Multi-source transfer learning method and device for scarce sample air conditioner energy consumption prediction and electronic equipment
By adopting a multi-source transfer learning method in air conditioner energy consumption prediction, using multiple source domain data sets and target domain data sets, the model parameters are dynamically adjusted, and the problem of sample scarcity and distribution differences is solved, and the performance of air conditioner energy consumption prediction is improved and generalization ability.
Patent Information
- Application Number
- CN202411071684.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-06
AI Technical Summary
In the prediction of air conditioner energy consumption, the existing technology is difficult to effectively predict the samples when samples are scarce, especially the data of newly built or majorly renovated buildings and long-term operating buildings, which leads to the increased difficulty of generalization during the training process, and a single transfer learning strategy Poor performance is performed under different distributions of different types of buildings and climatic conditions.
Using multi-source transfer learning method, by introducing model-level regular functions into the objective function, dynamically adjusting the transfer learning model parameters, using multiple source domain data sets to train the base prediction model, and building the transfer model on the target domain, and performing weighting Prediction, combining distribution consistency error and model complexity regularization terms, optimize the loss function to improve migration accuracy and generalization performance.
It effectively overcomes the problem of sample scarcity, improves the performance of air conditioner energy consumption prediction, enhances the migration accuracy and generalization capabilities of the model, and provides an effective solution for air conditioner system energy consumption prediction in sample scarcity scenarios.
Smart Images

Figure BDA0004981051840000041 
Figure BDA0004981051840000042 
Figure BDA0004981051840000043
Abstract
Description
Technical Field
[0001] The present invention relates to a method, a device and an electronic device for predicting air conditioning energy consumption by using a multi-source transfer learning method when available sample data is scarce, and belongs to the application field of data mining technology in energy management systems. Background Art
[0002] As building energy conservation and sustainable development issues are gaining more and more attention, energy efficiency management in the building sector has become an important research direction for new power systems. The energy consumption of air-conditioning systems is an important part of the entire building emission reduction value chain, and it is also a controllable part. Its load has energy storage characteristics, and accurate prediction of it will help achieve large-scale coordination and normalized regulation of flexible loads, thereby solving the problem of shortage of regulation resources in new power systems.
[0003] Energy consumption prediction of air conditioning system is a complex systemic problem that is nonlinear, time-space related, and affected by the complex interaction of internal and external parameters and user behavior. At present, machine learning and deep learning technologies have become the core tools for energy and power system research by efficiently capturing the internal laws of complex systems. They have achieved rich research results and provided better performance for prediction than traditional methods.
[0004] However, the above prediction studies are all based on the fact that the research subjects have sufficient labeled instance sample data, but in actual situations, the historical data of building loads and the number of sampling samples are often missing in different situations. Especially for newly built or significantly renovated buildings, the initial data is often insufficient to support in-depth predictions. For long-term operating buildings, technology upgrades and system maintenance may lead to data breakpoints, which also cause the problem of insufficient samples. For the above training process, the insufficient number of samples will increase the difficulty of generalization and cause overfitting of the trained network parameters. Therefore, for prediction scenarios with limited samples, relying purely on the above learning methods may not be the best choice.
[0005] At present, transfer learning is a mainstream method to solve the problem of sample scarcity. Solution transfer learning is to provide a better initial point for the target domain through pre-training, and then fine-tune the learned knowledge. However, these methods usually assume that the data distribution of the source domain and the target domain is similar. When the distribution between the source domain and the target domain does not match, a single transfer learning strategy may produce poor transfer effects or even lead to negative transfer. In the field of energy consumption prediction of air-conditioning systems, researchers are faced with data from different types of buildings (such as commercial buildings, residential buildings, schools, etc.) and different climatic conditions (such as cold areas, hot areas, humid areas, etc.). These data sets have significant distribution differences due to their different characteristics, making it difficult for a single transfer strategy to cope with them.
