A method for predicting building thermal load of existing residential buildings based on machine learning

Abstract

This invention belongs to the field of heating technology, specifically relating to a machine learning-based method for predicting the heat load of existing residential buildings. It defines the overall framework of this machine learning-based method. A temperature and humidity correction model is established by integrating measured historical data, a standard climate database for buildings, and meteorological forecast data. A dynamic heat load index model is then established based on multi-dimensional building classification, and finally, the total heat load is calculated. This method solves the problems of large local biases in meteorological data and failure to consider individual building differences and dynamic changes in traditional forecasting. It achieves a shift from broad-area, extensive forecasting to localized, refined forecasting, providing a reliable data foundation for on-demand and precision heating, and helping to reduce energy waste and improve user comfort.

Description

Technical Field

[0001] This invention belongs to the field of heating technology, specifically relating to a method for predicting the heat load of existing residential buildings based on machine learning. Background Technology

[0002] With social development and improved living standards, users have increasingly higher demands for heating quality. However, most existing building heating systems are outdated and generally employ an extensive heating model, lacking accurate forecasting of heating load and unable to provide heating on demand. Most heating systems rely solely on recent weather forecasts and operational experience to control heating, failing to achieve precise heating. This results in fluctuating room temperatures, significant energy waste, severely impacting user experience, exacerbating user complaints, and intensifying conflicts between users and heating companies.

[0003] It is generally believed that the heat load index for centralized heating in residential buildings is 40-80 kWh / m² / a. In reality, the residential heat load index is a dynamic value, related not only to the outdoor climate but also to the building type (insulation performance of the building envelope), building orientation, and building height. However, for a given existing residential building, the outdoor temperature is constantly changing; simultaneously, the heat transfer coefficient of the building envelope also varies with temperature, humidity, and the building's age. Therefore, accurately determining the heat load index value is a prerequisite for on-demand and precise heating.

[0004] The main technical problems existing in the prior art include: (1) Meteorological forecast data is used for wide-area forecasting based on cities. The cities cover a large area (such as Hulunbuir City, which covers 260,000 square kilometers), resulting in a serious deviation between the predicted temperature and humidity and the actual local climate of the heat station. Direct application to a single heat station results in a huge error; (2) The heat load index only considers the rough differences in geographical areas and does not fully consider the different influences of different building types (energy-saving level), building orientation, and unit location (corner unit / middle unit) on the heat transfer coefficient of the building envelope; (3) The heat load index is mostly a static empirical value and cannot be adaptively updated with the increase of the building's service life, changes in the thermal insulation performance of the building envelope, and dynamic changes in outdoor temperature and humidity; (4) The frequency of temperature forecasting is low (usually in hours), which is not matched with the lag of the heat system regulation, resulting in a large lag in regulation.

[0005] To address the aforementioned technical issues, this invention proposes a machine learning-based method for predicting the heat load of existing residential buildings, enabling localized, refined, and dynamic heat load prediction to meet the needs of on-demand and precise heating. Summary of the Invention

[0006] This invention aims to address the problems in existing technologies, such as large local deviations in meteorological forecast data, failure to consider individual building differences and dynamic changes in heat load indicators, and low forecast frequency. It provides a heat load forecasting method for existing residential buildings based on machine learning, so as to achieve high-precision, fine-grained, adaptive iterative heat load forecasting and provide decision-making basis for smart heating.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A machine learning-based method for predicting the heat load of existing residential buildings includes the following steps: Step S1: Obtain historical temperature and humidity data measured near the heating station, standard meteorological database for buildings, and meteorological department's predicted future climate data. Establish a temperature and humidity correction model based on machine learning. The temperature and humidity correction model is used to output half-hourly temperature and humidity data for the next 24 hours at half-hour intervals. Step S2: Classify existing residential buildings according to three dimensions: building type, building orientation, and unit location to obtain several types of units; for each type of unit, calculate the heat load index per unit area by measuring its heat consumption and heating area, and establish a heat load index model based on machine learning; the heat load index model represents the dynamic relationship between the heat load index per unit area and outdoor temperature and humidity data. Step S3: Substitute the half-hourly temperature and humidity data obtained in step S1 into the heat load index model of the corresponding unit in step S2 to obtain the heat load index value per unit area for each unit. Calculate the corresponding heat load prediction value based on the heat load index value per unit area for each unit. Step S4: Weight the predicted heat load values ​​of all types of units obtained in step S3 to obtain the predicted total heat load value of the heating station.

[0008] Furthermore, the specific method for establishing the temperature and humidity correction model in step S1 includes: The support vector regression algorithm is used, and the temperature and humidity data include dry-bulb temperature and wet-bulb temperature; The inputs to the temperature and humidity correction model are the dry-bulb and wet-bulb temperatures predicted by the meteorological department for each time in the next 24 hours, the dry-bulb and wet-bulb temperatures at the corresponding time in the standard meteorological database for buildings, the difference between the dry-bulb temperature predicted by the meteorological department and the dry-bulb temperature in the standard meteorological database for buildings at the same time, the difference between the wet-bulb temperature predicted by the meteorological department and the wet-bulb temperature in the standard meteorological database for buildings at the same time, and the historical measured temperature and humidity data of the heating station. The output of the temperature and humidity correction model is the dry-bulb temperature correction factor and the wet-bulb temperature correction factor at the corresponding time. The model training process involves collecting daily meteorological forecast data from the meteorological department over the past few years, corresponding data from the standard meteorological database for buildings, and actual temperature and humidity data measured on the current day. The ratio of the actual measured temperature and humidity data to the corresponding data from the standard meteorological database is used as the label, and the meteorological forecast data and the corresponding data from the standard meteorological database are used as inputs to train a support vector regression model. Based on this, the predicted temperature and humidity values ​​for the vicinity of the heating station for the next 24 hours are obtained.

