Energy consumption reconstruction effect measurement method for building energy saving reconstruction

Abstract

The application discloses a kind of energy consumption reconstruction effect measurement methods for building energy-saving reconstruction, the method obtains building historical energy consumption sample and carries out climate standardization processing, obtains standardization unit energy consumption as the supervision target value of model training;According to the two-layer grouping of building industry and energy supply system type, the energy consumption prediction model of each system type is trained respectively to construct knowledge base, and in training, through counterfactual sample and physical constraint, the model output meets the thermodynamic law;With industry and system type as joint key value, load model to infer current system energy consumption, and switch candidate system type in turn to carry out multi-scheme deduction, after climate reverse correction, the energy-saving amount of each reconstruction scheme is output;After cold and heat splitting and abnormal identification to actual monthly energy consumption, normal sample is backflowed, to update model in rolling. The present application only needs basic features and energy consumption data to model, and the prediction accuracy is high, supports multi-scheme quantitative comparison and selection and model self-evolution.

Description

Technical Field

[0001] This invention relates to the field of building energy conservation data analysis technology, and in particular to a method for calculating the energy consumption effect of building energy conservation retrofits. Background Technology

[0002] Building energy supply systems (heating / cooling) are the core focus of building energy-saving retrofits. Accurate prediction of the energy consumption of the heating and cooling sources after the retrofit for different schemes is a necessary prerequisite for quantifying energy savings and selecting the optimal retrofit scheme. Among existing building energy consumption assessment technologies, the closest to this invention are physical mechanism simulation technology and the index and empirical formula method. Their technical solutions, principles, and core problems are as follows: (I) Physical Mechanism Simulation Technology This technology is currently the recognized industry benchmark method in the field of building energy consumption assessment, and it plays a dominant role in new building design, green building certification, and major energy-saving renovation projects. Mainstream tools include DeST developed by Tsinghua University. Its core principle is based on the first and second laws of thermodynamics, Fourier's law of heat transfer, fluid mechanics, and other fundamental physical laws. It constructs a complete set of differential equations that include the entire process of one-dimensional / three-dimensional heat transfer in the building envelope, heat and moisture exchange of indoor air, solar radiation heat gain, energy conversion and transport of HVAC systems, and heat dissipation of terminal equipment. By integrating the equations at hourly (usually 1 hour or 15 minutes) time steps, it solves the dynamic thermal process of the building throughout the year, and finally outputs the simulation results of the building's heating / cooling energy consumption for hourly, daily, monthly, and annual periods.

[0003] This technology suffers from the following fundamental drawbacks that are directly related to the problem solved by this invention: 1) Incompatible with the renovation assessment needs of over 90% of existing buildings in my country: The core principle of mechanistic simulation technology dictates that the accuracy of its model depends entirely on the completeness and precision of the input physical parameters. To establish an accurate energy consumption model, complete as-built drawings (including architectural, structural, and HVAC specifications), precise thermal parameters (heat transfer coefficient, specific heat capacity, etc.) of each layer of the building envelope, model numbers and detailed performance parameters of all HVAC system equipment, hourly indoor occupancy / equipment / lighting load data, and hourly meteorological data are required. However, the vast majority of existing buildings in my country (especially those built before 2000) lack complete as-built drawings and have not undergone thermal performance testing of their building envelopes, making effective modeling impossible.

[0004] 2) Extremely low efficiency and extremely high application threshold: Due to the need for manual input of a large amount of physical details of buildings and systems, the modeling cycle for a single ordinary public building can take 3-7 days, and complex buildings can even take several weeks; a single annual energy consumption simulation takes several hours. Moreover, the modeling and simulation process must be completed by professional engineers with in-depth HVAC expertise and rich software operation experience. The learning cycle for mainstream simulation software usually takes 3-6 months, and front-line technicians in ordinary energy-saving renovation projects cannot use it independently, making it difficult to meet the batch assessment needs of large-scale energy-saving renovation of existing buildings in my country.

[0005] 3) Large prediction error after renovation of existing buildings: For renovation projects of existing buildings, even if modeling can be completed, the simulation results are usually between 15% and 25% from the actual energy consumption, and in some projects even more than 30%, because it is impossible to accurately obtain parameters such as the actual aging state of the building, the actual degree of decay of the thermal performance of the building envelope, the airtightness of the building, and the actual energy consumption habits of users. This makes it impossible to provide a reliable basis for the decision-making of renovation plans.

[0006] (II) Indicators and Empirical Formulas The core principle of this technology is to collect sample data of buildings that have completed energy consumption audits or energy-saving renovations, and use statistical methods to analyze the correlation between energy consumption per unit area of ​​buildings and basic characteristic parameters. Then, it establishes an energy consumption benchmark database by region, building type, and energy supply system, and summarizes empirical formulas for the average energy saving rate of various typical energy-saving renovation measures.

[0007] In practical engineering applications, assessors only need to obtain basic characteristic information of the building to be assessed (including building type, building area, year of construction, city, energy supply system type, etc.) and annual total energy consumption bill data. By querying the energy consumption benchmark value of the corresponding category or substituting into the preset empirical formula, they can quickly estimate the current energy consumption level of the building. At the same time, by comparing with the industry-standard experience value of renovation effect, they can make a preliminary judgment on the expected energy saving potential of different energy-saving measures.