[0006] Therefore, the present invention proposes a multi-source transfer learning method for air-conditioning energy consumption prediction with scarce samples. By adding a model-level regularization function in the objective function, knowledge sharing between multiple source domains is achieved and the migration accuracy and generalization performance of the model are improved, providing an effective solution for air-conditioning system energy consumption prediction in scenarios with scarce samples. Summary of the invention
[0007] The invention purpose of this application is: by adopting a multi-source transfer learning framework, the problem of sample scarcity is effectively overcome, the air-conditioning energy consumption prediction performance under scarce sample conditions is improved, and a reference basis is provided for achieving large-scale coordination and normalized regulation of building flexible loads.
[0008] This application discloses a multi-source transfer learning method for air conditioning energy consumption prediction with scarce samples, which adopts the following steps:
[0009] Preprocess the data;
[0010] Determine a loss function for a multi-source transfer learning model based on the preprocessed data;
[0011] Dynamically adjust the parameters of the transfer learning model according to the distribution difference between the source domain and the target domain;
[0012] The transfer learning model is used to predict the air conditioning energy consumption in the target domain and obtain the predicted energy consumption value.
[0013] Furthermore, the data preprocessing includes:
[0014] Collect multiple source domain data sets and target domain data sets, including different types of air conditioning energy consumption data, building parameter data, and auxiliary meteorological data;
[0015] The previous values of adjacent features are used to interpolate missing values in energy consumption and meteorological time series data, and a method based on quantile statistics is used to remove noise data and outliers;
[0016] Transform all feature values in the source and target domain datasets to a common time scale and value range;
[0017] Sequence segments are selected with the required sliding window length as the input and output data sequences of the experimental data set, and are divided into training sets, validation sets, and test sets according to a certain ratio.
[0018] Furthermore, the determination of the loss function includes introducing a model-level regularizer to quantitatively evaluate the differences between multiple source domains and the target domain on the basis of the objective function of the multi-source base prediction model.
[0019] Furthermore, the method for determining the loss function for the multi-source transfer learning model based on the preprocessed data is:
[0020] A neural network model is used to train each source domain data set to obtain the corresponding base prediction model as the main part of the loss function;
[0021] A consistency error regularization term is introduced into the loss function to measure the distribution difference between the source domain model and the target domain model;
[0022] A regularization term is introduced into the loss function to control the complexity of the model and prevent overfitting.
[0023] Furthermore, the multi-source base prediction model is:
[0024] For a sufficient sample source domain D s1 ,D s2 ,...,D sm , train a base prediction model on each source domain, denoted as f1(x)...,f i (x),...,f m (x). Build a new migration model f on the target domain t (x), its prediction function is recorded as:
[0025] f t (x) = ∑ω i f i (x) (1)
[0026] Among them, ω i is the weight associated with source domain i.
[0027] Furthermore, using a model-level regularizer to quantitatively evaluate the difference between the source domain and the target domain includes: introducing a distribution mapping based on the maximum mean difference, wherein the distribution mapping of the maximum mean difference is used to measure the distribution difference between the source domain and the target domain, and the specific calculation formula is as follows:
[0028]
[0029] Where X={x1,x2,...x n} and Y={y1,y2,...,y m} represents the source domain and target domain data, k(·) is the Gaussian radial basis function, and σ is the width of the Gaussian kernel function.
[0030] Furthermore, the calculation process of the loss function is as follows:
[0031] First, define the loss function to optimize the mapping process; for the target domain D t The feature space Xt and label space Y t , define the loss function L LOSS as follows:
[0032] L LOSS=L TASK +αR1+βR2 (3)
[0034] Among them, L TASK is the model prediction consistency error. Here, the mean square error is used to measure the difference between the predicted value and the true value obtained by weighting the prediction model; α and β are the weights of the regularization terms, which control the strength of different regularizers and the model complexity respectively;
[0035] R1 is the distribution consistency error regularization term, which is used to measure the label distribution difference between source domain models. Its calculation formula is:
[0036]
[0037] R2 is a regularization term that controls the complexity of the model and is used to prevent the model from overfitting. Its calculation formula is as follows:
[0038]
[0039] Furthermore, the method for dynamically adjusting the parameters of the transfer learning model according to the distribution difference between the source domain and the target domain is:
[0040] During the entire training process, a gradient descent-based optimization algorithm is used for optimization;
[0041] In each round of iteration, the model parameters are updated by optimizing the loss function;
[0042] Use the cross-validation method to adjust the values of the regularization term weights α and β;
[0043] The loss on the validation set is continuously monitored. Once the loss stops decreasing or starts to increase after a certain number of times, the iteration is stopped. The model weights and mapping functions at the time of stopping the iteration are used as the final model results.
