Building energy consumption prediction method and device based on machine learning

Abstract

The invention relates to a building energy consumption prediction method and device based on machine learning, and the method comprises the steps: S1, obtaining a benchmark building energy consumption model which is established based on an existing standard specification and referring to typical office building cases, and is used for obtaining month-by-month energy consumption based on the value of each key parameter; s2, sampling each key parameter by adopting a Latin hypercube sampling method to obtain a plurality of first samples, obtaining corresponding month-by-month energy consumption according to each first sample based on a reference building energy consumption model, and combining each first sample with the corresponding month-by-month energy consumption to generate a second sample; s3, training a building energy consumption prediction model by using the second sample; and S4, performing energy consumption prediction based on the trained building energy consumption prediction model. Compared with the prior art, the building energy consumption prediction method has the advantages that the calculation amount is reduced while the building energy consumption prediction accuracy is improved, and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of building energy conservation, and in particular to a method and device for predicting building energy consumption based on machine learning. Background Art

[0002] With the rapid economic development and the continuous advancement of urbanization, energy consumption in the construction industry continues to increase, and building energy conservation has become a key issue in addressing climate change. Building energy consumption can effectively reflect and represent the operating status of buildings. Accurate analysis and evaluation of it is a practical need for building energy-saving transformation and smart energy management.

[0003] At present, building energy consumption prediction methods are mainly divided into two categories: physical simulation methods and data-driven methods. The physical simulation method simulates the energy usage of buildings based on the principles of thermodynamics, and can more accurately consider the operation and use of buildings, so the energy consumption simulation results are more accurate. However, the physical simulation method is more complicated to operate, relies on a large amount of detailed building information, and the model calibration process requires users to have high professional knowledge. On the other hand, the data-driven method uses historical energy consumption data and machine learning algorithms to reveal the potential linear or nonlinear relationship between input and output variables, and has the advantages of low computational cost and fast prediction speed. The data-driven method is very friendly to non-professionals, but it has high requirements on the amount of data, and is prone to model overfitting problems, which affects the accuracy of the prediction results. Summary of the invention

[0004] The purpose of the present invention is to provide a method and device for predicting building energy consumption based on machine learning.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A building energy consumption prediction method based on machine learning, comprising:

[0007] Step S1: Obtain a benchmark building energy consumption model, wherein the benchmark building energy consumption model is established based on existing standard specifications and with reference to typical office building cases, and is configured to obtain monthly energy consumption based on the values ​​of various key parameters;

[0008] Step S2: Sampling each key parameter using a Latin hypercube sampling method to obtain a plurality of first samples, obtaining a corresponding monthly energy consumption according to each first sample based on the benchmark building energy consumption model, and combining each first sample with the corresponding monthly energy consumption to generate a second sample;

[0009] Step S3: using the second sample to train a building energy consumption prediction model;

[0010] Step S4: Perform energy consumption prediction based on the trained building energy consumption prediction model.

[0011] The benchmark building energy consumption model is developed based on EnergyPlus, SketchUp and Openstudio.

[0012] The key parameters include:

[0013] Building characteristic parameters, including at least the heat transfer coefficient of the exterior wall, the heat transfer coefficient of the exterior window, the heat transfer coefficient of the roof and the building orientation;

[0014] The user behavior parameters include at least personnel density, lighting power density, equipment power density, cooling temperature setting value and heating temperature setting value.

[0015] The step S2 comprises:

[0016] Step S2-1: Determine the variation range of each key parameter according to literature records and standard specifications;

[0017] Step S2-2: Divide the variation range of each key parameter into multiple small intervals of equal width, and randomly extract a sample value in each small interval for each key parameter, ensuring that each small interval is extracted only once, and obtain multiple non-repeating first samples by arranging and combining the sample values ​​of each key parameter between the small intervals, thereby fully covering the parameter space;

[0018] Step S2-3: inputting the obtained first sample into the benchmark building energy consumption model, and generating corresponding monthly energy consumption through simulation of the benchmark building energy consumption model;

[0019] Step S2-4: Combine each first sample with the corresponding monthly energy consumption to generate a second sample.