[0009] Furthermore, the building types in step S2 include non-energy-saving and one-step energy-saving buildings, two-step energy-saving buildings, three-step energy-saving buildings, and four-step energy-saving buildings; the building orientation includes north-south orientation, east-west orientation, southeast orientation, and northwest orientation; the unit location includes corner units and middle units; and 32 types of units are formed by combining the three dimensions of building type, building orientation, and unit location.

[0010] Furthermore, in step S2, when measuring the heat consumption of each type of unit, an acceptable indoor temperature non-guarantee rate of 5% is set. The indoor temperature of each household during the measurement period is statistically analyzed, and household data with temperatures below 18°C ​​and accounting for less than 5% of the total are excluded. The average indoor temperature of the remaining households is taken as the effective indoor temperature, and the heat consumption is measured under this condition.

[0011] Furthermore, in step S2, the specific method for establishing the heat load index model includes: Modeling is performed using multivariate nonlinear regression or BP neural network; the temperature and humidity data include dry-bulb temperature and wet-bulb temperature. The input to the heat load index model is the outdoor dry-bulb temperature and wet-bulb temperature, and the output of the heat load index model is the heat load index per unit area. Using multiple sets of measured temperature and humidity data and their corresponding heat load index values ​​as training samples, the model parameters are trained by minimizing the mean square error. The specific functional form of the heat load index model is automatically learned and determined by the algorithm from the data.

[0012] Furthermore, the heat load index model has an iterative update mechanism. Specifically, after the end of each heating season, the newly collected measured temperature and humidity data and the corresponding heat load index data are added to the historical dataset, and the heat load index model is retrained or incrementally learned to obtain the updated heat load index model parameters.

[0013] Furthermore, the temperature and humidity data includes dry-bulb temperature and wet-bulb temperature; in step S1, the iterative update method for the building standard meteorological database is to take the arithmetic mean of the dry-bulb temperature at a certain moment in the previous year's building standard meteorological database and the corresponding dry-bulb temperature measured near the heating station, and use this as the dry-bulb temperature of the building standard meteorological database; and to take the arithmetic mean of the wet-bulb temperature at a certain moment in the previous year's building standard meteorological database and the corresponding wet-bulb temperature measured near the heating station, and use this as the wet-bulb temperature of the building standard meteorological database.

[0014] Furthermore, in step S3, the predicted heat load value for each type of unit is calculated by multiplying the predicted heat load index per unit area of ​​each type of unit by the total heating area of ​​the corresponding type of unit.

[0015] Furthermore, in step S4, the formula for calculating the predicted total heat load of the heating station is as follows: in, This is the predicted total heat load of the heating station, in W. For the number of the class unit, ; The number of units in the i-th class; , where is the predicted value of the heat load index per unit area of ​​the i-th type of unit under the predicted temperature and humidity, in W / m². Let be the area of ​​the i-th type of unit building, in m².

[0016] Furthermore, in step S1, the measured temperature and humidity data near the heating station are recorded every 30 minutes; the standard meteorological database for buildings is updated annually.

[0017] The beneficial effects of this invention are: The technical solution of this invention defines the overall framework of a machine learning-based method for predicting the heat load of existing residential buildings. It establishes a temperature and humidity correction model by integrating historical measured data, a standard building database, and meteorological forecast data, and builds a dynamic heat load index model based on multi-dimensional building classification. Finally, it aggregates and calculates the total heat load. This method solves the problems of large local deviations in meteorological data and failure to consider individual building differences and dynamic changes in traditional forecasting. It realizes the transformation from broad-area, extensive forecasting to localized, refined forecasting, providing a reliable data foundation for on-demand and precision heating, and helping to reduce energy waste and improve user comfort. Attached Figure Description

[0018] Figure 1 The diagram shows a flowchart of the steps of a machine learning-based method for predicting the heat load of existing residential buildings according to the present invention. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments: like Figure 1 As shown, the present invention provides a method for predicting the heat load of existing residential buildings based on machine learning, comprising the following steps: Step S1: Obtain historical temperature and humidity data measured near the heating station, standard meteorological database for buildings, and meteorological department's predicted future climate data. Establish a temperature and humidity correction model based on machine learning. The temperature and humidity correction model is used to output half-hourly temperature and humidity data for the next 24 hours at half-hour intervals. Step S2: Classify existing residential buildings according to three dimensions: building type, building orientation, and unit location to obtain several types of units; for each type of unit, calculate the heat load index per unit area by measuring its heat consumption and heating area, and establish a heat load index model based on machine learning; the heat load index model represents the dynamic relationship between the heat load index per unit area and outdoor temperature and humidity data. Step S3: Substitute the half-hourly temperature and humidity data obtained in step S1 into the heat load index model of the corresponding unit in step S2 to obtain the heat load index value per unit area for each unit. Calculate the corresponding heat load prediction value based on the heat load index value per unit area for each unit. Step S4: Weight the predicted heat load values ​​of all types of units obtained in step S3 to obtain the predicted total heat load value of the heating station.