[0008] This method, based on the statistical average law of the population, has the following fundamental drawbacks that are directly related to the problem solved by this invention: 1) The coupling relationships between features are unclear, failing to reflect the actual energy consumption impact mechanism: Indicators and empirical formulas treat each influencing factor as an independent variable, assuming their impact on energy consumption is a simple linear superposition, completely ignoring the complex nonlinear coupling effects between features in the building energy consumption system. For example, the form of the terminal unit affects energy consumption in conjunction with multiple factors such as local climate conditions and user energy consumption habits. Because these complex coupling relationships cannot be characterized, the prediction accuracy of empirical formulas drops sharply when dealing with buildings with different combinations of features.

[0009] 2) Low assessment accuracy, unable to support engineering decisions: The index method and empirical formula method reflect the average energy consumption level and average renovation effect of a certain type of building, completely ignoring the significant differences in the actual thermal performance of the building envelope and users' energy consumption habits of individual buildings. For different buildings of the same category, their actual energy consumption may differ by 2-3 times, and the renovation effect may also vary greatly. Therefore, the assessment results of this method are usually more than 30%, and for some special buildings, even more than 50%, making it impossible to use as a basis for accurate assessment of specific renovation scheme decisions.

[0010] 3) Requires a massive sample size to discover reliable patterns, and patterns for new scenarios cannot be quickly established: The core of this method is to extract statistical patterns from a large number of historical samples. Only when the sample size reaches a sufficient scale can the statistical results eliminate the influence of individual differences and random errors, reflecting the true trend of the overall situation. For any new building type, energy supply system form, or renovation measure, hundreds or even thousands of engineering samples are needed to establish a relatively reliable empirical formula. Summary of the Invention

[0011] Therefore, the purpose of this invention is to provide a method for measuring the energy consumption effect of building energy-saving renovation, which solves the technical problems of existing physical mechanism simulation technology being too dependent on complete input data and unable to adapt to existing buildings, as well as the low accuracy of the index empirical formula method and its inability to support quantitative comparison of multiple schemes. By integrating climate standardization processing, grouping modeling according to business type and system type, counterfactual learning and physical constraints, it can achieve rapid and accurate prediction of energy consumption after renovation and quantitative evaluation of multiple schemes under the condition of limited data.

[0012] To achieve the above objectives, the present invention provides a method for calculating the energy consumption retrofit effect of building energy-saving renovation, comprising: S1. Obtain historical energy consumption samples of different building types, standardize the historical energy consumption samples, and calculate the standardized unit energy consumption. S2. The standardized historical energy consumption samples are divided into multiple subsets according to building type. Within each subset, they are further grouped according to energy supply system type. In each group, the standardized unit energy consumption is used as the supervision target to train energy consumption prediction models for each energy supply system type. The energy consumption prediction models are physically constrained by a custom loss function so that the output energy consumption prediction values ​​satisfy physical laws while fitting the real data. The trained energy consumption prediction models are associated with the corresponding energy supply system type and building type label to form a building cold and heat source knowledge base. S3. Based on the building type label and current energy supply system type of the building to be evaluated, load the corresponding energy consumption prediction model from the building cold and heat source knowledge base to perform reasoning and obtain the predicted unit energy consumption under the current energy supply system type. S4. Reload the energy consumption prediction models corresponding to other candidate energy supply system types under the same business type in the building cold and heat source knowledge base, and perform reasoning to obtain the unit energy consumption of candidate schemes for other candidate energy supply system types; compare the unit energy consumption of the candidate schemes with the predicted unit energy consumption, and calculate and output the energy saving and energy saving rate of each candidate scheme.

[0013] Further preferably, the building type is used to identify the functional use of the building, and includes at least office, school, hospital, and stadium types. The historical energy consumption sample includes historical energy consumption data of building heating and cooling sources and meteorological data. The historical energy consumption samples are standardized, and the standardized unit energy consumption is calculated, including: Typical meteorological years are extracted based on the meteorological data, and the start and end dates of the standard season are determined based on the number of heating days or cooling days under the typical meteorological year. Calculate the seasonal ratio coefficient of the actual cooling or heating season start and end dates of each sample relative to the standard season start and end dates, based on the actual cooling or heating season start and end dates of each sample. The original energy consumption is normalized annually based on the climate correction factor corresponding to the year. Divide the total energy consumption by the energy consumption area to obtain the energy consumption per unit area, and then divide it by the climate correction factor and the seasonal proportionality factor to obtain the standardized unit energy consumption. The standardization process also includes standardizing the heating temperature and heating duration, and for buildings to be evaluated in non-training benchmark cities, calculating city correction factors based on typical meteorological years in the target city.

[0014] More preferably, the energy supply system type includes at least air source heat pump system, ground source heat pump system, water source heat pump system, multi-split system, chiller system, gas boiler system and direct-fired system; the building heat source knowledge base is stored separately as a heating model library and a cooling model library.

[0015] Furthermore, in S2, the step of training energy consumption prediction models for each type of energy supply system in each group, using the standardized unit energy consumption as the supervision target, includes the following training process: For each real sample, a counterfactual sample is generated, and the real sample and the counterfactual sample are input into the energy consumption prediction model. When the samples are grouped and trained according to the type of energy supply system, the XGBoost regression model or the neural network model is used for training the system type with a sample size that reaches the preset threshold, and the K-nearest neighbor regression or fixed benchmark regression is used for modeling the system type with a sample size that does not reach the preset threshold.