[0020] In step S2-2, the variation range of each key parameter is divided into a plurality of small intervals of equal width.

[0021] The sample values ​​extracted from each small interval are:

[0022]

[0023] Where: sv ij is the jth sample value of the i-th key parameter, a i is the lower limit of the variation range of the i-th key parameter, b i is the upper limit of the variation range of the i-th key parameter, N i is the number of small intervals where the i-th key parameter is divided, ∈ ij is a random disturbance ranging from 0 to 1.

[0024] The step S2-4 specifically includes:

[0025] Step S2-4-1: preprocessing the generated monthly energy consumption data set, wherein the preprocessing process includes converting energy consumption into energy consumption intensity;

[0026] Step S2-4-2: Combine each first sample, the corresponding month and its energy consumption intensity to generate a second sample.

[0027] The step S3 comprises:

[0028] Step S3-1: Map all second samples to the root node and define a mapping function q(x i )=0, where x i is the i-th second sample;

[0029] Step S3-2: Initialize the leaf node set to be segmented L={0}, and the current leaf node set only includes the root node;

[0030] Step S3-3: Select a leaf node l from the leaf node set for segmentation;

[0031] Step S3-4: for a leaf node l, traverse all possible segmentation points s in the leaf node and calculate the segmentation gain Gain;

[0032] Step S3-5: Select the segmentation point s with the largest segmentation gain as the segmentation point to segment the leaf node l, and update the mapping between samples and nodes:

[0033]

[0034] Among them, l L is the left node after the split, l R is the right node after splitting;

[0035] Step S3-6: Repeat steps S3-3 to S3-5 until the maximum node segmentation depth is reached or the segmentation gain Gain ≤ 0;

[0036] Step S3-6: Assign weight to each node.

[0037] The split gain Gain is:

[0038]

[0039] G=∑ i∈I g i

[0040] H=∑ i∈I h i

[0041] G R =GG L

[0042] H R =HH L

[0043] Where: G L is the sum of the first-order gradients of all second samples in the left node, H L is the sum of the second-order gradients of all second samples in the left node, λ is the regularization parameter of the weight, γ is the regularization parameter of the number of leaf nodes, G is the sum of the first-order gradients of all second samples in node l, H is the sum of the second-order gradients of all second samples in node l, g i is the first-order gradient of the second sample i, h i is the second-order gradient of the second sample i, I is the set of all second samples of node l, G R is the sum of the first-order gradients of all second samples in the right node, H R is the sum of the second-order gradients of all second samples in the right node.

[0044] The weight of the node in step S3-6 is:

[0045]

[0046] Where: j is the weight of node j, G j is the sum of the first-order gradients of all second samples in node j, H j is the sum of the first-order gradients of all second samples in node j.

[0047] A building energy consumption prediction device based on machine learning includes a memory, a processor, and a program stored in the memory, characterized in that the processor implements the above-mentioned method when executing the program.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. By integrating the advantages of physical simulation and data-driven methods, the accuracy and practicality of building energy consumption prediction are effectively improved. By developing a benchmark building energy consumption model and combining Latin hypercube sampling with EnergyPlus-Python co-simulation technology, large-scale energy consumption simulation can be performed efficiently, thus effectively overcoming the limitations of traditional physical simulation methods in data processing and operational complexity.

[0050] 2. The generated building energy consumption dataset can comprehensively cover various daily operation conditions of office buildings, providing a data basis for the training of the extreme gradient boosting algorithm, thereby reducing the risk of model overfitting and significantly enhancing the adaptability and coverage of the model.

[0051] 3. The user-friendly nature of machine learning models greatly lowers the technical threshold, enabling non-professionals to effectively analyze and manage energy usage, thus providing greater possibilities for the widespread application of energy forecasting technology.