[0020] As can be seen from the above description, the present invention has the following beneficial effects: The technical solution of this invention defines the overall framework of a machine learning-based method for predicting the heat load of existing residential buildings. It establishes a temperature and humidity correction model by integrating historical measured data, a standard building database, and meteorological forecast data, and builds a dynamic heat load index model based on multi-dimensional building classification. Finally, it aggregates and calculates the total heat load. This method solves the problems of large local deviations in meteorological data and failure to consider individual building differences and dynamic changes in traditional forecasting. It realizes the transformation from broad-area, extensive forecasting to localized, refined forecasting, providing a reliable data foundation for on-demand and precision heating, and helping to reduce energy waste and improve user comfort.

[0021] Furthermore, the specific method for establishing the temperature and humidity correction model in step S1 includes: The support vector regression algorithm is used, and the temperature and humidity data include dry-bulb temperature and wet-bulb temperature; The inputs to the temperature and humidity correction model are the dry-bulb and wet-bulb temperatures predicted by the meteorological department for each time in the next 24 hours, the dry-bulb and wet-bulb temperatures at the corresponding time in the standard meteorological database for buildings, the difference between the dry-bulb temperature predicted by the meteorological department and the dry-bulb temperature in the standard meteorological database for buildings at the same time, the difference between the wet-bulb temperature predicted by the meteorological department and the wet-bulb temperature in the standard meteorological database for buildings at the same time, and the historical measured temperature and humidity data of the heating station. The output of the temperature and humidity correction model is the dry-bulb temperature correction factor and the wet-bulb temperature correction factor at the corresponding time. The model training process involves collecting daily meteorological forecast data from the meteorological department, corresponding data from the standard meteorological database for buildings, and actual temperature and humidity data for each day of the past year. The ratio of the actual measured temperature and humidity data to the corresponding data from the standard meteorological database for buildings is used as the label, and the meteorological forecast data and the corresponding data from the standard meteorological database for buildings are used as inputs to train a support vector regression model. Based on this, the predicted temperature and humidity values ​​for the vicinity of the heating station for the next 24 hours are obtained.

[0022] As described above, the specific method for establishing the temperature and humidity correction model is defined, including the use of support vector regression algorithm, explicit input and output features, and training process. By combining meteorological forecast data with localized standard database data and training with measured data as labels, the deviation between the wide-area forecast data of the meteorological department and the local actual climate of the heating station can be effectively eliminated. This results in the output of high-precision dry-bulb temperature correction coefficients and wet-bulb temperature correction coefficients, significantly improving the accuracy of half-hourly temperature and humidity forecasts for the next 24 hours and providing more accurate input parameters for subsequent heat load calculations.

[0023] Furthermore, the building types in step S2 include non-energy-saving and one-step energy-saving buildings, two-step energy-saving buildings, three-step energy-saving buildings, and four-step energy-saving buildings; the building orientation includes north-south orientation, east-west orientation, southeast orientation, and northwest orientation; the unit location includes corner units and middle units; and 32 types of units are formed by combining the three dimensions of building type, building orientation, and unit location.

[0024] As described above, the specific dimensions and combinations of building classifications are defined, combining three dimensions—building type (4 types), building orientation (4 types), and unit location (2 types)—to form 32 unit categories. This refined classification fully reflects the differences in heat transfer coefficients of building envelopes for buildings with different energy efficiency ratings, buildings with different orientations, and units in different locations (corner units / middle units). This allows for the establishment of differentiated heat load index models for each unit category, avoiding the prediction errors caused by the "one-size-fits-all" heat load index in traditional methods, and significantly improving the targeting and accuracy of heat load prediction.

[0025] Furthermore, in step S2, when measuring the heat consumption of each type of unit, an acceptable indoor temperature non-guarantee rate of 5% is set. The indoor temperature of each household during the measurement period is statistically analyzed, and the data of households with a temperature below 18°C ​​accounting for 5% of the total number are removed. The average indoor temperature of the remaining households is taken as the effective indoor temperature, and the heat consumption is measured under this condition.

[0026] As described above, the conditions for introducing a heating non-guarantee rate when measuring heat consumption are limited. An acceptable indoor temperature non-guarantee rate of 5% is set, excluding data from households with temperatures below 18℃ and less than 5% of the total, and households with temperatures above 22℃ and less than 5% of the total. This limitation makes the heat consumption measurement conditions more scientific, objective, and operable, avoiding interference from abnormal room temperatures of individual users (such as open windows or unauthorized adjustments), ensuring the representativeness and reliability of the measured data, and thus providing more accurate sample data for the subsequent training of the heat load index model.

[0027] Furthermore, in step S2, the specific method for establishing the heat load index model includes: Modeling is performed using multivariate nonlinear regression or BP neural network; the temperature and humidity data include dry-bulb temperature and wet-bulb temperature. The input to the heat load index model is the outdoor dry-bulb temperature and wet-bulb temperature, and the output of the heat load index model is the heat load index per unit area. Using multiple sets of measured temperature and humidity data and their corresponding heat load index values ​​as training samples, the model parameters are trained by minimizing the mean square error. The specific functional form of the heat load index model is automatically learned and determined by the algorithm from the data.