[0016] Furthermore, the counterfactual samples are generated in the following way: keeping the main building conditions unchanged, starting from each real training sample, the end form switching, building age progression and / or insulation status switching are performed respectively to construct counterfactual samples that only change a single physical feature. The counterfactual samples share the same main building conditions and the same building type with the corresponding real samples. The counterfactual samples and real samples together constitute the training set. Physical prior constraints are embedded into the model parameter optimization process through a custom loss function. While fitting the real data, the output of the counterfactual samples satisfies the preset physical law boundary.

[0017] Furthermore, the physical constraints include: Constraint 1: Under the same conditions, the predicted energy consumption of the counterfactual sample after the building age is progressive shall not be lower than the predicted energy consumption of the original true sample. Constraint 2: Under the same conditions, the energy consumption prediction value of the counterfactual sample after the insulation state is switched from "no insulation" to "with insulation" is not higher than the energy consumption prediction value of the original true sample. Constraint 3: The predicted energy consumption ratio between different terminal types conforms to the preset engineering experience boundary. Furthermore, the specific method for applying the physical constraints is as follows: For the XGBoost tree model, the physical constraints are embedded into the training process by using a custom objective function and monotonic constraints. For the XGBoost tree model, by constructing twin samples and introducing a prior loss term, the model can minimize regression error while avoiding violation of physical constraints. The custom loss function and the prior loss term participate in gradient calculation together with the regression loss of standardized unit energy consumption during backpropagation, driving the model parameters to optimize in the direction of fitting the real data and satisfying physical laws.

[0018] Furthermore, it also includes S5, which splits the actual monthly energy consumption data submitted by users into cold and hot categories to obtain cooling energy consumption and heating energy consumption, performs anomaly identification on the monthly energy consumption curve, and archives normal samples after anomaly identification into the database with their building type tags, and periodically merges new normal samples with the original training samples to update each energy consumption prediction model.

[0019] Furthermore, the process of splitting the actual monthly energy consumption data submitted by users into hot and cold components includes: decomposing the monthly comprehensive energy consumption data submitted by users into basic energy consumption, cooling energy consumption, and heating energy consumption; the specific splitting method is as follows: combining the monthly total energy consumption curve, the monthly heating and cooling days under typical meteorological years, the cooling and heating season intervals, and building operation information, and for school-related businesses, adjusting the basic energy consumption coefficients according to the operating characteristics of winter and summer vacations; after splitting, the total annual cooling and heating capacity and the corresponding energy consumption index per unit area are obtained, and stored separately according to heating meters and cooling meters. The splitting results and their corresponding building type tags are used for subsequent rolling update training after anomaly identification and screening.

[0020] Furthermore, the anomaly identification process for the monthly energy consumption curve includes the following steps: training a classification model using real monthly energy consumption data and synthetic anomaly data; extracting peak-valley features, fluctuation features, slope features, continuous identical value features, bimodal features, and similarity features with the standard trend from the monthly energy consumption curve to form an anomaly identification feature set; simultaneously calling the energy consumption prediction model to calculate the predicted cooling and heating energy consumption of the building under the current conditions, comparing it with the energy consumption breakdown results, and identifying abnormal energy consumption inputs; using the classification model to output anomaly probabilities and anomaly labels, saving the anomaly identification results and energy consumption breakdown results together, and prioritizing the removal of samples marked as anomalies in subsequent rolling updates, only incorporating normal samples into the training set.

[0021] The energy consumption retrofitting effect calculation method for building energy-saving renovation disclosed in this application has at least the following advantages compared with existing technologies: This invention employs personalized modeling and prediction based on the specific characteristics of each building to be evaluated, fully considering individual differences. Simultaneously, through energy consumption data standardization correction, physical constraints, and counterfactual learning, the prediction accuracy is significantly improved. Actual engineering verification shows that the prediction error of this invention can be controlled within 10%, providing a reliable basis for decision-making on specific renovation plans.

[0022] This invention utilizes counterfactual learning technology to generate three types of counterfactual twin samples for each real sample (end-stage form switching, building age progression, and building envelope insulation status switching), effectively expanding the training sample size. Simultaneously, by imposing physical prior constraints on the model through a custom objective function, the model can learn general building thermophysical laws based on a limited number of real samples, thus achieving high-accuracy predictions even with a small sample size. Furthermore, this invention establishes a data rolling update mechanism; as new samples accumulate, the model can automatically perform incremental training, quickly adapting to new building types, energy supply system forms, and renovation measures.

[0023] This invention constructs a multi-system energy consumption knowledge base, groups samples according to energy supply system type, and matches a corresponding prediction model to each energy supply system. During evaluation, the system can simultaneously output the predicted energy consumption of the building under standard operating conditions under the current energy supply system and multiple target modified energy supply systems. It automatically calculates the energy savings and energy efficiency of each modification scheme and quantitatively ranks multiple modification schemes to help owners select the optimal scheme. This method can not only predict the modification effect of parameter adjustments within the same type of system, but also accurately predict the energy consumption changes when switching between different energy supply systems.

[0024] This invention establishes a complete closed-loop self-evolving mechanism for data acquisition, cleaning, training, prediction, and feedback. Through monthly hot / cold energy consumption separation technology and abnormal energy consumption identification, the real energy consumption data submitted by users is automatically cleaned and stored in the database. A rolling update training mechanism periodically incorporates newly added normal samples into the training set to retrain the model, achieving hot updates without service interruption. As system runtime increases and sample size accumulates, the energy consumption prediction accuracy of the modified model will continuously improve, achieving long-term self-evolution for real-world business scenarios. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the energy consumption retrofitting effect calculation method for building energy conservation retrofitting provided by the present invention.

