Training method of energy consumption probability prediction model, energy consumption probability prediction method, computer equipment and readable storage medium

Abstract

The invention provides a training method of an energy consumption probability prediction model, an energy consumption probability prediction method, computer equipment and a readable storage medium, and relates to the technical field of building energy management and intelligent prediction. The method comprises the following steps: performing feature extraction on historical building data of a target building in a preset historical time period to obtain sample energy consumption related features; constructing a sample data set according to the sample energy consumption related characteristics and the real energy consumption probability; dividing the sample data set into a plurality of sub-data sets; respectively training the plurality of initial energy consumption prediction models according to the plurality of sub-data sets to obtain a plurality of trained energy consumption prediction sub-models; and according to the plurality of trained energy consumption prediction sub-models, constructing a target energy consumption prediction model. According to the method, the target energy consumption prediction model is obtained according to the multi-sample energy consumption related features, so that the prediction interval coverage degree is high, the method adapts to dynamic load changes, and then the accurate prediction requirement for building energy consumption is met.

Description

Technical Field

[0001] The present application relates to the technical field of building energy management and intelligent prediction, and specifically to a training method for an energy consumption probability prediction model, an energy consumption probability prediction method, a computer device, and a readable storage medium. Background Art

[0002] As the global ecological environment continues to deteriorate, building energy consumption has become a global concern. HVAC systems account for approximately 40% of a building's total energy consumption. Building carbon emissions contribute 25% to 33% of all carbon emissions. Furthermore, frequent high temperatures have led to a surge in electricity demand for HVAC systems in buildings, exacerbating the load on the power grid. Therefore, accurately predicting building HVAC energy consumption is crucial for both building energy conservation and the stable operation of the power grid.

[0003] Currently, most building energy simulation software relies on a fixed schedule defined by the American Society of Heating, Refrigerating and Air-Conditioning Engineers to simply define building energy use.

[0004] However, this approach ignores the randomness of weather changes and fluctuations in pedestrian flow, resulting in low coverage of the prediction interval and an error range of up to 50%-150% between predicted energy consumption and actual energy consumption. It is unable to adapt to dynamic load changes and thus cannot achieve the demand for accurate prediction of building energy consumption. Summary of the Invention

[0005] The purpose of the embodiments of the present application is to provide a training method for an energy consumption probability prediction model, an energy consumption probability prediction method, a computer device and a readable storage medium, so as to obtain a target energy consumption prediction model based on the energy consumption-related characteristics of multiple samples, so that the prediction interval coverage is high and it adapts to dynamic load changes, thereby realizing the accurate prediction demand for building energy consumption.

[0006] In order to achieve the above objectives, the technical solutions adopted in the embodiments of the present application are as follows: In a first aspect, an embodiment of the present application provides a method for training an energy consumption probability prediction model, comprising: Extracting features from historical building data of a target building in a preset historical time period to obtain sample energy consumption-related features of the preset historical time period; Constructing a sample data set based on the sample energy consumption related characteristics of the preset historical time period and the actual energy consumption probability of the HVAC in the target building during the preset historical time period; Dividing the sample data set into a plurality of sub-data sets, wherein the plurality of sub-data sets respectively correspond to a plurality of sub-time periods within the preset historical time period; Training a plurality of initial energy consumption prediction models according to the plurality of sub-data sets to obtain a plurality of trained energy consumption prediction sub-models; A target energy consumption prediction model is constructed based on the multiple trained energy consumption prediction sub-models.

[0007] Optionally, the sample energy consumption related features include: sample linear correlation features and sample nonlinear correlation features; and the feature extraction of historical building data of the target building in a preset historical time period to obtain the sample energy consumption related features of the preset historical time period includes: According to the true energy consumption probability, a preset linear correlation analysis method is used to extract features from the historical building data to obtain sample linear correlation features; According to the true energy consumption probability, a preset nonlinear correlation analysis method is used to extract features of the historical building data to obtain sample nonlinear correlation features; The sample linear correlation feature and the sample nonlinear correlation feature are fused to obtain the sample energy consumption correlation feature.

[0008] Optionally, the method of extracting features from the historical building data using a preset linear correlation analysis method based on the actual energy consumption probability to obtain sample linear correlation features includes: Using the preset linear correlation analysis method, respectively calculating the Pearson correlation coefficient of each type of characteristic data in the historical building data in the true energy consumption probability; The sample linear correlation features are screened from various types of feature data in the historical building data according to the Pearson correlation coefficient.

[0009] Optionally, the extracting features of the historical building data using a preset nonlinear correlation analysis method based on the actual energy consumption probability to obtain sample nonlinear correlation features includes: Using the preset nonlinear correlation analysis method, respectively calculating the Spearman rank correlation coefficient of each type of characteristic data in the historical building data in the true energy consumption probability; The sample nonlinear correlation features are screened from various types of feature data in the historical building data according to the Spearman rank correlation coefficient.

[0010] Optionally, each initial energy consumption prediction model integrates: an initial long short-term memory (LSTM) network and an initial conformal quantile regression (CQR) prediction model; The method of training a plurality of initial energy consumption prediction models according to the plurality of sub-data sets to obtain a plurality of trained energy consumption prediction sub-models includes: Using the initial LSTM network in each initial energy consumption prediction model to perform time series processing on the sample energy consumption related features in each sub-data set to obtain time series feature information corresponding to each sub-data set; Using the initial CQR prediction model in each initial energy consumption prediction model, perform energy consumption prediction on the time series feature information corresponding to each sub-data set to obtain a predicted energy consumption probability interval corresponding to each sub-data set; According to the actual energy consumption probability of the sub-time period corresponding to each sub-data set and the corresponding predicted energy consumption probability interval, the initial LSTM network and the initial CQR prediction model in each initial energy consumption prediction model are trained to obtain a trained energy consumption prediction sub-model, and the trained energy consumption prediction sub-model includes: a trained LSTM network and a trained CQR prediction network.

[0011] Optionally, constructing a target energy consumption prediction model based on the multiple trained energy consumption prediction sub-models includes: The multiple trained energy consumption prediction sub-models are subjected to integrated learning training to obtain the target energy consumption prediction model.

[0012] Optionally, performing ensemble learning training on the multiple trained energy consumption prediction sub-models to obtain the target energy consumption prediction model includes: Using each trained energy consumption prediction sub-model to predict the sample energy consumption related features of each sub-time period, to obtain the predicted energy consumption probability interval of each trained energy consumption prediction sub-model for each sub-time period; Obtaining a prediction deviation of each trained energy consumption prediction sub-model for each sub-time period based on the predicted energy consumption probability interval of each trained energy consumption prediction sub-model for each sub-time period and the actual energy consumption probability of each sub-time period; Constructing a prediction residual set based on the prediction deviations of the multiple trained energy consumption prediction sub-models for each sub-time period; The target energy consumption prediction model is constructed according to the prediction residual set.

[0013] In a second aspect, an embodiment of the present application provides an energy consumption probability prediction method, comprising: Extracting features from the predicted building data of the building to be predicted in a preset future time period to obtain features related to predicted energy consumption in the preset future time period; A pre-trained target energy consumption prediction model is used to perform energy consumption prediction based on the predicted energy consumption related features of the preset future time period, and a predicted HVAC energy consumption probability interval of the building to be predicted in the preset future time period is obtained.