[0028] As described above, the specific method for establishing the heat load index model is defined. Multiple nonlinear regression or a shallow neural network (BP neural network) is used for modeling. The model inputs are dry-bulb and wet-bulb temperatures, and the output is the heat load index per unit area. The function form is automatically learned and determined by the algorithm from the data. This method does not require manual pre-setting of the function form and can adaptively fit the complex nonlinear relationship between outdoor temperature and humidity and heat load per unit area. Compared with traditional empirical formulas, it has stronger fitting and generalization capabilities, improving the prediction accuracy of the heat load index model.

[0029] Furthermore, the heat load index model has an iterative update mechanism. Specifically, after the end of each heating season, the newly collected measured temperature and humidity data and the corresponding heat load index data are added to the historical dataset, and the heat load index model is retrained or incrementally learned to obtain the updated heat load index model parameters.

[0030] As described above, an iterative update mechanism for the heat load index model is defined, meaning that the model is retrained or incrementally learned using newly collected measured data after each heating season. As the building's service life increases, the thermal insulation performance of the building envelope changes (e.g., aging of insulation materials, moisture absorption, etc.). This iterative update mechanism enables the model to reflect the year-on-year changes in the heat transfer performance of the building envelope in a timely manner, maintaining the model's timeliness and accuracy, and achieving the dynamic adaptive capability of the heat load index model.

[0031] Furthermore, the temperature and humidity data includes dry-bulb temperature and wet-bulb temperature; in step S1, the iterative update method for the building standard meteorological database is to take the arithmetic mean of the dry-bulb temperature at a certain moment in the previous year's building standard meteorological database and the corresponding dry-bulb temperature measured near the heating station, and use this as the dry-bulb temperature of the new building standard database; and to take the arithmetic mean of the wet-bulb temperature at a certain moment in the previous year's building standard meteorological database and the corresponding wet-bulb temperature measured near the heating station, and use this as the wet-bulb temperature of the new building standard database.

[0032] As described above, the iterative update method for the standard meteorological database used in buildings is defined by taking the arithmetic mean of the previous year's standard database data and the measured data near the heating station as the new database data. This update method is simple and easy to implement, and can gradually incorporate local measured climate characteristics from the heating station year by year, thus gradually localizing the standard database, eliminating systematic deviations between the general standard database and local actual climate, providing more accurate benchmark data for temperature and humidity correction models, and indirectly improving the accuracy of temperature and humidity predictions.

[0033] Furthermore, in step S3, the predicted heat load value for each type of unit is calculated by multiplying the predicted heat load index per unit area of ​​each type of unit by the total heating area of ​​the corresponding type of unit.

[0034] As described above, the specific calculation method for the predicted heat load of each type of unit is defined, namely, multiplying the predicted heat load index per unit area by the total heating area of ​​that type of unit. This calculation method is logically clear, computationally efficient, and can quickly obtain the predicted heat load of each type of unit, facilitating subsequent aggregate calculation of the total heat load, thus demonstrating the operability and practicality of the method of this invention.

[0035] Furthermore, in step S4, the formula for calculating the predicted total heat load of the heating station is as follows: in, This is the predicted total heat load of the heating station, in W. For the number of the class unit, ; The number of units in the i-th class; , where is the predicted value of the heat load index per unit area of ​​the i-th type of unit under the predicted temperature and humidity, in W / m². Let be the area of ​​the i-th type of unit building, in m².

[0036] As described above, a specific formula for calculating the total heat load forecast of a heating station is defined. The total heat load forecast is obtained by weighted summation of the heat load forecasts for 32 types of units. This formula has a clear structure and well-defined parameters (including the number of units in each type, the predicted heat load per unit area, and the area). It accurately reflects the comprehensive heat demand of all types of buildings under the jurisdiction of the heating station, providing a direct and quantitative basis for decision-making regarding the scheduling and operation of the heating station. This helps to achieve on-demand production by the heat source plant and on-demand allocation by the heating station, thereby achieving the goal of energy conservation and consumption reduction.

[0037] Furthermore, in step S1, the measured temperature and humidity data near the heating station are recorded every 30 minutes; the standard meteorological database for buildings is updated annually.

[0038] As described above, the frequency of temperature and humidity measurements (every 30 minutes) and the update cycle of the standard meteorological database for buildings are limited (once a year). Measurements recorded at a 30-minute granularity match the hysteresis characteristics of thermal system regulation, providing high-temporal-resolution training data for the temperature and humidity correction model; the annual database update balances data timeliness and computational cost. This limitation ensures clear and operable periodic standards for data acquisition and model updates, guaranteeing the feasibility of the method in practical engineering.

[0039] The following are several preferred embodiments or application embodiments to help those skilled in the art better understand the technical content of the present invention and the technical contributions made by the present invention compared with the prior art: Example 1 like Figure 1 As shown in the figure, this embodiment provides a machine learning-based method for predicting the heat load of existing residential buildings, applied to a typical heating station in a northern city. The heating station covers 12 buildings in an existing residential community, including non-energy-efficient buildings and three-step energy-saving buildings, with a total heating area of ​​approximately 80,000 square meters.

[0040] I. Predicting outdoor temperature and humidity Based on the standard meteorological database for buildings, historical climate data recorded by heating stations, and future climate forecasts from meteorological departments, relatively accurate half-hourly temperature and humidity data are obtained through machine learning calculations.