[0044] A prediction device for a multi-source transfer learning method for scarce sample air conditioning energy consumption prediction, the prediction device comprising:
[0045] A data preprocessing module, used for preprocessing data;
[0046] The loss function module is used to introduce a model-level regularizer to quantitatively evaluate the differences between multiple source domains and target domains based on the objective function of the multi-source base prediction model;
[0047] The transfer learning model module is used to predict the air conditioning energy consumption in the target domain and obtain the energy consumption prediction value.
[0048] The data preprocessing module preprocesses the data including:
[0049] Collect multiple source domain data sets and target domain data sets, including different types of air conditioning energy consumption data, building parameter data, and auxiliary meteorological data;
[0050] The previous values of adjacent features are used to interpolate missing values in energy consumption and meteorological time series data, and a method based on quantile statistics is used to remove noise data and outliers;
[0051] Transform all feature values in the source and target domain datasets to a common time scale and value range;
[0052] Sequence segments are selected with the required sliding window length as the input and output data sequences of the experimental data set, and are divided into training sets, validation sets, and test sets according to a certain ratio.
[0053] The loss function module is used to introduce a model-level regularizer to quantitatively evaluate the differences between multiple source domains and target domains based on the objective function of the multi-source base prediction model;
[0054] The multi-source prediction model is:
[0055] For a sufficient sample source domain D s1 ,D s2 ,...,D sm , train a base prediction model on each source domain, denoted as f1(x)...,f i (x),...,f m (x). Build a new migration model f on the target domain t (x), its prediction function is recorded as:
[0056] f t (x) = ∑ω i f i (x) (1)
[0057] Among them, ω i is the weight associated with source domain i.
[0058] The dynamically adjusting the parameters of the transfer learning model according to the distribution difference between the source domain and the target domain includes:
[0059] During the entire training process, a gradient descent-based optimization algorithm is used for optimization;
[0060] In each round of iteration, the model parameters are updated by optimizing the loss function;
[0061] Use the cross-validation method to adjust the values of the regularization term weights α and β;
[0062] The loss on the validation set is continuously monitored. Once the loss stops decreasing or starts to increase after a certain number of times, the iteration is stopped. The model weights and mapping functions at the time of stopping the iteration are used as the final model results.
[0063] Furthermore, an electronic device includes one or more processors for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors execute the multi-source transfer learning method for scarce sample air conditioning energy consumption prediction.
[0064] Furthermore, a storage medium is provided, wherein a computer program is stored in the storage medium, wherein the computer program is configured to execute the multi-source transfer learning method for scarce sample air conditioning energy consumption prediction when running.
[0065] The present invention proposes a multi-source transfer learning method for air-conditioning energy consumption prediction with scarce samples. By adding a model-level regularization function in the objective function, knowledge sharing between multiple source domains is achieved and the migration accuracy and generalization performance of the model are improved, providing an effective solution for air-conditioning system energy consumption prediction in sample-scarce scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0066] Figure 1 It is a flow chart of a multi-source transfer learning method for scarce sample air conditioning energy consumption prediction in the present invention;
[0067] Figure 2 It is a schematic diagram of the multi-source domain transfer learning framework structure in the present invention.
[0068] Figure 3 It is a structural principle diagram of a prediction device of a multi-source transfer learning method for predicting air conditioning energy consumption with scarce samples in the present invention.
[0069] Figure 4 This is a schematic diagram of the structure of a computer device in some embodiments of the present invention. DETAILED DESCRIPTION
[0070] In order to more clearly illustrate the technical solution of the present invention, the present invention will be further described below; obviously, the following description is only a part of the embodiments, and for ordinary technicians in this field, without paying creative work, the technical solution of the present invention can also be applied to other similar scenarios based on these; in order to more clearly illustrate the technical solution of the present invention, the following description is combined with the attached Figure 1-4 The technical solution of the present invention is further described in detail.