[0014] In a third aspect, an embodiment of the present application provides a computer device comprising: a processor; a memory for storing processor executable instructions; wherein the processor runs the executable instructions to implement the training method of the energy consumption probability prediction model as described in any one of the first aspect and the energy consumption probability prediction method described in the second aspect.

[0015] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the training method of the energy consumption probability prediction model as described in any one of the first aspects and the steps of the energy consumption probability prediction method as described in the second aspect are executed.

[0016] Compared with the existing technology, the energy consumption probability prediction model training method, energy consumption probability prediction method, computer device and readable storage medium provided by this application have the following technical effects: The present application provides a training method for an energy consumption probability prediction model, an energy consumption probability prediction method, a computer device, and a readable storage medium, relating to the field of building energy management and intelligent prediction technology. The method includes: extracting features from historical building data of a target building in a preset historical time period to obtain sample energy consumption-related features of the preset historical time period, which helps to reduce the dimension of subsequent sample data, avoid interference of irrelevant data on the subsequent target energy consumption prediction model, and improve the training efficiency and accuracy of the target energy consumption prediction model. Based on the sample energy consumption-related features of the preset historical time period and the actual energy consumption probability of the HVAC in the target building in the preset historical time period, a sample data set is constructed to ensure that all data in the data set are related to energy consumption prediction; the sample data set is divided into multiple sub-data sets, each of which corresponds to multiple sub-time periods within the preset historical time period, so that each sub-data set has energy consumption characteristics of a specific time period; based on the multiple sub-data sets, multiple initial energy consumption prediction models are trained to obtain multiple trained energy consumption prediction sub-models, each of which has good prediction capabilities in its corresponding sub-time period. These energy consumption prediction sub-models are integrated to construct a target energy consumption prediction model, which can combine the advantages of each energy consumption prediction sub-model and adapt to a wider range of energy consumption scenarios. At the same time, the target energy consumption prediction model is obtained based on the energy consumption-related characteristics of multiple samples, so that the prediction interval coverage is high and it can adapt to dynamic load changes, thereby realizing the accurate prediction needs of building energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 A schematic diagram of the structure of a training system for an energy consumption probability prediction model provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of a computer device provided in an embodiment of the present application; Figure 3 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 1 ; Figure 4 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 2 ; Figure 5 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 3 ; Figure 6 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 4 ; Figure 7 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 5 ; Figure 8 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 6 ; Figure 9 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 7 ; Figure 10 A flow chart of an energy consumption probability prediction method provided in an embodiment of the present application; Figure 11 A schematic diagram of the structure of a training device for an energy consumption probability prediction model provided in an embodiment of the present application; Figure 12 A schematic diagram of the structure of an energy consumption probability prediction device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0019] The following will be combined with the accompanying drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. The following embodiments of the present application are described in detail with reference to the accompanying drawings. The following embodiments and features in the embodiments can be combined with each other unless there is any conflict.

[0020] Figure 1 This is a structural diagram of a training system for an energy consumption probability prediction model provided in an embodiment of the present application. Figure 1 As shown, the energy consumption probability prediction model training system 100 may include: a feature extraction and sample data set construction unit 110 , an energy consumption prediction sub-model construction unit 120 and a target energy consumption prediction model construction unit 130 .

[0021] The feature extraction and sample data set construction unit 110 is used to extract features from the historical building data, construct a sample data set, and obtain multiple sub-data sets Ie. The energy consumption prediction sub-model construction unit 120 is used to respectively perform multiple initial energy consumption prediction models based on the multiple sub-data sets Ie. The target energy consumption prediction model construction unit 130 is used to obtain multiple trained energy consumption prediction sub-models LCQR according to the multiple trained energy consumption prediction sub-models. , construct the target energy consumption prediction model LECQR.

[0022] It should be noted that the training system 100 of the above-mentioned energy consumption probability prediction model can be integrated into a computer device 200. The computer device 200 can be selected according to actual conditions. For example, the computer device 200 can be a computer, etc.

[0023] Further, Figure 2 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. Figure 2 As shown, the computer device 200 may include: a processor 210 and a memory 220 .

[0024] The memory 220 stores machine-executable instructions that can be executed by the processor 210. That is, when the computer device 200 is running, the machine-readable instructions are executed, and the processor 210 and the memory 220 communicate via a bus. The processor 210 can execute the machine-executable instructions to implement the energy consumption probability prediction model training method and the energy consumption probability prediction method.

[0025] The memory 220, processor 210, and various bus components are electrically connected to each other, directly or indirectly, to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines. The mobile storage device includes at least one software functional module that can be stored in the memory 220 in the form of software or firmware or embedded in the operating system (OS) of the microgrid controller. The processor 210 is configured to execute executable modules stored in the memory 220, such as the training method for the energy consumption probability prediction model of the mobile storage medium and the software functional modules and computer programs included in the energy consumption probability prediction method.

[0026] Among them, the memory 220 can be, but is not limited to, random access memory (RAM), read only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0027] The training method of the energy consumption probability prediction model and the energy consumption probability prediction method provided in the embodiments of the present application can be executed by the processor 210 in the computer device 200. The following continues to explain and illustrate in detail the training method of the energy consumption probability prediction model and the energy consumption probability prediction method provided in the above embodiments of the present application in combination with the accompanying drawings. Figure 3 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 1 .like Figure 3 As shown, the method may include: S301. Extract features from historical building data of a target building in a preset historical time period to obtain sample energy consumption-related features of the preset historical time period.

[0028] The preset historical time period can be selected according to actual conditions. For example, the preset historical time period can be the past three months.

[0029] In one possible implementation, please refer to the above Figure 1, feature extraction is performed on the time data and environmental data from the historical building data of the target building during a preset historical period to obtain the sample energy consumption-related features of the preset historical period. Environmental data may include: ambient temperature, humidity, wind speed, wind direction, and pedestrian flow. Time data may include: day of the week, hour, holidays, and weekdays.

[0030] S302: Construct a sample data set based on the sample energy consumption related characteristics of the preset historical time period and the actual energy consumption probability of the HVAC in the target building during the preset historical time period.

[0031] In one possible implementation method, the sample energy consumption-related features of a preset historical time period are combined with the actual energy consumption probability of the HVAC in the target building during the preset historical time period to construct a sample data set. This can enable the subsequent target energy consumption prediction model to consider both the relevant factors of energy consumption and the uncertainty of energy consumption during the training process, which helps to improve the target energy consumption prediction model's ability to predict energy consumption probability.

[0032] It should be noted that the actual energy consumption probability of the preset historical time period here is for the purpose of explaining that the sample energy consumption related features obtained in the preset historical time period are features associated with energy consumption prediction.

[0033] S303: Divide the sample data set into multiple sub-data sets, where the multiple sub-data sets respectively correspond to multiple sub-time periods within a preset historical time period.

[0034] In one possible implementation, the sample data set is divided according to a preset historical time period to obtain multiple sub-data sets, which may include: test set samples , training set samples ~e and validation set samples The multiple sub-data sets correspond to multiple sub-time periods within a preset historical time period.

[0035] Among them, the test set samples Data used to evaluate the performance of the energy consumption prediction sub-model. Test set sample In the test set samples Used after training and validation are completed to evaluate test set samples The generalization ability of the training set samples ~e is used to train the data of the initial energy consumption prediction model. The initial energy consumption prediction model is trained by learning the training set samples. The relationship between the features and labels in ~e is used to construct internal parameters. The data used to adjust the hyperparameters of the energy consumption prediction sub-model and evaluate the performance of the energy consumption prediction model. It is usually used during training to prevent overfitting.