[0041] The specific steps are as follows: (1) The dry-bulb and wet-bulb temperatures near the heating station were measured and recorded every 30 minutes. An automatic weather station, including dry-bulb and wet-bulb temperature sensors, was installed in an open area near the heating station, and the data acquisition frequency was matched with the heating system control cycle.

[0042] (2) By combining the standard meteorological database for buildings and the historical temperature and humidity measured by heating stations over the past year, a self-iterably updating standard meteorological database for buildings (updated annually) is established. The calculation method is as follows: In the formula It is the dry-bulb temperature (°C) at a certain moment in the previous year's standard meteorological database for building use. It is the dry-bulb temperature (°C) measured at a certain moment near the heating station. It is the dry-bulb temperature (°C) at a certain moment in the updated standard meteorological database for building applications. It is the wet-bulb temperature (°C) at a certain moment in the previous year's building standard meteorological database. It is the wet-bulb temperature (°C) measured at a certain moment near the heating station. This is the wet-bulb temperature (°C) at a specific moment in the updated standard meteorological database for buildings. This iterative update is performed once a year after the heating season ends, incorporating the previous year's measured data into the standard database.

[0043] (3) Obtain the forecast temperature and humidity data for the next 24 hours from the meteorological department. Obtain the hourly dry-bulb temperature and wet-bulb temperature forecast values ​​for the next 24 hours issued by the local meteorological station through API interface or manual input.

[0044] (4) Compare the meteorological forecast data with the standard meteorological database for construction in the corresponding period, establish a temperature and humidity correction model based on the support vector regression (SVR) algorithm, and obtain the dry-bulb temperature correction coefficient for the next 24 hours. and wet-bulb temperature correction factor .

[0045] The specific modeling method is as follows: Input feature: The dry-bulb temperature predicted by the meteorological department for the i-th time in the next 24 hours. and wet-bulb temperature The dry-bulb temperature corresponding to the i-th time in the standard meteorological database for building applications. and wet-bulb temperature ; and the difference between the two. and .

[0046] Output characteristic: Dry-bulb temperature correction coefficient at time i in the next 24 hours and wet-bulb temperature correction factor .

[0047] Training process: Collect meteorological forecast data from the meteorological department for each day of the past year, corresponding data from the standard meteorological database for buildings, and actual temperature and humidity data measured on that day. Use the ratio of the actual measured temperature and humidity to the corresponding data from the standard meteorological database for buildings as the label. and The SVR model is trained using meteorological forecast data and corresponding data from a standard meteorological database for construction as input. The optimal kernel function (radial basis function RBF in this embodiment) and hyperparameters (penalty coefficient C=10, kernel parameter γ=0.1) are selected through cross-validation. It is important to note that the temperature data is converted to absolute temperature values ​​(K, not °C) during processing to avoid temperature oscillations at 0 °C.

[0048] (5) Based on the temperature and humidity correction coefficients and the standard meteorological database for building applications, the following data points were obtained for the next 24 hours: 48 dry-bulb temperatures and 48 wet-bulb temperatures (one data point per half hour): In this embodiment, the predicted local dry-bulb temperature of the heating station is -9.2℃ to -4.5℃, which is 0.8℃ higher than the average wide-area forecast value (-10℃ to 5℃) by the meteorological department. After correction, it is closer to the actual local climate.

[0049] II. Determination of Heat Load Index per Unit Area Based on the building characteristics of residential communities and measured heat load data, heat load index values ​​are derived, and a dynamic model of heat load index is established based on machine learning technology.

[0050] (6) Classify the residential buildings supplied by the heating station.

[0051] Buildings are categorized into four types based on building type: non-energy-efficient and first-stage energy-efficient buildings (built before 1995), second-stage energy-efficient buildings (built between 1995 and 2005), third-stage energy-efficient buildings (built between 2005 and 2015), and fourth-stage energy-efficient buildings (built after 2015). They are also categorized into four types based on building orientation: north-south, east-west, southeast, and northwest. Finally, they are categorized into two types based on unit location: corner units and middle units. Combining these three dimensions results in a total of 4 × 4 × 2 = 32 unit types.

[0052] The community in this embodiment includes two types of buildings: non-energy-saving buildings and three-step energy-saving buildings. It involves two types of buildings: north-south facing and east-west facing, as well as two types of units: corner units and middle units. There are actually 2×2×2=8 types of units, and the number of other types is 0.

[0053] (7) Measure the heat consumption of the unit building.

[0054] For each type of building unit, ultrasonic heat meters are installed on the supply and return water pipes of the heating system to measure flow rate and supply and return water temperature, recording data every 30 minutes. The measurement period is selected during the stable heating season (e.g., January), and measurements are taken continuously for 7 days.

[0055] During the measurement process, the concept of heating non-guarantee rate is introduced: when measuring the heat consumption of each type of unit, the heat consumption is classified and statistically analyzed according to the indoor temperature, and an acceptable indoor temperature non-guarantee rate of 5% is set. For residential indoor temperatures between 18-22℃, all are considered effective indoor temperatures, and their heat consumption index is directly calculated; data with indoor temperatures below 18℃ and above 22℃ within 5% of the total are excluded. Otherwise, for the area heat consumption index with indoor temperatures below 18℃, the calculated heat consumption is increased by 5% for every 1℃ decrease; for the area heat consumption index with indoor temperatures above 22℃, the calculated heat consumption is decreased by 5% for every 1℃ increase. If the indoor temperature is below 18℃ and within 5% of the total, the heat consumption of the above three parts is added together to obtain the heat consumption of the unit building.