[0026] Figure 2 A schematic diagram of the structure of the energy consumption knowledge base provided by this invention.

[0027] Figure 3 This invention provides a detailed flowchart for calculating energy consumption in building energy-saving renovations. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] like Figure 1 As shown, one embodiment of the present invention provides a method for calculating the energy consumption retrofit effect of building energy-saving renovation, which includes the following steps: S1. Obtain historical energy consumption samples of different building types, standardize the historical energy consumption samples, and calculate the standardized unit energy consumption; the building type is used to identify the functional use type of the building, including at least office type, school type, hospital type, and stadium type.

[0030] The historical energy consumption samples include historical energy consumption sample data of building heating and cooling sources and meteorological data; the historical energy consumption sample data of building heating and cooling sources can also be analyzed monthly according to the natural year to form monthly energy consumption data.

[0031] The historical energy consumption sample data is standardized, and the standardized unit energy consumption is calculated, including: Based on the meteorological data, a typical meteorological year is extracted, and the start and end dates of the standard season are determined according to the number of heating or cooling days under the typical meteorological year; specifically including: Calculate Heating Degree Days (HDD) and Cooling Degree Days (CDD) (monthly / annual). Set heating reference temperature (Commonly used 18℃), Refrigeration reference temperature (Usually 26℃).

[0032] : Average daily outdoor temperature on day i; : Heating degree days; the higher the value, the colder the winter and the higher the heating demand. Cooling degree days: The higher the value, the hotter the summer and the higher the cooling demand.

[0033] Sum the results over the statistical period to obtain the actual number of heating days for the target period. Actual number of cooling days in the target period (Weather data for the current year to be corrected); simultaneously read the reference heating degree days for the standard meteorological year. Reference number of cooling days in a standard meteorological year (Typical meteorological year TMY); Calculate the seasonal ratio coefficient of the actual cooling or heating season start and end dates of each sample relative to the standard season start and end dates, based on the actual cooling or heating season start and end dates of each sample. The original energy consumption is normalized annually based on the climate correction factor corresponding to the year. Divide the total energy consumption by the energy consumption area to obtain the energy consumption per unit area, and then divide it by the climate correction factor and the seasonal proportionality factor to obtain the standardized unit energy consumption. Because the training set samples will have different heating durations, some are data from 2026 and some are data from 2025, but the climate in 2026 and 2025 is different. To train under the same standard, we need to use heating / cooling degree-days, because we believe that the energy consumption gap is proportional to the number of degree-days.

[0034] Assuming a building in City A has an energy consumption of T and an area of ​​A during the 2025 heating season, the energy consumption per unit area is u = T / A. The building's heating season runs from October 30th to April 1st, while City A's standard heating season is from November 15th to March 15th of the following year. An example of standardizing the energy consumption per unit area is as follows: Actual operating period: October 30 to April 1. The cumulative number of heating days within the actual heating season window is obtained by summing up the actual weather data within this period. ; Standard seasonal window: November 15 to March 15. The cumulative value of heating degree days within the standard seasonal window is obtained by summing the values ​​of typical meteorological years within this window. ; Seasonal proportion coefficient ; Example values: The actual heating period is from October 30th to April 1st, during which time the total ; The standard seasonal window is from November 15 to March 15, during a typical meteorological year. ; Seasonal proportion coefficient ; Meaning: The actual number of heating windows exceeded the standard number (October 30th to November 14th, March 16th to April 1st), and HDD and energy consumption also occurred during this period; it needs to be addressed through... The energy consumption of the actual window is converted to the degree-day level of the standard window.

[0035] The raw energy consumption is normalized annually based on the climate correction factor corresponding to the year, specifically including: the actual number of heating days in the target period. : The actual number of heating days corresponding to this heating season (the actual HDD within the standard season window); Reference heating degree days for a standard meteorological year.

[0036] Annual climate correction coefficient ; Example values: 2025-2026 winter standard season window realities Typical meteorological year reference After normalization ; Meaning: This winter is colder than a typical meteorological year, and the number of days in the same period is greater.

[0037] Final standardized energy consumption per unit area (sample feature labels used for model training): Substitute the numbers from the example above: ; in, Here, A represents the energy consumption per unit area, and A represents the energy consumption area. The standardization process also includes standardizing and correcting heating temperature and duration according to current national standards, and calculating urban correction factors for buildings to be evaluated in non-training benchmark cities based on typical meteorological years in the target city. Current national standards include, for example: GB 50736-2012 "Code for Design of Heating, Ventilation and Air Conditioning of Civil Buildings"; GB 55015-2021 "General Code for Energy Conservation and Renewable Energy Utilization in Buildings," etc.

[0038] It should be further clarified that, for example, the training dataset consists entirely of buildings in City A; the city correction factor is only used during the inference phase when evaluating buildings in other cities (Cities B, C, D, etc.); the city correction factor is used during training with local samples from City A. No cross-city corrections will be made.

[0039] All the correction formulas can be expressed as: This serves as the benchmark constant for the standard heating season in a typical meteorological year in City A. For cities outside the city to be evaluated, the number of heating degree days for their typical meteorological year is calculated using the same standard seasonal time window as City A.