[0036] S304 : training a plurality of initial energy consumption prediction models respectively according to the plurality of sub-data sets to obtain a plurality of trained energy consumption prediction sub-models.

[0037] In one possible implementation, multiple initial energy consumption prediction models are respectively Training is performed so that each initial energy consumption prediction model (like Figure 1 in ) Focusing on the energy consumption characteristics of different sub-time periods helps improve the initial energy consumption prediction model The adaptability to energy consumption changes in different time periods is used to obtain multiple trained energy consumption prediction sub-models LCQR.

[0038] S305: Construct a target energy consumption prediction model based on the multiple trained energy consumption prediction sub-models.

[0039] In one possible implementation, based on the test set samples The energy consumption prediction sub-model LCQR is input to obtain the outputs of multiple trained energy consumption prediction sub-models LCQR, and then the target energy consumption prediction model LECQR is obtained.

[0040] The present application provides a training method for an energy consumption probability prediction model, which extracts features from the historical building data of the target building in a preset historical time period to obtain sample energy consumption related features of the preset historical time period, which helps to reduce the dimension of subsequent sample data, avoid the interference of irrelevant data on the subsequent target energy consumption prediction model, and improve the training efficiency and accuracy of the target energy consumption prediction model. According to the sample energy consumption related features of the preset historical time period and the actual energy consumption probability of the HVAC in the target building in the preset historical time period, a sample data set is constructed to ensure that the data in the data set are all related to energy consumption prediction; the sample data set is divided into multiple sub-data sets, and the multiple sub-data sets correspond to multiple sub-time periods within the preset historical time period, so that each sub-data set has energy consumption characteristics of a specific time period; according to the multiple sub-data sets, multiple initial energy consumption prediction models are trained respectively to obtain multiple trained energy consumption prediction sub-models, and each energy consumption prediction sub-model has good prediction capabilities in its corresponding sub-time period. These energy consumption prediction sub-models are integrated to construct a target energy consumption prediction model, which can combine the advantages of each energy consumption prediction sub-model and adapt to a wider range of energy consumption scenarios. At the same time, the target energy consumption prediction model is obtained based on the energy consumption-related characteristics of multiple samples, so that the prediction interval coverage is high and it can adapt to dynamic load changes, thereby realizing the accurate prediction needs of building energy consumption.

[0041] Optionally, the above-mentioned sample energy consumption-related features include: sample linear correlation features and sample non-linear correlation features. Among them, the sample energy consumption-related features may include: ambient temperature, daily maximum temperature, daily minimum temperature, perceived temperature, relative humidity, dew point temperature, daily total rainfall, 10-minute average wind speed, gust wind speed, wind direction, wind direction degree, hours of the day, days of the week, days of the month, week number, holiday characteristics, lighting and socket energy consumption, pedestrian flow, and historical HVAC energy consumption, etc. Among them, the sample linear correlation feature is a linear sample feature among the above-mentioned sample energy consumption-related features, for example, the operating time is a sample linear correlation feature; the sample non-linear correlation feature is a non-linear sample feature among the above-mentioned sample energy consumption-related features, for example, the ambient temperature sample non-linear correlation feature.

[0042] Figure 4 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 2 .like Figure 4 As shown, in the above method, feature extraction is performed on the historical building data of the target building in a preset historical time period to obtain the sample energy consumption related features of the preset historical time period, which may include: S401. Based on the actual energy consumption probability, a preset linear correlation analysis method is used to extract features from the historical building data to obtain sample linear correlation features.

[0043] The preset linear correlation analysis method may be selected according to actual conditions. For example, the preset linear correlation analysis method may be selected as a Pearson correlation coefficient analysis method.

[0044] In one possible implementation method, based on the actual energy consumption probability, a preset linear correlation analysis method (such as the Pearson correlation coefficient analysis method) is used to extract features from the time data and environmental data in the historical building data to obtain sample linear correlation features. This is to eliminate the interference of some accidental factors based on the actual energy consumption probability, making the extracted sample linear correlation features more representative and reliable, thereby providing an accurate basis for subsequent energy consumption predictions.

[0045] S402. Based on the actual energy consumption probability, a preset nonlinear correlation analysis method is used to extract features from the historical building data to obtain sample nonlinear correlation features.

[0046] The preset nonlinear correlation analysis method may be selected according to actual conditions. For example, the preset nonlinear correlation analysis method may be selected as the Spearman rank correlation coefficient analysis method.

[0047] In one possible implementation method, in order to comprehensively evaluate the correlation between features, a preset nonlinear correlation analysis method (such as the Spearman rank correlation coefficient analysis method) can be used to extract features of the time data and environmental data in the historical building data based on the actual energy consumption probability, and obtain sample nonlinear correlation features to better adapt to the complex and changeable actual situation, thereby improving the model's adaptability to energy consumption prediction under different working conditions.

[0048] S403: Fusing the sample linear correlation features and the sample nonlinear correlation features to obtain the sample energy consumption correlation features.

[0049] One possible implementation involves fusing sample linear correlation features with sample nonlinear correlation features. The linear correlation features provide a stable and easily understood foundation for energy consumption relationships, while the nonlinear correlation features complement the description of complex energy consumption relationships. The resulting sample energy consumption correlation features provide richer information for the subsequent target energy consumption prediction model, helping to improve the model's energy consumption prediction accuracy and reduce prediction errors. Whether under routine operating conditions or in extreme weather conditions, the fused sample energy consumption correlation features enable the target energy consumption prediction model to make more accurate energy consumption forecasts.

[0050] The present application provides a training method for an energy consumption probability prediction model, wherein the sample energy consumption related features can be composed of sample linear related features and sample nonlinear related features; based on the true energy consumption probability, a preset linear correlation analysis method is used to extract features from historical building data to obtain sample linear related features; based on the true energy consumption probability, a preset nonlinear correlation analysis method is used to extract features from historical building data to obtain sample nonlinear related features; the sample linear related features and the sample nonlinear related features are fused to obtain sample energy consumption related features. Thus, the sample energy consumption related features obtained in the present application make the target energy consumption prediction model no longer limited to dependence on specific linear or nonlinear relationships, and can better adapt to different types of buildings, different usage scenarios and different environmental conditions. This enhances the generalization ability of the target energy consumption prediction model, enabling it to effectively perform energy consumption prediction in a wider range.

[0051] Figure 5 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 3 .like Figure 5 As shown, in the above method, based on the actual energy consumption probability, a preset linear correlation analysis method is used to extract features from historical building data to obtain sample linear correlation features, which may include: S501. Using a preset linear correlation analysis method, calculate the Pearson correlation coefficient between the actual energy consumption probability and various characteristic data in the historical building data.

[0052] In one possible implementation method, a preset linear correlation analysis method (such as the Pearson correlation coefficient analysis method) is used to calculate the Pearson correlation coefficients of the actual energy consumption probability and various characteristic data in the historical building data. .

[0053] The Pearson correlation coefficient can be expressed according to the following formula (1): Formula (1) Wherein, i and j in the above formula (1) are both positive integers from 1 to 19, that is, . and Represent the values ​​of different sample energy consumption related features. Specifically, They represent ambient temperature, daily maximum temperature, daily minimum temperature, perceived temperature, relative humidity, dew point temperature, daily total rainfall, 10-minute average wind speed, gust wind speed, wind direction, wind direction degree, hour of the day, day of the week, day of the month, week number, holiday characteristics, lighting and socket energy consumption, pedestrian flow, and HVAC historical energy consumption.