[0052] 4. Based on the benchmark building energy consumption model, the monthly energy consumption is obtained by setting the monthly information as a virtual variable. Combined with the building characteristic parameters and the user behavior parameters, compared with the method of directly setting temperature and humidity as variables, it can solve the problem of interference of unconventional combinations caused by sampling combinations of temperature and humidity on model training. In this way, the synergistic effect of the three key energy consumption influencing factors of climate, building characteristics, and human behavior can be fully considered, and the accuracy and applicability of building energy consumption prediction can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic flow chart of the main steps of the method of the present invention;

[0054] Figure 2 This is a structural diagram of the extreme gradient boosting algorithm used in the present invention.

[0055] Figure 3 This is a structural diagram of the building energy consumption prediction device of the present invention. DETAILED DESCRIPTION

[0056] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0057] A building energy consumption prediction method based on machine learning, such as Figure 1 As shown, including:

[0058] Step S1: Obtain a benchmark building energy consumption model, wherein the benchmark building energy consumption model is established based on existing standard specifications and with reference to typical office building cases, and is configured to obtain monthly energy consumption based on the values ​​of various key parameters;

[0059] In this embodiment, the benchmark building energy consumption model is developed based on EnergyPlus, SketchUp and Openstudio.

[0060] Specifically, EnergyPlus, SketchUp and Openstudio were used to develop a benchmark building energy consumption model. EnergyPlus was used as the main building energy consumption simulation tool, SketchUp was used to create and visualize the building geometry model, and OpenStudio was used to modify the structural properties, materials, indoor loads and HVAC system properties of the building model.

[0061] Generally, the benchmark building energy consumption model requires three main input parameters: building characteristics, user behavior and meteorological conditions. Building characteristics include building geometry, functional layout and enclosure composition, and the values ​​of these parameters are mainly determined based on literature collection, map websites and street view data. User behavior parameters such as occupancy rate, lighting and equipment power density, personnel heat dissipation, cooling and heating set temperatures are set according to local building energy conservation standards and other standard specifications. Meteorological conditions use the hourly typical meteorological year data set from the EnergyPlus official website.

[0062] In most cases, the key parameters include:

[0063] Building characteristic parameters, including at least the heat transfer coefficient of the exterior wall, the heat transfer coefficient of the exterior window, the heat transfer coefficient of the roof and the building orientation;

[0064] User behavior parameters, including at least personnel density, lighting power density, equipment power density, cooling temperature setting value and heating temperature setting value;

[0065] Meteorological condition parameters, including at least temperature and humidity.

[0066] Step S2: using the Latin hypercube sampling method to sample each key parameter to obtain a plurality of first samples, obtaining the corresponding monthly energy consumption according to each first sample based on the benchmark building energy consumption model, and combining each first sample with the corresponding monthly energy consumption to generate a second sample;

[0067] The Latin hypercube sampling method is used to sample parameters to generate a large sample parameter data set covering a variety of energy usage scenario combinations. The EnergyPlus-Python joint simulation integration is used to perform automated multi-scenario building energy simulation to generate a building energy consumption data set that comprehensively covers all types of daily operating conditions of office buildings.

[0068] First, the variable parameters and their possible range of variation are determined. The key parameters of the three types of building energy consumption influencing factors identified, namely the variables of the benchmark building energy consumption model, include building characteristic parameters (exterior wall heat transfer coefficient, exterior window heat transfer coefficient, roof heat transfer coefficient, building orientation, etc.), user behavior parameters (staff density, lighting power density, equipment power density, cooling temperature setting value, heating temperature setting value, etc.) and meteorological conditions (temperature, humidity, etc.). The building envelope characteristics are described by the heat transfer coefficients of the exterior walls, roofs and windows, which are affected by the thickness of the insulation layer. Therefore, the thickness of the insulation layer of the benchmark model is modified to change the heat transfer coefficient. According to the changes and constraints in the relevant literature and standards, the potential value range of the insulation layer thickness and the heat transfer coefficient is determined. The building orientation has a significant impact on the indoor temperature, and its possible range is set to 0-360 degrees. The value range of the user behavior parameters is determined based on a large number of literature surveys.