[0040] S2. The standardized historical energy consumption samples are divided into multiple subsets according to building type. Within each subset, they are further grouped according to energy supply system type. In each group, the standardized unit energy consumption is used as the supervision target to train energy consumption prediction models for each energy supply system type. The energy consumption prediction models are physically constrained by a custom loss function so that the output energy consumption prediction values ​​satisfy physical laws while fitting the real data. The trained energy consumption prediction models are associated with the corresponding energy supply system type and building type label to form a building cold and heat source knowledge base. The energy supply system types include at least air source heat pump systems, ground source heat pump systems, water source heat pump systems, multi-split systems, chiller systems, gas boiler systems, and direct-fired systems; the building heat and cold source knowledge base is stored separately as a heating model library and a cooling model library.

[0041] In each group, using the standardized unit energy consumption as the supervision target, energy consumption prediction models for each type of energy supply system are trained, including the following training process: For each real sample, a counterfactual sample is generated, and the real sample and the counterfactual sample are input into the energy consumption prediction model. When the samples are grouped and trained according to the type of energy supply system, the XGBoost regression model or the neural network model is used for training the system type with a sample size that reaches the preset threshold, and the K-nearest neighbor regression or fixed benchmark regression is used for modeling the system type with a sample size that does not reach the preset threshold.

[0042] The counterfactual samples are generated in the following way: keeping the main building conditions unchanged, starting from each real training sample, the end form switching, building age progression and / or insulation status switching are performed respectively to construct counterfactual samples that only change a single physical feature. The counterfactual samples share the same main building conditions and the same building type with the corresponding real samples. The counterfactual samples and real samples together constitute the training set. Physical prior constraints are embedded into the model parameter optimization process through a custom loss function. While fitting the real data, the output of the counterfactual samples satisfies the preset physical law boundary.

[0043] The physical constraints include: Constraint 1: Under the same conditions, the predicted energy consumption of the counterfactual sample after the building age is progressive shall not be lower than the predicted energy consumption of the original true sample. Constraint 2: Under the same conditions, the energy consumption prediction value of the counterfactual sample after the insulation state is switched from "no insulation" to "with insulation" is not higher than the energy consumption prediction value of the original true sample. Constraint 3: The predicted energy consumption ratio between different terminal types conforms to a preset engineering experience boundary; the specific method for applying the physical constraint is as follows: For the XGBoost tree model, the physical constraints are embedded into the training process by using a custom objective function and monotonic constraints. For the XGBoost tree model, by constructing twin samples and introducing a prior loss term, the model can minimize regression error while avoiding violation of physical constraints. The custom loss function and the prior loss term participate in gradient calculation together with the regression loss of standardized unit energy consumption during backpropagation, driving the model parameters to optimize in the direction of fitting the real data and satisfying physical laws.

[0044] By unifying the above three constraints into the existing prior loss framework, the total prior loss is the sum of the losses of the three constraints.

[0045] 1) Constraint-based loss (building age) For each original sample Generate counterfactual twin samples Only the building age has been increased; all other features have been completely copied.

[0046] Constraint 1 Loss Value ; when Constraint satisfaction ; If the predicted energy consumption decreases as the age increases, the loss activation will result in a penalty. Used to avoid floating-point noise This represents the original number of samples.

[0047] 2) Constraint on two losses (increase external wall insulation) Generate counterfactual twin samples Only the insulation status was changed from no insulation to insulation; all other features were copied from the original sample.

[0048] Constraint 2 Loss Value ; when If the energy consumption increases after increasing insulation, a penalty will be triggered. The number of counterfactual twin samples.

[0049] 3) Constraints on three losses (end-point form ratio engineering boundary) Terminal A: Counterfactual sample; : Only the terminal B is replaced, and all other features remain unchanged; The ratio of predicted energy consumption at terminal B to terminal A ; Constraint Three Loss Values ; The ratio exceeds the upper limit of the engineering scope. / Lower World A loss will occur; if the ratio falls within the interval, the loss is 0. This represents the total number of sample pairs.

[0050] Total loss: in, Standardized unit energy consumption regression loss (MSE, from real labeled samples). : These are the loss weight hyperparameters for the three physical constraints.

[0051] S3. Based on the building type tag and current energy supply system type of the building to be evaluated, the system loads the corresponding energy consumption prediction model from the building's heating and cooling source knowledge base for inference, and obtains the predicted unit energy consumption under the current energy supply system type. When an external client initiates a prediction request, the system receives the building's original JSON data through the interface service. This data typically includes the building name, building year, city, building area, insulation status, number of people, heating and cooling source system type, terminal type, start and end dates of the cooling / heating season, daily operating hours, energy price, and pre-renovation cost information, and is standardized as input. For non-training benchmark cities (City A), the system will also calculate a city correction factor based on the typical meteorological year of the target city, correcting the standard prediction results to the results under the conditions of the target city.

[0052] S4. Reload the energy consumption prediction models corresponding to other candidate energy supply system types under the same business type in the building cold and heat source knowledge base, and perform reasoning to obtain the unit energy consumption of candidate schemes for other candidate energy supply system types; compare the unit energy consumption of the candidate schemes with the predicted unit energy consumption, and calculate and output the energy saving and energy saving rate of each candidate scheme; Based on the main conditions of the same building, inference is performed sequentially on multiple candidate system models to simulate the energy consumption performance under different renovation schemes. For example, for heating, if the existing system in a certain park is a gas boiler, then during inference, the expected energy consumption model of all heating systems other than the gas boiler should be constructed. If a corresponding model already exists for the target system, the model is used directly for prediction; for systems lacking a model, an alternative estimate is calculated based on the results of benchmark systems such as air source systems and preset conversion coefficients, thereby avoiding the loss of solutions due to insufficient training samples. Ultimately, the system generates a set of predicted energy consumption and costs for the "current system" and "various alternative systems," providing users with a basis for horizontal comparison before and after the transformation.