[0054] In the above formula (1), n ​​represents the number of samples of energy consumption related features of each sample (i.e., the number of observations), and and Respectively and The average value of .

[0055] S502. Filter sample linear correlation features from various feature data in the historical building data based on the Pearson correlation coefficient.

[0056] In one possible implementation, the linear correlation features of samples are screened from various feature data in the historical building data according to the Pearson correlation coefficient expressed in the above formula (1). The Pearson correlation coefficient is a common method for measuring the strength of the linear relationship between variables, but it is sensitive to outliers.

[0057] The training method of the energy consumption probability prediction model provided by the present application adopts a preset linear correlation analysis method to respectively calculate the Pearson correlation coefficient between the actual energy consumption probability and various feature data in the historical building data; based on the Pearson correlation coefficient, sample linear correlation features are screened from various feature data in the historical building data, so that by setting a certain Pearson correlation coefficient threshold, features that are linearly correlated with energy consumption can be effectively screened out from numerous historical building data features, thereby avoiding the interference of irrelevant features on the subsequent target energy consumption prediction model, and enabling the target energy consumption prediction model to focus on factors that truly affect energy consumption. Therefore, the target energy consumption prediction model that can be constructed based on the screened sample linear correlation features in the present application can more accurately capture changes in the linear relationship between energy consumption and related features.

[0058] Figure 6 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 4 .like Figure 6 As shown, in the above method, based on the actual energy consumption probability, a preset nonlinear correlation analysis method is used to extract features from historical building data to obtain sample nonlinear correlation features, which may include: S601. Using a preset nonlinear correlation analysis method, respectively calculate the Spearman rank correlation coefficient between the actual energy consumption probability and various characteristic data in the historical building data.

[0059] In one possible implementation method, a preset nonlinear correlation analysis method (such as the Spearman rank correlation coefficient analysis method) is used to calculate the Spearman rank correlation coefficient between the actual energy consumption probability and various characteristic data in the historical building data. .

[0060] Among them, the Spearman rank correlation coefficient can be expressed according to the following formula (2): Formula (2) in, represents the rank difference of each pair of observations. It can be expressed by the following formula (3): Formula (3) The result of the above formula (3) can be input into the above formula (2) to obtain the Spearman rank correlation coefficient .

[0061] S602. Filter sample nonlinear correlation features from various feature data in the historical building data based on the Spearman rank correlation coefficient.

[0062] In one possible implementation, the nonlinear correlation features of samples are screened from various characteristic data in the historical building data according to the Spearman rank correlation coefficient expressed in the above formula (2). The Spearman rank correlation coefficient mainly measures the monotonic relationship between two variables and does not require the data to satisfy a linear relationship or a specific distribution form.

[0063] The training method of the energy consumption probability prediction model provided by the present application adopts a preset nonlinear correlation analysis method to respectively calculate the Spearman rank correlation coefficient between the actual energy consumption probability and various characteristic data in the historical building data; based on the Spearman rank correlation coefficient, the sample nonlinear correlation features are screened from various characteristic data in the historical building data, so that by screening out the sample nonlinear correlation features, richer input information can be provided for the target energy consumption prediction model. Therefore, when constructing the target energy consumption prediction model, the present application considers both linear and nonlinear correlation sample features, which can make the target energy consumption prediction model more optimized. The target energy consumption prediction model can process linear and nonlinear relationships separately according to the characteristics of different features, thereby improving the prediction accuracy and stability of the target energy consumption prediction model.

[0064] Optionally, each of the above initial energy consumption prediction models integrates: an initial long short-term memory (LSTM) network and an initial shape-preserving quantile regression (CQR) prediction model.

[0065] The initial Long Short-Term Memory (LSTM) network is a new architecture for recurrent neural networks designed to overcome the problem of vanishing or exploding gradients. The internal structure of the initial LSTM network is based on a series of interconnected units.

[0066] Figure 7 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 5 .like Figure 7 As shown, in the above method, multiple initial energy consumption prediction models are trained according to multiple sub-data sets to obtain multiple trained energy consumption prediction sub-models, which may include: S701. Use the initial LSTM network in each initial energy consumption prediction model to perform time series processing on the sample energy consumption related features in each sub-data set to obtain the time series feature information corresponding to each sub-data set.

[0067] In one possible implementation, the initial LSTM network in each initial energy consumption prediction model is used to perform time series processing on the sample energy consumption related features in each sub-dataset to obtain the time series feature information corresponding to each sub-dataset.

[0068] The initial LSTM network is equipped with a cell state feature to store information, which can learn temporal information across multiple time scales. Each cell in the initial LSTM network contains three main gates: forget gate, input gate, and output gate. The forget gate controls the previous moment. The input gate identifies the historical features that are crucial for predicting HVAC energy consumption at the current time step t. The output gate controls the current hidden state The initial long short-term memory LSTM network can be expressed by the following formula (4): Formula (4) in, Represents the input information at time step t; , and Represents the states of the input gate, forget gate, and output gate at time step t respectively; the activation function of the fully connected layer is the sigmoid function ; represents the hidden state from the previous time step t-1; w represents the weights of different gates; b represents the bias terms of different gates.

[0069] The initial LSTM network is also equipped with a candidate unit, which provides new information to the memory unit. This information may replace or supplement the existing information in the memory unit. The calculation of this candidate unit is similar to the gate calculation steps in the above formula (4), but uses the tangent function. At time step t, it can be expressed by the following formulas (5), (6), and (7): Formula (5) Formula (6) Formula (7) in, Represents the generation of candidate information in the candidate unit; Indicates the unit state at the current moment; Indicates the hidden state at the current moment.

[0070] S702 : Using the initial CQR prediction model in each initial energy consumption prediction model, perform energy consumption prediction on the time series feature information corresponding to each sub-data set to obtain a predicted energy consumption probability interval corresponding to each sub-data set.

[0071] In one possible implementation, in the training set samples Based on ~e, each initial energy consumption prediction model is used The initial CQR (Conformalized Quantile Regression, shape-preserving quantile regression model) prediction model in each sub-dataset training set sample The corresponding time series feature information is used to predict energy consumption and obtain the training set samples in each sub-dataset The corresponding predicted energy consumption probability interval. Among them, the initial CQR prediction model is a probability prediction model based on statistical theory, which combines the advantages of shape-preserving prediction and quantile regression (QR) model. It can construct an effective prediction interval for heteroscedastic data without assuming the data distribution. In addition, it can adapt to local changes in the data. Therefore, it is suitable for the probabilistic prediction of HVAC energy consumption with complex data distribution. Among them, the initial CQR prediction model can be expressed by the following formula (8) using shape-preserving prediction and quantile regression: Formula (8) Among them, i ; and represent the upper quantile and lower quantile respectively; f represents the energy consumption prediction sub-model LCQR; represents the input features, represents the true value; represents the lower quantile regression function; represents the upper quantile regression function.

[0072] in, . Where a represents the confidence interval.

[0073] According to the above formula (8), the energy consumption of the time series feature information corresponding to each sub-data set is predicted, and the predicted energy consumption probability interval corresponding to each sub-data set is obtained. . It can be expressed by the following formula (9).