[0056] (8) Determine the heating area of ​​each unit building.

[0057] The heating area (unit: m²) of each unit building is obtained based on building ownership documents or on-site measurements. In this embodiment, the area of ​​each unit building ranges from 2000 to 5000 m².

[0058] (9) The heat load index of each unit is calculated.

[0059] Calculate the heat load index per unit area for each unit using the following formula: In the formula, This represents the heat load index per unit area (W / m²) for a certain unit. The heat consumption (W) of a certain unit. The heating area (m²) of a certain unit. In this embodiment, the heat load per unit area of ​​a non-energy-saving building's corner unit is approximately 65W / m² when the outdoor temperature is -7℃, while that of a middle unit in a three-step energy-saving building is approximately 42W / m² at the same temperature.

[0060] (10) Establish a model for the change of heat load index value of a certain unit.

[0061] Based on machine learning technology, and according to the temperature, humidity and heat load index values ​​obtained in step (1), a heat load index change model is established: The specific modeling method is as follows: A backpropagation (BP) neural network with one hidden layer is used for modeling. The number of nodes in the input layer is 2 (dry-bulb temperature). and wet-bulb temperature The hidden layer has 10 nodes, the activation function is ReLU, and the output layer has 1 node (heat load per unit area index). The output layer has no activation function.

[0062] The multiple sets of temperature and humidity values ​​and their corresponding heat load index values ​​obtained in step (7) are used as training samples (approximately 336 sets of data are obtained for each class of units in this embodiment, covering 7 days × 24 hours × 2.5 hours). The model parameters are trained by minimizing the mean squared error (MSE) using the Adam optimizer with a learning rate of 0.001 and 500 training epochs. The specific functional form of the model is automatically learned and determined by the algorithm from the data, without the need for manual pre-setting.

[0063] Iterative update mechanism: After the heating season ends each year, the newly collected measured temperature and humidity data and corresponding heat load index data are added to the historical dataset. The model is then incrementally learned (the online gradient descent method is used to update the neural network weights) to obtain the updated model parameters, so that the model can reflect the changes in the thermal insulation performance of the building envelope with the service life.

[0064] III. Determining the predicted heat load value at a certain moment (11) Predict the heat load index value of a certain unit at a certain moment.

[0065] Based on the temperature and humidity data predicted in step (5) at a certain moment (e.g., dry-bulb temperature of -7.2℃ and wet-bulb temperature of -8.5℃ at 10:00 AM on January 15, 2024), substitute it into the heat load index model of the corresponding type of unit obtained in step (10) to obtain the predicted value of the heat load index per unit area of ​​the unit at that moment. .

[0066] (12) Calculate the heat load value of a certain unit at a certain moment.

[0067] For a given unit, its heat load value is calculated using the following formula: In the formula, The predicted heat load (W) for the unit. This represents the total heating area (m²) of the unit.

[0068] (13) Calculate the predicted total heat load of the heating station.

[0069] The total heat load of the heating station can be calculated using the following formula: In the formula, It is the predicted total heat load (W) of the heating station. It is the unit type number ( ), It is the number of units of the i-th class. It is the predicted value (W / m²) of the heat load index per unit area of ​​the i-th type of unit under the predicted temperature and humidity. It is the area (m²) of the i-th type of unit building.

[0070] In this embodiment, the predicted heat load values ​​for the eight types of units are calculated and summed to obtain a total predicted heat supply of 115.3 MWh for the heating station on January 15, 2024.

[0071] IV. Verification and Results The predicted total heat load of the heating station was compared with the actual heat supply. Actual meter readings showed that the actual heat supply for that day was 117.8 MWh. In contrast, the traditional method (based solely on the meteorological department's dry-bulb temperature forecast, without distinguishing building categories) predicted the total heat supply of the heating station for that day to be 124.6 MWh.

[0072] The prediction error of traditional methods is The prediction error of the method of this invention is Based on a heating season of 120 days, the method of this invention can achieve an energy saving rate of approximately 3% to 5%, significantly improve the stability of indoor temperature for users, and reduce the user complaint rate by approximately 40%.

[0073] Example 2 This embodiment is basically the same as Embodiment 1, except that a multivariate nonlinear regression is used instead of a BP neural network when establishing the heat load index model.

[0074] Specifically, for each type of unit, based on the outdoor dry-bulb temperature and wet-bulb temperature The independent variable is the heat load per unit area. As the dependent variable, a multinomial regression model (maximum order 3, including interaction terms) is used for fitting. The model form is: Solving regression coefficients using the least squares method This embodiment uses the Akaike Information Criterion (AIC) to select the model and automatically determine the optimal number of terms. This method is faster and suitable for scenarios with small amounts of data or high real-time requirements.

[0075] The remaining steps and parameters are the same as in Example 1, and will not be repeated here.

[0076] Example 3 This embodiment is basically the same as embodiment 1, except that the heat consumption measurement is processed in step (7).