[0053] It also includes S5, which splits the actual monthly energy consumption data submitted by users into cold and hot categories to obtain cooling energy consumption and heating energy consumption, performs anomaly identification on the monthly energy consumption curve, and archives normal samples after anomaly identification into the database with their building type tags, and periodically merges new normal samples with the original training samples to update each energy consumption prediction model.

[0054] The process of splitting the actual monthly energy consumption data submitted by users into cold and hot components includes: decomposing the monthly comprehensive energy consumption data submitted by users into basic energy consumption, cooling energy consumption, and heating energy consumption; the specific splitting method is as follows: combining the monthly total energy consumption curve, the monthly heating and cooling days under typical meteorological years, the cooling and heating season intervals, and building operation information, and for school-type businesses, adjusting the basic energy consumption coefficient according to the operating characteristics of winter and summer vacations; after splitting, the total annual cooling and heating capacity and the corresponding energy consumption index per unit area are obtained, and stored separately according to heating meters and cooling meters. The splitting results and their corresponding building type tags are used for subsequent rolling update training after anomaly identification and screening.

[0055] The anomaly identification process for the monthly energy consumption curve includes the following steps: training a classification model using real monthly energy consumption data and synthetic anomaly data; extracting peak-valley features, fluctuation features, slope features, continuous identical value features, bimodal features, and similarity features with the standard trend from the monthly energy consumption curve to form an anomaly identification feature set; simultaneously calling the energy consumption prediction model to calculate the predicted cooling and heating energy consumption of the building under the current conditions, comparing it with the energy consumption breakdown results, and identifying abnormal energy consumption inputs; using the classification model to output anomaly probabilities and anomaly labels, saving the anomaly identification results and energy consumption breakdown results together, and prioritizing the removal of samples marked as anomalies in subsequent rolling updates, only incorporating normal samples into the training set.

[0056] The following six features are constructed from the monthly energy consumption curve: whether there are consecutive months with completely identical energy consumption; peak-to-valley ratio (highest monthly energy consumption ÷ lowest monthly energy consumption); standard deviation of monthly energy consumption; slope of annual energy consumption trend; cosine similarity with standard energy consumption trend; and total deviation from standard trend.

[0057] Training an anomaly detection and classification model includes the following steps: Prepare the sample set: It includes real historical monthly energy consumption samples, and at the same time, it artificially generates / samples synthetic abnormal data (simulating fault modes such as instrument jamming, missing data, jumps, and trend reversals), and labels them as normal / abnormal; thus forming the training dataset.

[0058] 1. For each sample, extract the above 6-dimensional features to construct an anomaly identification feature set;

[0059] 2. Train a binary classification model (random forest); the model input is a 6-dimensional feature set, and the output is: anomaly probability p (0~1) and binary anomaly label (normal / abnormal).

[0060] The anomaly identification results (anomaly probability, anomaly label) and energy consumption breakdown results are bound together and stored in the database. During subsequent model rolling updates and iterations: samples marked as anomalies are preferentially removed; only samples marked as normal are included in the training set, and anomaly samples are not included in the training to avoid contaminating the energy consumption prediction model with dirty data.

[0061] like Figure 3 As shown, step one: collect raw data, including the following four types of basic data: Basic building information: business type, area, age, building envelope, insulation type, and terminal type.

[0062] Energy consumption data: Historical monthly energy consumption of water, electricity, and gas.

[0063] Environmental data: monthly HDD heating days, CDD cooling days, and historical outdoor temperatures.

[0064] Equipment parameters: Existing heat / cold source system type, equipment COP, and system annual degradation rate.

[0065] Step Two: Systematic breakdown of building energy consumption, including: Breaking down by energy type: water consumption, electricity consumption, and gas consumption; Based on energy consumption attributes: basic energy consumption, heating energy consumption, and cooling energy consumption.

[0066] Determining the type of construction industry includes: For ordinary buildings: basic energy consumption is calculated based on a constant value throughout the year; School buildings: Introduce winter and summer vacation / winter and summer vacation activity intensity coefficients to correct basic energy consumption during holidays; Due to the unique nature of the school sector (students' winter and summer vacations), the basic energy consumption under the school sector fluctuates accordingly. Therefore, according to: in, It refers to the intensity of activities during winter break. ; It's about the intensity of summer activities. ; Basic energy consumption for winter break Daily basic energy consumption This is the basic energy consumption for summer vacation; Regarding the varying heating and cooling energy consumption each month, we believe that the amount of energy consumed is related to the number of degree-days in the local area that year, that is: ; Based on this, monthly energy consumption can be expressed as: ; Here, Factor represents the actual monthly holiday activity intensity coefficient; This represents some uncertain amounts each month.

[0067] Establish a correlation model: heating energy consumption ≈ HDD, cooling energy consumption ≈ CDD, and construct a monthly energy consumption calculation formula.

[0068] Step 3: Calculate the starting point (baseline value) before the modification. Based on historical energy consumption data uploaded by users, the initial energy consumption baseline before the renovation is determined.

[0069] Step 4: Energy consumption standardization correction (correction based on national standards), including: Heating temperature correction: Based on an indoor temperature of 20℃ and a standard outdoor temperature of -0.6℃, adjust the heating energy consumption over the years; Heating duration adjustment: Standardize heating duration to eliminate the impact of yearly differences.