[0069] Secondly, Latin hypercube sampling was used to extract parameter combination samples from the above multidimensional distribution, generating a total of 5000 different parameter combinations, representing different scenarios that may be encountered during building operation.

[0070] In this embodiment, there are 9 input parameters, namely: exterior wall heat transfer coefficient, exterior window heat transfer coefficient, roof heat transfer coefficient, building orientation, occupant density, lighting power density, equipment power density, cooling temperature set point, and heating temperature set point.

[0071] Specifically, step S2 includes:

[0072] Step S2-1: Determine the variation range of each key parameter according to literature records and standard specifications;

[0073] Step S2-2: Divide the variation range of each key parameter into multiple small intervals of equal width, and randomly extract a sample value in each small interval for each key parameter, ensuring that each small interval is extracted only once, and obtain multiple non-repeating first samples by arranging and combining the sample values ​​of each key parameter between the small intervals, thereby fully covering the parameter space;

[0074] The sample values ​​drawn from each small interval are:

[0075]

[0076] Where: sv ij is the jth sample value of the i-th key parameter, a i is the lower limit of the variation range of the i-th key parameter, b i is the upper limit of the variation range of the i-th key parameter, N i is the number of small intervals where the i-th key parameter is divided, ∈ ij is a random disturbance ranging from 0 to 1.

[0077] In this way, if the small intervals into which the key parameters are divided are all N, a set of N sample values ​​is formed:

[0078] SV={(sv 1j ,sv 2j ,…,sv 9j )|j=1,2,…,N}

[0079] Step S2-3: inputting the obtained first sample into the benchmark building energy consumption model, and generating corresponding monthly energy consumption through simulation of the benchmark building energy consumption model;

[0080] After that, EnergyPlus and Python are combined for automatic joint simulation. First, the Python script dynamically retrieves key parameters, inputs them into the EnergyPlus input file, and uses the os.system() command to execute the EnergyPlus simulation task and transfer the required parameters for energy consumption simulation. After the simulation is completed, the script captures the output data and finally generates a monthly building energy consumption dataset. The joint simulation process S is defined as follows:

[0081] S=Θ→CSV

[0082] Here, Θ is the set of model parameters and CSV is the output file of simulation results.

[0083] The model parameter θ can be composed of a series of parameters θ1, θ2, …, θ n Composition, each parameter θ i Corresponds to a specific value or setting in the simulation file (benchmark building energy consumption model).

[0084] Each iteration of the simulation can be represented as obtaining a set of parameter combinations from the sample set X, applying the parameter combination to modify the simulation file and running the simulation. The process can be represented as follows:

[0085]

[0086] S(modefiedfile k →result k )

[0087] Collect(result1,result2,…,result m )

[0088] Where f is a function that accepts a base file and a set of parameters as input and outputs a modified file; S is the EnergyPlus simulation process that takes the modified file as input and outputs a result file.

[0089] Step S2-4: Combining each first sample with the corresponding monthly energy consumption to generate a second sample, specifically including:

[0090] Step S2-4-1: preprocessing the generated monthly energy consumption data set, wherein the preprocessing process includes converting energy consumption into energy consumption intensity;

[0091] Specifically, first, the generated building energy consumption dataset is preprocessed. ① Normalization: Convert building energy consumption into energy consumption intensity to reduce the impact of building size on energy consumption; ② Categorical variable conversion: The energy consumption simulation results are output by month to form categorical variables, and unique hot encoding is used to consider the impact of meteorological conditions. ③ Dataset division: The dataset is randomly divided into three parts, 70% for training, 15% for verification, and 15% for testing. The exterior wall heat transfer coefficient, exterior window heat transfer coefficient, roof heat transfer coefficient, building orientation, personnel density, lighting power density, equipment power density, cooling temperature set point, and heating temperature set point are used as the input of the machine learning model, and the building energy consumption intensity is used as the output of the machine learning model to construct an extreme gradient boosting machine learning prediction model.

[0092] Step S2-4-2: Combine each first sample, the corresponding month and its energy consumption intensity to generate a second sample.