[0074] Formula (9) Among them, the above formula (9) It can be expressed by the following formula (10), where the prediction deviation in the above formula (9) is It can be expressed by the following formula (11): Formula (10) Formula (11) S703. According to the actual energy consumption probability of the sub-time period corresponding to each sub-data set and the corresponding predicted energy consumption probability interval, the initial LSTM network and the initial CQR prediction model in each initial energy consumption prediction model are trained to obtain a trained energy consumption prediction sub-model.

[0075] Among them, the post-training energy consumption prediction sub-model includes: post-training LSTM network and post-training CQR prediction network.

[0076] In one possible implementation, the actual energy consumption probability of each sub-data set corresponding to the sub-time period and the corresponding predicted energy consumption probability interval are used to calculate the energy consumption probability of each initial energy consumption prediction model. The initial LSTM network and the initial CQR prediction model in are trained to obtain a trained energy consumption prediction sub-model LCQR.

[0077] For example, refer to Figure 1 , the training set samples The features in ~e are input into the initial long short-term memory LSTM network to obtain the time series feature information corresponding to each sub-dataset. Then, the time series feature information corresponding to each sub-dataset is input into the shape-preserving prediction and quantile regression to obtain the interval corresponding to the above formula (8). Then, according to the above formula (8), the energy consumption of the time series feature information corresponding to each sub-dataset is predicted to obtain the predicted energy consumption probability interval corresponding to each sub-dataset. , and then form a trained energy consumption prediction sub-model LCQR.

[0078] This application provides a training method for probabilistic energy consumption prediction models. Each initial energy consumption prediction model integrates an initial long short-term memory (LSTM) network and an initial shape-preserving quantile regression (CQR) prediction model. The initial LSTM network excels at processing time series data and can capture the time series information in sample energy consumption-related features. The initial shape-preserving quantile regression (CQR) prediction model can perform probabilistic energy consumption predictions and provide prediction intervals. This initial energy consumption prediction model enables the model to both process time series data and perform probabilistic energy consumption predictions, improving overall prediction capabilities. First, the initial LSTM network in each initial energy consumption prediction model performs time series processing on the energy consumption-related features of the samples in each sub-dataset, obtaining the time series feature information corresponding to each sub-dataset. Then, the initial CQR prediction model in each initial energy consumption prediction model is used to perform energy consumption prediction on the time series feature information corresponding to each sub-dataset, obtaining the predicted energy consumption probability interval corresponding to each sub-dataset, thereby increasing the reliability of the prediction results. Finally, based on the actual energy consumption probability and the corresponding predicted energy consumption probability interval for each sub-time period of each sub-dataset, the initial LSTM network and initial CQR prediction model in each initial energy consumption prediction model are trained to obtain a trained energy consumption prediction sub-model. This trained energy consumption prediction sub-model is continuously optimized and can be composed of the trained LSTM network and the trained CQR prediction network. Since each energy consumption prediction sub-model is trained based on a different sub-dataset, it can learn energy consumption characteristics under different time periods and different operating conditions, further improving the energy consumption prediction sub-model's ability to predict building energy consumption in different scenarios and adapting to dynamic changes in building energy consumption.

[0079] Optionally, in the above method, constructing a target energy consumption prediction model based on multiple trained energy consumption prediction sub-models may include: The multiple trained energy consumption prediction sub-models are trained through ensemble learning to obtain the target energy consumption prediction model.

[0080] In one possible implementation, the energy consumption prediction probability interval is obtained by using the above formula (9) for multiple trained energy consumption prediction sub-models LCQR: , then for each training set sample Corresponding predicted energy consumption probability interval Perform ensemble learning training, for example, for each training set sample Corresponding predicted energy consumption probability interval Calculate the mean and obtain multiple mean predicted energy consumption probability intervals , and then predict the energy consumption probability interval based on multiple means Generate the target energy consumption prediction model LECQR.

[0081] The training method for the energy consumption probability prediction model provided in this application performs ensemble learning training on multiple trained energy consumption prediction sub-models to obtain a target energy consumption prediction model. Therefore, this application may train different trained energy consumption prediction sub-models based on different feature subsets, data subsets, or using different algorithm parameters. Through ensemble learning, these sub-models can complement each other, thereby improving the overall prediction accuracy of energy consumption in different scenarios.

[0082] Figure 8 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 6 .like Figure 8 As shown, in the above method, multiple trained energy consumption prediction sub-models are integrated and trained to obtain a target energy consumption prediction model, which may include: S801. Use each trained energy consumption prediction sub-model to predict the sample energy consumption related features of each sub-time period to obtain the predicted energy consumption probability interval of each trained energy consumption prediction sub-model for each sub-time period.

[0083] In one possible implementation, based on the validation set samples , use each trained energy consumption prediction sub-model LCQR to predict the energy consumption related features of samples in each sub-time period, and obtain the validation set samples of each trained energy consumption prediction sub-model Forecast energy consumption probability interval for each sub-time period .

[0084] S802. Obtain the prediction deviation of each trained energy consumption prediction sub-model for each sub-time period based on the predicted energy consumption probability interval of each trained energy consumption prediction sub-model for each sub-time period and the actual energy consumption probability of each sub-time period.

[0085] In one possible implementation, the energy consumption probability interval of each sub-time period is predicted according to each trained energy consumption prediction sub-model LCQR. , and the actual energy consumption probability of each sub-time period, the prediction deviation of each trained energy consumption prediction sub-model LCQR for each sub-time period can be obtained according to the following formula (12) and the following formula (13): .

[0086] Formula (12) Formula (13) in, and .

[0087] S803: Construct a prediction residual set based on the prediction deviations of the multiple trained energy consumption prediction sub-models for each sub-time period.

[0088] S804: Construct a target energy consumption prediction model based on the prediction residual set.

[0089] In one possible implementation, the prediction deviation of each sub-time period is obtained based on multiple trained energy consumption prediction sub-models. , construct the prediction residual set [ , ], and since the prediction interval generated by the target energy consumption prediction model LECQR is constructed by LECQR using the estimated quantile function and absolute consistency score and asymmetric consistency score, then according to the prediction residual set [ , ], the energy consumption prediction interval of the target energy consumption prediction model LECQR is obtained using the following formula (14): .

[0090] Formula (14) in, is an aggregate function; represents the residual set.

[0091] The residual set The T predicted values ​​can be used to update the corresponding T original values ​​in the residual set, that is, , .

[0092] The training method of the energy consumption probability prediction model provided in the present application uses each trained energy consumption prediction sub-model to predict the sample energy consumption related features of each sub-time period, and obtains the predicted energy consumption probability interval of each trained energy consumption prediction sub-model for each sub-time period; according to the predicted energy consumption probability interval of each trained energy consumption prediction sub-model for each sub-time period and the actual energy consumption probability of each sub-time period, the prediction deviation of each trained energy consumption prediction sub-model for each sub-time period is obtained, so that by comparing the predicted energy consumption probability interval and the actual energy consumption probability of each trained energy consumption prediction sub-model for each sub-time period, the prediction deviation of each energy consumption prediction sub-model in different time periods can be accurately obtained. This comparison method based on probability intervals reflects the accuracy of model prediction more comprehensively than the comparison of single values ​​because it takes into account the uncertainty of energy consumption. Then, based on the prediction deviations of multiple trained energy consumption prediction sub-models for each sub-time period, a prediction residual set is constructed. This residual set contains the error information of each sub-model in different time periods, reflecting the advantages and disadvantages of different energy consumption prediction sub-models in specific time periods. Based on the prediction residual set, a target energy consumption prediction model is constructed, leveraging the information in the residual set to adjust and optimize the model. Thus, the target energy consumption prediction model provided by this application combines the advantages of each energy consumption prediction sub-model, reducing prediction errors and improving overall prediction accuracy. In practical applications, more accurate energy consumption predictions can help building energy management systems rationally arrange energy supply and equipment operation, thereby reducing energy costs.