[0070] Step 5: Construct a knowledge base for the energy consumption of the heat source system (adaptive sample size modeling). Figure 2 As shown, the specific process includes the following: Categorization and organization: The samples are categorized into 12 mainstream system types, including air source heat pumps, ground source heat pumps, direct-fired engines, and gas boilers. Data preprocessing: feature derivation and one-hot encoding of category features.

[0071] Sample size determination & adaptive model selection: Sample size > 4 → Use XGBoost regression model for training; 2 ≤ Sample size ≤ 4 → Use the K-Nearest Neighbors (KNN) model; Sample size = 1 → Fix the energy consumption per unit area of ​​this sample as the baseline value. Sample size = 0 → Skip system modeling; Complete the import of various heat source system models into the database.

[0072] Step Six: Determine if the sample data is sufficient. If sufficient, proceed to Step Seven. If insufficient, execute "Small Sample System Energy Consumption Estimation": using the air source heat pump energy consumption as the baseline value air_base_val. Small sample system energy consumption estimation is a method to provide an estimated energy consumption value after modification when there is insufficient training data for the target energy supply system, making it impossible to train a reliable prediction model. Based on the target system type, COP, and pipe loss coefficient, the energy consumption per unit area is calculated according to the energy conservation formula, and the final output is the estimated system energy consumption value. COP and pipe loss coefficient can be set empirically.

[0073] Step 7: Simulate the renovation plan and calculate the starting point of energy consumption after the renovation. Enter the energy-saving renovation plan and the target replacement system type.

[0074] The corresponding system prediction model is invoked to calculate the energy consumption of heating, cooling, lighting, water supply, and photovoltaic power in parallel.

[0075] By integrating multi-dimensional energy consumption indicators, the overall energy consumption baseline after the renovation is obtained.

[0076] Step 8: Calculate the interannual trend of energy consumption (equipment aging and degradation) Read the system's inherent annual decay rate α.

[0077] Based on the coefficient of performance (COP) decay formula (the ratio of cooling / heating output energy to input energy), the annual energy consumption before and after the renovation is estimated.

[0078] When calculating the annual upward trend of cooling / heating mode before / after the renovation, the impact of temperature and major events on overall energy consumption is not considered; only the marginal benefits caused by unit aging are considered.

[0079] Therefore, daily energy consumption has the following relationship: The annual energy consumption can then be expressed as: Given the daily load, assuming the COP remains stable throughout the year, the energy consumption in year i can be expressed as: Since the impact of temperature and major events on overall energy consumption is not considered, and assuming that the total energy consumption in year k equals the baseline energy consumption without modification, the COP in year i can be expressed as: Performance coefficient in year i Then the energy consumption in the kth year can be expressed as: The annual degradation rate of the system is shown in Table 1 below, which lists the annual degradation rates of common energy-consuming systems: Table 1: Annual Degradation Rate of Common Energy Consumption Systems interannual decay rate 0.8%~1.2% 1%-3% 0.5%~5% 0.5%~1.0% 0.5%~5% 0.8%~1.2% 0.1~1.2% 0 0.5%~1.0% 0 0.8%~1.2% Generate a long-term energy consumption change curve.

[0080] Step Nine: Energy Saving and Long-Term Benefit Calculation Compare the energy consumption baselines before and after the renovation, plus the annual trend of change.

[0081] Calculate the cumulative energy savings corresponding to the region below the energy consumption change curve.

[0082] Output the long-term benefit assessment results of energy-saving renovation.

[0083] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for calculating the energy consumption effect of building energy-saving renovation, characterized in that, include: S1. Obtain historical energy consumption samples of different building types, standardize the historical energy consumption samples, and calculate the standardized unit energy consumption. S2. The standardized historical energy consumption samples are divided into multiple subsets according to building type. Within each subset, they are further grouped according to energy supply system type. In each group, the standardized unit energy consumption is used as the supervision target to train energy consumption prediction models for each energy supply system type. The energy consumption prediction models are physically constrained by a custom loss function so that the output energy consumption prediction values ​​satisfy physical laws while fitting the real data. The trained energy consumption prediction models are associated with the corresponding energy supply system type and building type label to form a building cold and heat source knowledge base. S3. Based on the building type label and current energy supply system type of the building to be evaluated, load the corresponding energy consumption prediction model from the building cold and heat source knowledge base to perform reasoning and obtain the predicted unit energy consumption under the current energy supply system type. S4. Reload the energy consumption prediction models corresponding to other candidate energy supply system types under the same business type in the building cold and heat source knowledge base, and perform reasoning to obtain the unit energy consumption of candidate schemes for other candidate energy supply system types; compare the unit energy consumption of the candidate schemes with the predicted unit energy consumption, and calculate and output the energy saving and energy saving rate of each candidate scheme.

2. The method for calculating the energy consumption retrofit effect of building energy-saving renovation according to claim 1, characterized in that, The building type is used to identify the functional use of a building, and includes at least office, school, hospital, and stadium types; the historical energy consumption sample includes historical energy consumption sample data of building heating and cooling sources and meteorological data; The historical energy consumption samples are standardized, and the standardized unit energy consumption is calculated, including: Typical meteorological years are extracted based on the meteorological data, and the start and end dates of the standard season are determined based on the number of heating days or cooling days under the typical meteorological year. Calculate the seasonal ratio coefficient of the actual cooling or heating season start and end dates of each sample relative to the standard season start and end dates, based on the actual cooling or heating season start and end dates of each sample. The original energy consumption is normalized annually based on the climate correction factor corresponding to the year. Divide the total energy consumption by the energy consumption area to obtain the energy consumption per unit area, and then divide it by the climate correction factor and the seasonal proportionality factor to obtain the standardized unit energy consumption. The standardization process also includes standardizing and correcting heating temperature and heating duration, and for buildings to be evaluated in non-training benchmark cities, calculating urban correction factors based on typical meteorological years in the target city, and using urban correction factors to correct non-training benchmark cities.