[0093] Step S3: Using the second sample to train the building energy consumption prediction model,

[0094] Development of machine learning models, including training, validation, and testing steps of machine learning models. Use Python libraries to implement model development. Combine grid search and ten-fold cross validation to optimize model hyperparameters and improve model accuracy. Evaluation indicators of the model development process include root mean square error (RMSE), determination coefficient (R2), and mean absolute error (MAE).

[0095] The extreme gradient boosting algorithm is used to construct a building energy consumption prediction model, such as Figure 2 As shown, including:

[0096] Step S3-1: Map all second samples to the root node and define a mapping function q(x i )=0, where x i is the i-th second sample;

[0097] Step S3-2: Initialize the leaf node set to be segmented L={0}, and the current leaf node set only includes the root node;

[0098] Step S3-3: Select a leaf node l from the leaf node set for segmentation;

[0099] Step S3-4: for a leaf node l, traverse all possible segmentation points s in the leaf node and calculate the segmentation gain Gain;

[0100] The split gain Gain is:

[0101]

[0102] G=∑ i∈I gi

[0103] H=∑ i∈I h i

[0104] G R =GG L

[0105] H R =HH L

[0106] Where: G L is the sum of the first-order gradients of all second samples in the left node, H L is the sum of the second-order gradients of all second samples in the left node, λ is the regularization parameter of the weight, γ is the regularization parameter of the number of leaf nodes, G is the sum of the first-order gradients of all second samples in node l, H is the sum of the second-order gradients of all second samples in node l, g i is the first-order gradient of the second sample i, h i is the second-order gradient of the second sample i, I is the set of all second samples of node l, G R is the sum of the first-order gradients of all second samples in the right node, H R is the sum of the second-order gradients of all second samples in the right node.

[0107] In addition, the first-order gradient of the second sample i and the second-order gradient of the second sample i are specifically as follows:

[0108]

[0109] Step S3-5: Select the segmentation point s with the largest segmentation gain as the segmentation point to segment the leaf node l, and update the mapping between samples and nodes:

[0110]

[0111] Among them, l L is the left node after the split, l R is the right node after splitting;

[0112] At this point, you can update the list of leaf nodes to be split:

[0113] L=(L\{l})∪{l L ,l R}

[0114] Step S3-6: Repeat steps S3-3 to S3-5 until the maximum node segmentation depth is reached or the segmentation gain Gain ≤ 0;

[0115] Step S3-6: Assign weight to each node.

[0116] The weight of the node in step S3-6 is:

[0117]

[0118] Where: j is the weight of node j, G j is the sum of the first-order gradients of all second samples in node j, H j is the sum of the first-order gradients of all second samples in node j.

[0119] Finally, the XGBoost regression model is constructed:

[0120] The model prediction output is:

[0121]

[0122] Among them, the prediction function f of each tree k (x) is defined as:

[0123] f k (x) = ω qk(x)

[0124] qk(x) is the leaf node to which the second sample x belongs in the k-th tree.

[0125] The objective function of the entire model is:

[0126]

[0127] in, is the loss function, T k is the number of leaf nodes in the kth tree, ω kj is the weight of the jth leaf node in the kth tree

[0128] Step S4: Perform energy consumption prediction based on the trained building energy consumption prediction model.

[0129] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program code.

Claims

1. A building energy consumption prediction method based on machine learning, characterized in that: include: Step S1: Obtain a benchmark building energy consumption model, wherein the benchmark building energy consumption model is established based on existing standard specifications and with reference to typical office building cases, and is configured to obtain monthly energy consumption based on the values ​​of various key parameters; Step S2: Sampling each key parameter using a Latin hypercube sampling method to obtain a plurality of first samples, obtaining a corresponding monthly energy consumption according to each first sample based on the benchmark building energy consumption model, and combining each first sample with the corresponding monthly energy consumption to generate a second sample; Step S3: using the second sample to train a building energy consumption prediction model; Step S4: Perform energy consumption prediction based on the trained building energy consumption prediction model.

2. The method for predicting building energy consumption based on machine learning according to claim 1, characterized in that: The benchmark building energy consumption model is developed based on EnergyPlus, SketchUp and Openstudio.

