number of neurons in the hidden layer of the first prediction sub-model is set to the first preset range of the number of neurons. The second prediction model inputs cross features. The cross features include temperature and air conditioning load interaction features and human flow and equipment power consumption combination features. The tree depth is set to the first preset range of tree depth. S33. Using the mean absolute percentage error as the loss function, a preset early stopping rule is adopted to prevent the model from overfitting. The model hyperparameters are adjusted through a grid search algorithm, and a parameter optimization algorithm is used to find the optimal combination of hyperparameters. After training, the model's prediction mean absolute percentage error is less than or equal to a preset error threshold. The time-sharing energy consumption prediction results for the fourth, fifth, and sixth preset durations are output.

4. The method for assessing the energy-saving potential of building parks based on machine learning according to claim 3, characterized in that, In step S4, a multi-dimensional energy-saving potential assessment index system is constructed. The energy-saving potential assessment index system includes technical, economic and scheduling dimensions. Under the technical dimension, two indicators are set: equipment energy-saving potential and system loss reduction potential. Under the economic dimension, two indicators are set: energy-saving benefits and investment payback period. Under the scheduling dimension, two indicators are set: adjustable load and response speed. The response speed is within a preset response speed threshold. In step S4, the weights of the evaluation indicators are determined by a combined weighting method, which combines the analytic hierarchy process (AHP) and the entropy weighting method. The AHP is used to determine the subjective weights. Within the subjective weighting system, the weight proportions corresponding to technical potential, economic potential, and scheduling potential are set as the first preset weight proportion, the second preset weight proportion, and the third preset weight proportion, respectively. The entropy weighting method calculates the objective weights based on the dispersion of each indicator in the historical transformation data. The dispersion can intuitively reflect the degree of difference between different indicator data, ensuring that the calculation basis of the objective weights is consistent with the actual transformation scenario. The combined weights are calculated according to a weighting formula, the expression of which is: In the formula, The final combined weight of a certain evaluation index. The subjective weights of this evaluation indicator, determined using the analytic hierarchy process, represent the subjective weights. The objective weights for this evaluation indicator are determined using the entropy weight method. The first preset weighting coefficient, The second preset weighting coefficient, and =1, to ensure the rationality and normalization of weight calculation, and to avoid the combined weight deviating from the actual needs due to coefficient imbalance; The overall energy-saving potential of the building complex is expressed by the formula: Calculate, where, The combined weights for the i-th evaluation index are calculated using the weighting formula described above. Let i be the quantified value of the i-th evaluation indicator. The total number of evaluation indicators; The energy-saving level is determined according to a preset level classification rule. The preset level classification rule is that when the comprehensive energy-saving potential is greater than the first potential threshold, it is determined to be a level 1 energy-saving potential; when it is between the second potential threshold and the first potential threshold, it is determined to be a level 2 energy-saving potential; and when it is less than the second potential threshold, it is determined to be a level 3 energy-saving potential.

5. The method for assessing the energy-saving potential of building parks based on machine learning according to claim 4, characterized in that, In step S5, the linkage with the virtual power plant dispatching platform involves first uploading the adjustable load characteristic data recorded in the user behavior profile and the adjustable load data calculated in step S4 to the virtual power plant dispatching platform. The data upload frequency is consistent with the first preset interval. After receiving the load adjustment instruction issued by the platform, the parameters of the park's power equipment are dynamically adjusted according to the preset equipment adjustment rules. The preset equipment adjustment rules specify the specific adjustment methods for the operating parameters of different types of equipment to ensure that the normal power consumption experience of the park is guaranteed while responding to the load dispatching needs of the virtual power plant.

6. The method for assessing the energy-saving potential of building parks based on machine learning according to claim 5, characterized in that, In step S6, the energy-saving renovation pilot parks include office parks, commercial parks and science and education parks. The building area difference between each type of park is greater than or equal to 5,000 square meters. The energy-saving level of the selected parks includes level one and level two. Each type of park includes one level one park and one level two park. In step S6, the actual data collected after the pilot transformation is implemented according to the energy-saving measures generated in step S5. The actual data collected for twelve months after the transformation is carried out continuously, including actual time-of-use energy consumption data, actual energy-saving benefit data and actual response data of virtual power plant linkage. The collection frequency is matched with the monthly electricity bill settlement cycle.

7. The method for assessing the energy-saving potential of building parks based on machine learning according to claim 6, characterized in that, In step S6, the pilot deviation verification performs consistency checks on the collected actual data, uses the three principles to remove outliers, and reprocesses the actual time-of-use energy consumption data using standardized rules to ensure consistency with the dimensions and statistical dimensions of the predicted data, avoiding misjudgments due to differences in statistical caliber. The preprocessed actual data is compared one-to-one with the time-of-use energy consumption prediction data and the comprehensive energy-saving potential assessment data to calculate the energy-saving potential deviation and the response speed deviation. If both the energy-saving potential deviation and the response speed deviation are less than or equal to the corresponding dynamic threshold, the verification is deemed successful. If the energy-saving potential deviation or the response speed deviation is greater than the corresponding dynamic threshold, the model and index optimization process is triggered.

8. The method for assessing the energy-saving potential of building parks based on machine learning according to claim 7, characterized in that, In step S6, for the first prediction sub-model, the number of hidden layer neurons is adjusted by ±16 within the first preset number of neurons. For the second prediction sub-model, the tree depth is adjusted by ±1 layer within the first preset tree depth. After adjustment, the model is retrained. Based on the actual data after the pilot park renovation, the actual threshold of user behavior tags is calculated. Two to three new parks that did not participate in the pilot are selected. The energy-saving level of the new parks covers level one and level two to ensure that the optimization effect is applicable to different potential scenarios. Steps S1-S4 are executed again to calculate the energy-saving potential deviation and response speed deviation of the new parks. If both deviations are less than or equal to the dynamic threshold, the optimization is deemed effective. If the energy-saving potential deviation or response speed deviation is greater than the corresponding dynamic threshold, the above parameter iteration and threshold correction steps are repeated until both deviations are less than or equal to the dynamic threshold.

9. A machine learning-based system for assessing the energy-saving potential of building parks, based on any one of claims 1-8, characterized in that, These include: an energy characteristic data acquisition module, a data preprocessing module, a generation module, an energy consumption prediction model module, an energy-saving measures and scheduling linkage module, and a pilot verification and model optimization module; The energy characteristic data acquisition module collects multi-dimensional energy characteristic data through a hardware interface unit, an environmental data access unit, and a historical data storage unit, and uses a time-series database to store historical transformation-related data. The data preprocessing module completes the cleaning and standardization of energy characteristic data through missing value processing, outlier removal, and data standardization units. The generation module uses pattern clustering, correlation feature filtering, and three-dimensional label generation unit to mine time-of-use electricity patterns, construct and output electricity user behavior profiles; The energy consumption prediction model module outputs time-of-use energy consumption prediction results for different future durations through feature engineering, hybrid prediction model construction and training optimization units. The energy-saving measures and scheduling linkage module achieves coordinated linkage between energy-saving measures generation and the scheduling platform through the measure matching unit and the virtual power plant linkage unit; The pilot verification and model optimization module completes pilot data collection, deviation verification, and model and label threshold optimization through pilot park management, actual data collection, deviation verification, and model index optimization units.

Images