[0093] To facilitate understanding of the above-mentioned training method of the energy consumption probability prediction model, the embodiment of the present application further provides an example of the process of the training method of the energy consumption probability prediction model, which is further described below with reference to the accompanying drawings. Figure 9 A schematic diagram of a training method for an energy consumption probability prediction model provided in an embodiment of the present application Figure 7 .like Figure 9 As shown, the embodiment of the present application provides a schematic diagram Figure 7 This may include: S901. Extract features from historical building data of a target building in a preset historical time period to obtain sample energy consumption-related features of the preset historical time period.

[0094] Specifically, refer to Figure 1 According to the Pearson correlation coefficient analysis method, refer to the above formula (1) to screen out the sample linear correlation features from the various feature data of the historical building data of the target building in the preset historical time period. At the same time, according to the Spearman rank correlation coefficient analysis method, refer to the above formula (2) to screen out the sample nonlinear correlation features from the various feature data of the historical building data of the target building in the preset historical time period. Then, the sample linear correlation features and the sample nonlinear correlation features are sorted and fused according to the preset time period or preset time point to obtain the sample energy consumption related features.

[0095] S902: Construct a sample data set based on the sample energy consumption related characteristics of the preset historical time period and the actual energy consumption probability of the HVAC in the target building during the preset historical time period, and divide the sample data set into multiple sub-data sets.

[0096] Specifically, the sample energy consumption related features of the preset historical time period are combined with the real energy consumption probability to construct a sample data set, and the sample data set is divided according to the preset historical time period to obtain multiple sub-data sets, which may include: test set samples , training set samples ~e and validation set samples .

[0097] S903 . Based on the multiple sub-data sets, train the multiple initial energy consumption prediction models respectively to obtain multiple trained energy consumption prediction sub-models.

[0098] Specifically, according to the training set samples in multiple sub-datasets ~e is input into the initial LSTM network, and the training set samples The energy consumption related features of the samples in ~e are processed in time series to obtain the time series feature information corresponding to each sub-dataset (refer to the above formula (4)). Then, the time series feature information corresponding to each sub-dataset is combined with the shape-preserving prediction and quantile regression models to obtain multiple initial CQR prediction models. , then multiple initial CQR prediction models The output results are stored in the validation set samples , based on the validation set samples Multiple trained energy consumption prediction sub-models LCQR are obtained.

[0099] S904: Construct a target energy consumption prediction model based on the multiple trained energy consumption prediction sub-models.

[0100] Specifically, the training set samples ~e is input into multiple trained energy consumption prediction sub-models LCQR, and the output results of multiple trained energy consumption prediction sub-models LCQR are obtained according to the above formula (9), that is, the training set samples The predicted energy consumption probability interval corresponding to ~e , then for each training set sample Corresponding predicted energy consumption probability interval Perform mean processing (refer to the above formula (12) and formula (13)) to obtain the prediction residual set [ , ], and then based on the prediction residual set [ , ], the target energy consumption prediction model LECQR and the output result of the target energy consumption prediction model LECQR are obtained through the above formula (14), that is, the energy consumption prediction interval .

[0101] The training method of the energy consumption probability prediction model provided by the present application extracts features from the historical building data of the target building in a preset historical time period to obtain sample energy consumption related features of the preset historical time period, constructs a sample data set based on the sample energy consumption related features of the preset historical time period and the actual energy consumption probability of the HVAC in the target building in the preset historical time period, and divides the sample data set into multiple sub-data sets; based on the multiple sub-data sets, trains multiple initial energy consumption prediction models respectively to obtain multiple trained energy consumption prediction sub-models; and constructs a target energy consumption prediction model based on the multiple trained energy consumption prediction sub-models. Therefore, the present application comprehensively considers the impact of different factors on HVAC energy consumption, adopts the fusion of linear correlation analysis method and nonlinear correlation analysis method, and screens out various features that have a great impact on HVAC energy consumption. And when performing probabilistic prediction of HVAC energy consumption, the present application introduces an integrated learning training strategy and improves the residual set of the prediction interval to construct a HVAC energy consumption probability prediction model LECQR that adapts to different time series lengths to obtain a high-precision prediction interval, thereby achieving the accurate prediction demand for building energy consumption.

[0102] Optionally, Figure 10 This is a flow chart of an energy consumption probability prediction method provided in an embodiment of the present application. Figure 10 As shown, the energy consumption probability prediction method may include: S1001. Extract features from predicted building data of a building to be predicted in a preset future time period to obtain predicted energy consumption-related features for the preset future time period.

[0103] S1002: Using a pre-trained target energy consumption prediction model, perform energy consumption prediction based on the predicted energy consumption related features of a preset future time period, and obtain a predicted HVAC energy consumption probability interval of the building to be predicted within the preset future time period.

[0104] In one possible implementation, feature extraction is first performed on the predicted building data for a preset future time period of the building to be predicted, so as to screen out key factors related to energy consumption from a large amount of building data and obtain the predicted energy consumption-related features for the preset future time period. Since the extracted predicted energy consumption-related features are for the preset future time period, these features can reflect the building's operating conditions and environmental conditions that may occur in the future. Then, the pre-trained target energy consumption prediction model LECQR is used to perform energy consumption prediction based on the extracted predicted energy consumption-related features, and the predicted HVAC energy consumption probability interval of the building to be predicted in the preset future time period can be obtained. Because the target energy consumption prediction model LECQR has undergone multiple rounds of training and optimization, it combines the advantages of multiple trained energy consumption prediction sub-models and can fully consider the complex relationships between various influencing factors. Therefore, it can more accurately give the possible range of energy consumption and provide a scientific basis for energy management.

[0105] For example, continue to refer to Figure 1 , the test set samples The sample energy consumption related features in the training are input into the energy consumption prediction sub-model LCQR, and then the output result of the energy consumption prediction sub-model LCQR is input into the target energy consumption prediction model LECQR to obtain the output result of the target energy consumption prediction model LECQR, that is, the energy consumption prediction interval .

[0106] Furthermore, according to the predicted energy consumption probability interval , select Prediction Intervals Coverage Probability (PICP), Prediction Intervals Normalized Average Width (PINAW), and Coverage Width-based Criterion (CWC) to predict the energy consumption probability interval to verify.

[0107] The interval prediction coverage measures the coverage of the prediction interval, ensuring that the true value has sufficient probability to fall within the prediction interval, which can be expressed by the following formula (15): Formula (15) Wherein, K in the above formula (15) can be expressed by the following formula (16): Formula (16) Among them, the above formula (16) is the predicted mean, is the prediction interval.

[0108] In the evaluation of interval prediction coverage, the interval prediction coverage should be as close as possible to the preset confidence level (such as 95%), indicating that the prediction interval has a high coverage.

[0109] The normalized average width of the prediction interval is used to calculate the relative average width, which measures the width of the prediction interval and ensures that the prediction interval is not too wide and loses its practical significance. It can be expressed by the following formula (17): Formula (17) In the evaluation of the normalized mean width of prediction intervals, PINAW should be as small as possible, indicating a narrower prediction interval and higher accuracy.