3. The method for predicting building energy consumption based on machine learning according to claim 1, characterized in that: The key parameters include: Building characteristic parameters, including at least the heat transfer coefficient of the exterior wall, the heat transfer coefficient of the exterior window, the heat transfer coefficient of the roof and the building orientation; The user behavior parameters include at least personnel density, lighting power density, equipment power density, cooling temperature setting value and heating temperature setting value.

4. The method for predicting building energy consumption based on machine learning according to claim 1, characterized in that: The step S2 comprises: Step S2-1: Determine the variation range of each key parameter according to literature records and standard specifications; Step S2-2: Divide the variation range of each key parameter into multiple small intervals of equal width, and randomly extract a sample value in each small interval for each key parameter, ensuring that each small interval is extracted only once, and obtain multiple non-repeating first samples by arranging and combining the sample values ​​of each key parameter between the small intervals, thereby fully covering the parameter space; Step S2-3: inputting the obtained first sample into the benchmark building energy consumption model, and generating corresponding monthly energy consumption through simulation of the benchmark building energy consumption model; Step S2-4: Combine each first sample with the corresponding monthly energy consumption to generate a second sample.

5. The method for predicting building energy consumption based on machine learning according to claim 4 is characterized in that: The sample values ​​extracted from each small interval are: Where: sv ij is the jth sample value of the ith key parameter, a i is the lower limit of the variation range of the i-th key parameter, b i is the upper limit of the variation range of the i-th key parameter, N i is the number of small intervals where the i-th key parameter is divided, ∈ ij is a random disturbance ranging from 0 to 1.

6. The method for predicting building energy consumption based on machine learning according to claim 4 is characterized in that: The step S2-4 specifically includes: Step S2-4-1: preprocessing the generated monthly energy consumption data set, wherein the preprocessing process includes converting energy consumption into energy consumption intensity; Step S2-4-2: Combine each first sample, the corresponding month and its energy consumption intensity to generate a second sample.

7. The method for predicting building energy consumption based on machine learning according to claim 6 is characterized in that: The step S3 comprises: Step S3-1: Map all second samples to the root node and define a mapping function q(x i )=0, where x i is the i-th second sample; Step S3-2: Initialize the leaf node set to be segmented L={0}, and the current leaf node set only includes the root node; Step S3-3: Select a leaf node l from the leaf node set for segmentation; Step S3-4: for a leaf node l, traverse all possible segmentation points s in the leaf node and calculate the segmentation gain Gain; Step S3-5: Select the segmentation point s with the largest segmentation gain as the segmentation point to segment the leaf node l, and update the mapping between samples and nodes: Among them, l L is the left node after the split, l R is the right node after splitting; Step S3-6: Repeat steps S3-3 to S3-5 until the maximum node segmentation depth is reached or the segmentation gain Gain ≤ 0; Step S3-6: Assign weight to each node.

8. The method for predicting building energy consumption based on machine learning according to claim 7 is characterized in that: The split gain Gain is: G=Σ i∈I g i H=∑ i∈I h i G R =GG L H R =H-H L Where: G L is the sum of the first-order gradients of all second samples in the left node, H L is the sum of the second-order gradients of all second samples in the left node, λ is the regularization parameter of the weight, γ is the regularization parameter of the number of leaf nodes, G is the sum of the first-order gradients of all second samples in node l, H is the sum of the second-order gradients of all second samples in node l, g i is the first-order gradient of the second sample i, h i is the second-order gradient of the second sample i, I is the set of all second samples of node l, G R is the sum of the first-order gradients of all second samples in the right node, H R is the sum of the second-order gradients of all second samples in the right node.

9. The method for predicting building energy consumption based on machine learning according to claim 7, characterized in that: The weight of the node in step S3-6 is: Where: j is the weight of node j, G j is the sum of the first-order gradients of all second samples in node j, H j is the sum of the first-order gradients of all second samples in node j.

10. A building energy consumption prediction device based on machine learning, comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 9 is implemented.

Images