[0077] In this embodiment, the handling of the indoor temperature non-guarantee rate is more refined: when measuring the indoor temperature of each household during the statistical measurement period, not only are households with temperatures below 18°C ​​excluded, but also households with temperatures above 22°C (possibly due to actively opening windows or unauthorized adjustments), with each exclusion accounting for 2.5%, totaling 5%. The average indoor temperature of the remaining households (accounting for 95% of the total) is taken as the effective indoor temperature, and this effective indoor temperature value is recorded when measuring heat consumption, serving as an input feature for the subsequent model (i.e., the indoor target temperature). This allows the heat load index model to consider the differences in users' indoor temperature setpoints during prediction, further improving prediction flexibility.

[0078] The remaining steps and parameters are the same as in Example 1, and will not be repeated here.

[0079] Example 4 This application provides a machine learning-based system for predicting the heat load of existing residential buildings, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the machine learning-based method for predicting the heat load of existing residential buildings.

[0080] Unlike existing technologies, the technical solution of this application embeds the heat load prediction method into a computer program stored in memory, which is then executed by a processor to achieve a fully automated heat load prediction process. This system can automatically acquire historical temperature and humidity data measured near the heating station, a standard meteorological database for buildings, and meteorological forecast data. Based on a support vector regression algorithm, it establishes a temperature and humidity correction model and outputs high-precision, high-temporal-resolution (half-hour interval) half-hourly temperature and humidity data. It automatically completes 32 detailed classifications of existing residential buildings according to three dimensions (building type, building orientation, and unit location) and establishes a dynamic heat load index model for each type of unit based on measured data. Finally, it automatically aggregates and calculates the predicted total heat load value for the heating station. This system can be directly deployed in the heating station monitoring system or a cloud-based heating management platform to achieve online, real-time, and automated prediction of heating load, significantly reducing manual intervention and maintenance costs, improving prediction efficiency and accuracy, and providing immediate decision support for on-demand control of smart heating systems. It has good engineering practicality and scalability.

[0081] Example 5 This application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the machine learning-based method for predicting the heat load of existing residential buildings.

[0082] This medium can be distributed and sold independently (e.g., USB flash drive, external hard drive, optical disc, etc.) or integrated into various computing devices (e.g., servers, industrial control computers, edge computing gateways, etc.). Users can quickly deploy the heat load prediction method of this invention by loading the program from this medium onto a processor for execution, without the need for redeveloping algorithms or performing complex system integration. This storage medium makes the method of this invention highly portable and easy to use, facilitating replication and promotion among different heating stations and heating companies, thus lowering the technical application threshold. Simultaneously, the programmed prediction process ensures the consistency and reproducibility of prediction results, avoiding subjective biases from human experience judgment, and providing fundamental support for the standardized and intelligent management of heating systems.

[0083] In summary, the technical solution of this invention combines historical measured temperature and humidity data from heating stations, a standard meteorological database for buildings, and meteorological forecast data. Based on a support vector regression algorithm, a temperature and humidity correction model is established, narrowing the prediction range from the city level to the heating station level, thus eliminating local biases in wide-area predictions. Simultaneously, using a half-hour prediction granularity matches the lag in heating system regulation, providing high temporal resolution input parameters for precise heating. This invention divides existing residential buildings into 32 unit types based on three dimensions: building type (4 energy-saving levels), building orientation (4 types), and unit location (2 types). It fully considers the differences in heat transfer coefficients of building envelopes for different energy-saving levels, the differences in solar radiation heat gain for buildings with different orientations, and the differences in heat transfer boundaries between end units and middle units, making the heat load index model more physically targeted and predictively accurate. This invention establishes a dynamic model of the heat load per unit area as a function of outdoor temperature and humidity based on measured data. The model is retrained or incrementally learned annually using newly collected data, enabling it to promptly reflect changes in the thermal insulation performance of building envelopes over their service life and in heat transfer performance under different temperature and humidity conditions, maintaining the model's timeliness and accuracy. This invention obtains the total heat load of a heating station by predicting the heat load of 32 different unit types and then weighting and summing the results. The formula is clear, and the parameters are well-defined, providing a direct and quantitative basis for decision-making regarding the scheduling and operation of heating stations. Practical application verification shows that the prediction error using this method can be reduced to 2.1%, a significant improvement compared to traditional methods (error 5.8%), achieving an energy saving rate of 3% to 5% and effectively improving the stability of indoor temperature for users.

[0084] The present invention has been described with reference to the foregoing embodiments and accompanying drawings; however, the foregoing embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, modifications and equivalents included within the spirit and scope of the claims are all included within the scope of the present invention.

Claims

1. A method for predicting the heat load of existing residential buildings based on machine learning, characterized in that, Includes the following steps: Step S1: Obtain historical temperature and humidity data measured near the heating station, standard meteorological database for buildings, and meteorological department's predicted future climate data. Establish a temperature and humidity correction model based on machine learning. The temperature and humidity correction model is used to output half-hourly temperature and humidity data for the next 24 hours at half-hour intervals. Step S2: Classify existing residential buildings according to three dimensions: building type, building orientation, and unit location to obtain several types of units; for each type of unit, calculate the heat load index per unit area by measuring its heat consumption and heating area, and establish a heat load index model based on machine learning; the heat load index model represents the dynamic relationship between the heat load index per unit area and outdoor temperature and humidity data. Step S3: Substitute the half-hourly temperature and humidity data obtained in step S1 into the heat load index model of the corresponding unit in step S2 to obtain the heat load index value per unit area for each unit. Calculate the corresponding heat load prediction value based on the heat load index value per unit area for each unit. Step S4: Weight the predicted heat load values ​​of all types of units obtained in step S3 to obtain the predicted total heat load value of the heating station.