3. The method for calculating the energy consumption retrofit effect of building energy-saving renovation according to claim 1, characterized in that, The energy supply system types include at least air source heat pump systems, ground source heat pump systems, water source heat pump systems, multi-split systems, chiller systems, gas boiler systems, and direct-fired systems; the building heat and cold source knowledge base is stored separately as a heating model library and a cooling model library.

4. The method for calculating the energy consumption retrofit effect of building energy-saving renovation according to claim 1, characterized in that, In S2, the step of training energy consumption prediction models for each type of energy supply system in each group, using the standardized unit energy consumption as the supervision target, includes the following training process: For each real sample, a counterfactual sample is generated, and the real sample and the counterfactual sample are input into the energy consumption prediction model. When the samples are grouped and trained according to the type of energy supply system, the XGBoost regression model or the neural network model is used for training the system type with a sample size that reaches the preset threshold, and the K-nearest neighbor regression or fixed benchmark regression is used for modeling the system type with a sample size that does not reach the preset threshold.

5. The method for calculating the energy consumption retrofit effect of building energy-saving renovation according to claim 4, characterized in that, The counterfactual samples are generated in the following way: keeping the main building conditions unchanged, starting from each real training sample, the end form switching, building age progression and / or insulation status switching are performed respectively to construct counterfactual samples that only change a single physical feature. The counterfactual samples share the same main building conditions and the same building type with the corresponding real samples. The counterfactual samples and real samples together constitute the training set. Physical prior constraints are embedded into the model parameter optimization process through a custom loss function. While fitting the real data, the output of the counterfactual samples satisfies the preset physical law boundary.

6. The method for calculating the energy consumption retrofit effect of building energy-saving renovation according to claim 5, characterized in that, The physical constraints include: Constraint 1: Under the same conditions, the predicted energy consumption of the counterfactual sample after the building age is progressive shall not be lower than the predicted energy consumption of the original true sample. Constraint 2: Under the same conditions, the energy consumption prediction value of the counterfactual sample after the insulation state is switched from "no insulation" to "with insulation" is not higher than the energy consumption prediction value of the original true sample. Constraint 3: The predicted energy consumption ratio between different terminal types conforms to the preset engineering experience boundary.

7. The method for calculating the energy consumption retrofit effect of building energy-saving renovation according to claim 4, characterized in that, The specific method for applying the physical constraints is as follows: For the XGBoost tree model, the physical constraints are embedded into the training process through a custom objective function and monotonic constraints; by constructing twin samples and introducing a prior loss term, the model can minimize regression error while avoiding violation of physical constraints; the custom loss function and the prior loss term participate in gradient calculation together with the regression loss of standardized unit energy consumption during backpropagation, driving the model parameters to optimize in the direction of fitting the real data and satisfying physical laws.

8. The method for calculating the energy consumption retrofit effect of building energy-saving renovation according to claim 1, characterized in that, It also includes S5, which splits the actual monthly energy consumption data submitted by users into cold and hot categories to obtain cooling energy consumption and heating energy consumption, performs anomaly identification on the monthly energy consumption curve, and archives normal samples after anomaly identification into the database with their building type tags, and periodically merges new normal samples with the original training samples to update each energy consumption prediction model.

9. The method for calculating the energy consumption retrofit effect of building energy-saving renovation according to claim 8, characterized in that, The process of splitting the actual monthly energy consumption data submitted by users into cold and hot components includes: decomposing the monthly comprehensive energy consumption data submitted by users into basic energy consumption, cooling energy consumption, and heating energy consumption; the specific splitting method is as follows: combining the monthly total energy consumption curve, the monthly heating and cooling days under typical meteorological years, the cooling and heating season intervals, and building operation information, and for school-type businesses, adjusting the basic energy consumption coefficient according to the operating characteristics of winter and summer vacations; after splitting, the total annual cooling and heating capacity and the corresponding energy consumption index per unit area are obtained, and stored separately according to heating meters and cooling meters. The splitting results and their corresponding building type tags are used for subsequent rolling update training after anomaly identification and screening.

10. The method for calculating the energy consumption retrofit effect of building energy-saving renovation according to claim 8, characterized in that, The anomaly identification process for the monthly energy consumption curve includes the following steps: training a classification model using real monthly energy consumption data and synthetic anomaly data; extracting peak and valley features, fluctuation features, slope features, continuous identical value features, bimodal features, and similarity features with the standard trend from the monthly energy consumption curve to form an anomaly identification feature set; simultaneously calling the energy consumption prediction model to calculate the predicted cooling and heating energy consumption of the building under the current conditions, comparing it with the energy consumption breakdown results, and identifying abnormal energy consumption inputs. The classification model outputs anomaly probabilities and anomaly labels. The anomaly identification results and energy consumption breakdown results are saved together. In subsequent rolling updates, samples marked as anomalies are preferentially removed, and only normal samples are included in the training set.

Images