[0110] The coverage width criterion integrates PICP and PINAW to balance coverage and width, providing a more comprehensive prediction interval performance verification. It can be expressed by the following formula (18): Formula (18) In the evaluation of the coverage width standard, CWC should be as small as possible, which means that the prediction interval width is as narrow as possible while meeting the preset coverage rate. It should be noted that in this application, the output result of the target energy consumption prediction model LECQR, that is, the energy consumption prediction interval In the evaluation, CWC needs to be as large as possible to make the energy consumption prediction interval The wider.

[0111] For example, this application also provides a specific application embodiment. The historical building data of the target building in the preset historical time period is the real data of a public building in Macau from March 1, 2021 to September 30, 2021. The historical building data constructs a sample dataset, and the sample dataset is divided into multiple sub-datasets, which are divided as follows: the training set sample contains data from March to August, and the training set sample contains data from March to August. ~e indicates that the validation set samples contain data from the first half of September 1st to 15th. The test set samples contain data from the second half of September 16 to 30, and are represented by express.

[0112] Furthermore, the training set samples ~e is divided into three subsets, each represented by a subset count b, corresponding to the following time periods: March to April, May to June, and July to August. According to the above formulas (1) and (2), the main correlation features are lighting and socket energy consumption. This is followed by the correlation with ambient temperature, daily maximum temperature, daily minimum temperature, perceived temperature, and dew point temperature. However, the correlation with relative humidity, daily total rainfall, 10-minute average wind speed, gust wind speed, wind direction, and week number is relatively weak, and their impact on energy consumption is negligible. Therefore, ambient temperature, daily maximum temperature, daily minimum temperature, perceived temperature, dew point temperature, hour of the day, day of the week, day of the month, holiday, lighting, socket energy consumption, HVAC energy consumption history, and traffic volume are selected as the main features affecting energy consumption.

[0113] Furthermore, the present application also provides a table showing the preset effects of each energy consumption prediction sub-model LCQR compared to the target energy consumption prediction model LECQR proposed by the present invention. Table 1 shows the prediction results of each energy consumption prediction sub-model and the target energy consumption prediction model provided in the embodiment of the present application. As shown in Table 1.

[0114]

[0115] Table 1 shows that in the evaluation of interval prediction coverage (PICP), the target energy consumption prediction model LECQR has a higher interval prediction coverage than the other energy consumption prediction sub-models LCQR. In the evaluation of prediction interval normalized average width (PINAW), the prediction intervals of the energy consumption prediction sub-models LCQR1 and LCQR2 are narrower, while the prediction interval of the target energy consumption prediction model LECQR is relatively wider. In the evaluation of coverage width criterion (CWC), the target energy consumption prediction model LECQR has a better interval prediction coverage than the other energy consumption prediction sub-models LCQR.

[0116] The energy consumption probability prediction method provided by the present application performs feature extraction on the predicted building data of the predicted building in a preset future time period, so as to accurately find the key features related to energy consumption from the complex data, and obtain the predicted energy consumption related features of the preset future time period. The predicted energy consumption related features fully consider various possible situations of future building operation. Using a pre-trained target energy consumption prediction model, energy consumption is predicted based on the predicted energy consumption related features of the preset future time period, and the predicted energy consumption probability interval of the HVAC of the predicted building in the preset future time period is obtained to effectively quantify the uncertainty in the energy consumption prediction. Therefore, the target energy consumption prediction model LECQR of the present application can effectively evaluate the probability of building energy consumption, reduce the impact of energy supply problems on the normal operation of the building, and enhance the stability and reliability of the building.

[0117] Based on the same inventive concept, an embodiment of the present application also provides a training device for an energy consumption probability prediction model and an energy consumption probability prediction device. Since the principle of solving the problem by the device in the embodiment of the present application is similar to the training method and energy consumption probability prediction method of the above-mentioned energy consumption probability prediction model in the embodiment of the present application, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.

[0118] Figure 11 This is a structural diagram of a training device for an energy consumption probability prediction model provided in an embodiment of the present application. Figure 11 As shown, the training device 1100 of the energy consumption probability prediction model may include: The first extraction module 1101 is used to extract features from the historical building data of the target building in a preset historical time period to obtain sample energy consumption related features of the preset historical time period; The first constructing module 1102 is configured to construct a sample data set based on the sample energy consumption related features of a preset historical time period and the actual energy consumption probability of the HVAC in the target building during the preset historical time period; A division module 1103 is configured to divide the sample data set into a plurality of sub-data sets, wherein the plurality of sub-data sets respectively correspond to a plurality of sub-time periods within a preset historical time period; A training module 1104 is configured to train a plurality of initial energy consumption prediction models based on a plurality of sub-data sets to obtain a plurality of trained energy consumption prediction sub-models; The second construction module 1105 is used to construct a target energy consumption prediction model based on multiple trained energy consumption prediction sub-models.

[0119] In an optional embodiment, the sample energy consumption related features include: sample linear correlation features and sample nonlinear correlation features; the first extraction module 1101 is specifically used to: based on the actual energy consumption probability, use a preset linear correlation analysis method to extract features from historical building data to obtain sample linear correlation features; based on the actual energy consumption probability, use a preset nonlinear correlation analysis method to extract features from historical building data to obtain sample nonlinear correlation features; and fuse the sample linear correlation features and the sample nonlinear correlation features to obtain sample energy consumption related features.

[0120] In an optional embodiment, the first extraction module 1101 is specifically used to: use a preset linear correlation analysis method to calculate the Pearson correlation coefficient between the actual energy consumption probability and various feature data in the historical building data; and based on the Pearson correlation coefficient, screen sample linear correlation features from various feature data in the historical building data.

[0121] In an optional embodiment, the first extraction module 1101 is specifically used to: use a preset nonlinear correlation analysis method to calculate the Spearman rank correlation coefficient between the actual energy consumption probability and various characteristic data in the historical building data; and screen sample nonlinear correlation features from various characteristic data in the historical building data based on the Spearman rank correlation coefficient.

[0122] In an optional embodiment, each initial energy consumption prediction model integrates: an initial long short-term memory (LSTM) network and an initial conformal quantile regression (CQR) prediction model; a training module 1104 is specifically used to: use the initial LSTM network in each initial energy consumption prediction model to perform time series processing on the sample energy consumption related features in each sub-data set to obtain the time series feature information corresponding to each sub-data set; use the initial CQR prediction model in each initial energy consumption prediction model to perform energy consumption prediction on the time series feature information corresponding to each sub-data set to obtain the predicted energy consumption probability interval corresponding to each sub-data set; according to the actual energy consumption probability of the sub-time period corresponding to each sub-data set, and the corresponding predicted energy consumption probability interval, the initial LSTM network and the initial CQR prediction model in each initial energy consumption prediction model are trained to obtain a trained energy consumption prediction sub-model, and the trained energy consumption prediction sub-model includes: a trained LSTM network and a trained CQR prediction network.

[0123] In an optional implementation, the second construction module 1105 is specifically configured to perform ensemble learning training on a plurality of trained energy consumption prediction sub-models to obtain a target energy consumption prediction model.