2. The method for predicting the heat load of existing residential buildings based on machine learning according to claim 1, characterized in that, The specific method for establishing the temperature and humidity correction model in step S1 includes: The support vector regression algorithm is used, and the temperature and humidity data include dry-bulb temperature and wet-bulb temperature; The inputs to the temperature and humidity correction model are the dry-bulb and wet-bulb temperatures predicted by the meteorological department for each time in the next 24 hours, the dry-bulb and wet-bulb temperatures at the corresponding time in the standard meteorological database for buildings, the difference between the dry-bulb temperature predicted by the meteorological department and the dry-bulb temperature in the standard meteorological database for buildings at the same time, the difference between the wet-bulb temperature predicted by the meteorological department and the wet-bulb temperature in the standard meteorological database for buildings at the same time, and the historical measured temperature and humidity data of the heating station. The output of the temperature and humidity correction model is the dry-bulb temperature correction factor and the wet-bulb temperature correction factor at the corresponding time. The model training process involves collecting meteorological forecast data from the meteorological department, corresponding data from the standard meteorological database for buildings, and actual temperature and humidity data measured on each day over the past few years. The ratio of the actual measured temperature and humidity data to the corresponding data from the standard meteorological database for buildings is used as the label, and the meteorological forecast data and the corresponding data from the standard meteorological database for buildings are used as inputs to train the support vector regression model.

3. The method for predicting the heat load of existing residential buildings based on machine learning according to claim 1, characterized in that, The building types in step S2 include non-energy-saving and one-step energy-saving buildings, two-step energy-saving buildings, three-step energy-saving buildings and four-step energy-saving buildings; the building orientation includes north-south orientation, east-west orientation, southeast orientation and northwest orientation; the unit location includes corner units and middle units; 32 types of units are formed by combining the three dimensions of building type, building orientation and unit location.

4. The method for predicting the heat load of existing residential buildings based on machine learning according to claim 1, characterized in that, In step S2, when measuring the heat consumption of each type of unit, an acceptable indoor temperature non-guarantee rate of 5% is set. The indoor temperature of each household during the measurement period is statistically analyzed. Household data with temperatures below 18℃ and accounting for less than 5% of the total, or above 22℃ and accounting for less than 5% of the total, are removed. The average indoor temperature of the remaining households is taken as the effective indoor temperature, and the heat consumption is measured under this condition.

5. The method for predicting the heat load of existing residential buildings based on machine learning according to claim 1, characterized in that, In step S2, the specific method for establishing the heat load index model includes: Modeling is performed using multivariate nonlinear regression or BP neural network; the temperature and humidity data include dry-bulb temperature and wet-bulb temperature. The input to the heat load index model is the outdoor dry-bulb temperature and wet-bulb temperature, and the output of the heat load index model is the heat load index per unit area. Using multiple sets of measured temperature and humidity data and their corresponding heat load index values ​​as training samples, the model parameters are trained by minimizing the mean square error. The specific functional form of the heat load index model is automatically learned and determined by the algorithm from the data.

6. The method for predicting the heat load of existing residential buildings based on machine learning according to claim 5, characterized in that, The heat load index model has an iterative update mechanism. Specifically, after the end of each heating season, the newly collected measured temperature and humidity data and the corresponding heat load index data are added to the historical dataset. The heat load index model is then retrained or incrementally learned to obtain the updated heat load index model parameters.

7. The method for predicting the heat load of existing residential buildings based on machine learning according to claim 1, characterized in that, The temperature and humidity data include dry-bulb temperature and wet-bulb temperature; In step S1, the iterative update method for the standard meteorological database for buildings is to take the arithmetic mean of the dry-bulb temperature at a certain moment in the standard meteorological database for buildings from the previous year and the corresponding dry-bulb temperature measured near the heating station, and use this as the dry-bulb temperature of the new database; and to take the arithmetic mean of the wet-bulb temperature at a certain moment in the standard meteorological database for buildings from the previous year and the corresponding wet-bulb temperature measured near the heating station, and use this as the wet-bulb temperature of the new database; thus forming a new standard climate database for buildings.

8. The method for predicting the heat load of existing residential buildings based on machine learning according to claim 1, characterized in that, In step S3, the predicted heat load value for each type of unit is calculated by multiplying the predicted heat load index per unit area of ​​each type of unit by the total heating area of ​​the corresponding type of unit.

9. The method for predicting the heat load of existing residential buildings based on machine learning according to claim 1, characterized in that, In step S4, the formula for calculating the predicted total heat load of the heating station is as follows: in, This is the predicted total heat load of the heating station, in W. For the number of the class unit, ; The number of units in the i-th class; , where is the predicted value of the heat load index per unit area of ​​the i-th type of unit under the predicted temperature and humidity, in W / m². Let be the area of ​​the i-th type of unit building, in m².

10. The method for predicting the heat load of existing residential buildings based on machine learning according to claim 1, characterized in that, In step S1, the measured temperature and humidity data near the heating station are recorded every 30 minutes; the standard meteorological database for buildings is updated annually.

Images