[0124] In an optional embodiment, the second construction module 1105 is specifically used to: use each trained energy consumption prediction sub-model to predict the sample energy consumption related features of each sub-time period, and obtain the predicted energy consumption probability interval of each trained energy consumption prediction sub-model for each sub-time period; obtain the prediction deviation of each trained energy consumption prediction sub-model for each sub-time period based on the predicted energy consumption probability interval of each trained energy consumption prediction sub-model for each sub-time period, and the actual energy consumption probability of each sub-time period; construct a prediction residual set based on the prediction deviation of multiple trained energy consumption prediction sub-models for each sub-time period; and construct a target energy consumption prediction model based on the prediction residual set.

[0125] Figure 12 This is a schematic diagram of the structure of an energy consumption probability prediction device provided in an embodiment of the present application. Figure 12 As shown, the energy consumption probability prediction device 1200 may include: The second extraction module 1201 is used to extract features from the predicted building data of the building to be predicted in a preset future time period to obtain features related to predicted energy consumption in the preset future time period; The prediction module 1202 is used to use a pre-trained target energy consumption prediction model to perform energy consumption prediction based on the predicted energy consumption related features of a preset future time period, and obtain a predicted HVAC energy consumption probability interval of the building to be predicted in the preset future time period.

[0126] It should be noted that for details not disclosed in the training device of the energy consumption probability prediction model and the energy consumption probability prediction device in the embodiment of the present application, please refer to the details disclosed in the training method of the energy consumption probability prediction model and the energy consumption probability prediction method in the embodiment of the present application, and the details will not be repeated here.

[0127] The above modules can be one or more integrated circuits configured to implement the above methods, such as one or more application-specific integrated circuits (ASICs), one or more microprocessors, or one or more field programmable gate arrays (FPGAs). For example, when a module is implemented by scheduling program code through a processing element, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules can be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0128] Optionally, an embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the processor executes the steps of the method for training an energy consumption probability prediction model for a mobile storage medium and the method for predicting energy consumption probability in the above-described embodiment. The specific implementation methods and technical effects are similar and will not be further described here.

[0129] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, various modifications and variations of the present application are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A training method for an energy consumption probability prediction model, characterized in that: include: Extracting features from historical building data of a target building in a preset historical time period to obtain sample energy consumption-related features of the preset historical time period; Constructing a sample data set based on the sample energy consumption related characteristics of the preset historical time period and the actual energy consumption probability of the HVAC in the target building during the preset historical time period; Dividing the sample data set into a plurality of sub-data sets, wherein the plurality of sub-data sets respectively correspond to a plurality of sub-time periods within the preset historical time period; Training a plurality of initial energy consumption prediction models according to the plurality of sub-data sets to obtain a plurality of trained energy consumption prediction sub-models; A target energy consumption prediction model is constructed based on the multiple trained energy consumption prediction sub-models.

2. The training method of the energy consumption probability prediction model according to claim 1, characterized in that: The sample energy consumption related features include: sample linear correlation features and sample nonlinear correlation features; the feature extraction of historical building data of the target building in a preset historical time period to obtain the sample energy consumption related features of the preset historical time period includes: According to the true energy consumption probability, a preset linear correlation analysis method is used to extract features from the historical building data to obtain sample linear correlation features; According to the true energy consumption probability, a preset nonlinear correlation analysis method is used to extract features of the historical building data to obtain sample nonlinear correlation features; The sample linear correlation feature and the sample nonlinear correlation feature are fused to obtain the sample energy consumption correlation feature.

3. The training method of the energy consumption probability prediction model according to claim 2, characterized in that: The method of extracting features from the historical building data based on the actual energy consumption probability and using a preset linear correlation analysis method to obtain sample linear correlation features includes: Using the preset linear correlation analysis method, respectively calculating the Pearson correlation coefficient between the actual energy consumption probability and various characteristic data in the historical building data; The sample linear correlation features are screened from various types of feature data in the historical building data according to the Pearson correlation coefficient.

4. The training method of the energy consumption probability prediction model according to claim 2, characterized in that: The method of extracting features from the historical building data based on the actual energy consumption probability and using a preset nonlinear correlation analysis method to obtain sample nonlinear correlation features includes: Using the preset nonlinear correlation analysis method, respectively calculating the Spearman rank correlation coefficient between the actual energy consumption probability and various characteristic data in the historical building data; The sample nonlinear correlation features are screened from various types of feature data in the historical building data according to the Spearman rank correlation coefficient.

5. The training method of the energy consumption probability prediction model according to claim 1, characterized in that: Each initial energy consumption prediction model integrates: an initial long short-term memory (LSTM) network and an initial conformal quantile regression (CQR) prediction model; The method of training the multiple initial energy consumption prediction models according to the multiple sub-data sets to obtain multiple trained energy consumption prediction sub-models includes: Using the initial LSTM network in each initial energy consumption prediction model to perform time series processing on the sample energy consumption related features in each sub-data set to obtain time series feature information corresponding to each sub-data set; Using the initial CQR prediction model in each initial energy consumption prediction model, perform energy consumption prediction on the time series feature information corresponding to each sub-data set to obtain a predicted energy consumption probability interval corresponding to each sub-data set; According to the actual energy consumption probability of the sub-time period corresponding to each sub-data set and the corresponding predicted energy consumption probability interval, the initial LSTM network and the initial CQR prediction model in each initial energy consumption prediction model are trained to obtain a trained energy consumption prediction sub-model, and the trained energy consumption prediction sub-model includes: a trained LSTM network and a trained CQR prediction network.

6. The training method of the energy consumption probability prediction model according to claim 1, characterized in that: The step of constructing a target energy consumption prediction model based on the plurality of trained energy consumption prediction sub-models includes: The multiple trained energy consumption prediction sub-models are subjected to integrated learning training to obtain the target energy consumption prediction model.

7. The training method of the energy consumption probability prediction model according to claim 6, characterized in that: The performing ensemble learning training on the plurality of trained energy consumption prediction sub-models to obtain the target energy consumption prediction model includes: Using each trained energy consumption prediction sub-model to predict the sample energy consumption related features of each sub-time period, to obtain the predicted energy consumption probability interval of each trained energy consumption prediction sub-model for each sub-time period; Obtaining a prediction deviation of each trained energy consumption prediction sub-model for each sub-time period based on the predicted energy consumption probability interval of each trained energy consumption prediction sub-model for each sub-time period and the actual energy consumption probability of each sub-time period; Constructing a prediction residual set based on the prediction deviations of the multiple trained energy consumption prediction sub-models for each sub-time period; The target energy consumption prediction model is constructed according to the prediction residual set.

8. A method for predicting energy consumption probability, characterized in that: include: Extracting features from the predicted building data of the building to be predicted in a preset future time period to obtain features related to predicted energy consumption in the preset future time period; A pre-trained target energy consumption prediction model is used to perform energy consumption prediction based on the predicted energy consumption related features of the preset future time period, and a predicted HVAC energy consumption probability interval of the building to be predicted in the preset future time period is obtained.

9. A computer device, characterized in that: include: processor; A memory for storing processor executable instructions; wherein, the processor implements the training method of the energy consumption probability prediction model as described in any one of claims 1 to 7 and the energy consumption probability prediction method as described in claim 8 by running the executable instructions.

10. A computer-readable storage medium, characterized in that The storage medium stores a computer program, which, when executed by a processor, executes the training method of the energy consumption probability prediction model according to any one of claims 1 to 8 and the steps of the energy consumption probability prediction method according to claim 8.

Images