[0071] Embodiment 1:
[0072] Refer to the attached Figure 1-2,The present invention provides a multi-source transfer learning method for ,prediction of air conditioning energy consumption with scarce samples, adopts the following steps;
[0073] Preprocess the data;
[0074] Determine a loss function for a multi-source transfer learning model based on the preprocessed data;
[0075] Dynamically adjust the parameters of the transfer learning model according to the distribution difference between the source domain and the target domain;
[0076] The transfer learning model is used to predict the air conditioning energy consumption in the target domain and obtain the predicted energy consumption value.
[0077] The data preprocessing includes:
[0078] Collect multiple source domain data sets and target domain data sets, including data collection, data filling and cleaning, data normalization, data sequence generation and division; including different types of air conditioning energy consumption data, building parameter data and auxiliary meteorological data;
[0079] The implementation steps of data acquisition are as follows: Collect data sets related to air conditioning energy consumption from multiple source domains and target domains. The source domain data sets should cover different types of building air conditioning, different geographical locations, and climate conditions to ensure diversity and breadth. The target domain data sets are data in a specific environment or scenario for the prediction task.
[0080] The implementation steps of data filling and cleaning are as follows: use the previous values of adjacent features to fill in the missing values in the energy consumption and meteorological time series data, and use a method based on quantile statistics to remove noise data and outliers.
[0081] The implementation steps of data normalization processing are as follows: all eigenvalues in the source domain and target domain data sets are converted to a common time scale, and minimum-maximum normalization is performed on all data sets so that the eigenvalues of different data sets are scaled to the range of 0 to 1.
[0082] Among them, the implementation steps of generating and dividing data sequences are as follows: the sample data of each source domain and target domain are preprocessed using the sliding window method. Sequence fragments are selected with a sliding window length Δt as the input data sequence of the experimental data set. These sequences reflect the air-conditioning system data samples in the past Δt days (i.e., [t-Δt, t] interval); accordingly, each input sequence has a corresponding label, indicating the target equipment power on the t+1 day. Furthermore, the data set is divided into a training set and a test set in a ratio of 8:2, and 10% of the data is extracted from the training set as a validation set to adjust the hyperparameters of the model.
[0083] The previous values of adjacent features are used to interpolate missing values in energy consumption and meteorological time series data, and a method based on quantile statistics is used to remove noise data and outliers;
[0084] Transform all feature values in the source and target domain datasets to a common time scale and value range;
[0085] Sequence segments are selected with the required sliding window length as the input and output data sequences of the experimental data set, and are divided into training sets, validation sets, and test sets according to a certain ratio.
[0086] Embodiment 2:
[0087] The present invention provides a multi-source transfer learning method for scarce sample air conditioning energy consumption prediction, which adopts the following steps:
[0088] The determination of the loss function includes introducing a model-level regularizer to quantitatively evaluate the differences between multiple source domains and target domains based on the objective function of the multi-source base prediction model.
[0089] Furthermore, the method for determining the loss function for the multi-source transfer learning model based on the preprocessed data is:
[0090] A neural network model is used to train each source domain data set to obtain the corresponding base prediction model as the main part of the loss function;
[0091] A consistency error regularization term is introduced into the loss function to measure the distribution difference between the source domain model and the target domain model;
[0092] A regularization term is introduced into the loss function to control the complexity of the model and prevent overfitting.
[0093] Furthermore, the multi-source base prediction model is:
[0094] For a sufficient sample source domain D s1 ,D s2 ,...,D sm , train a base prediction model on each source domain, denoted as f1(x)...,f i (x),...,f m (x). Build a new migration model f on the target domain t (x), its prediction function is recorded as:
[0095] f t (x) = ∑ω i f i (x) (1)
[0096] Among them, ω i is the weight associated with source domain i.
[0097] Furthermore, using a model-level regularizer to quantitatively evaluate the difference between the source domain and the target domain includes: introducing a distribution mapping based on the maximum mean difference, wherein the distribution mapping of the maximum mean difference is used to measure the distribution difference between the source domain and the target domain, and the specific calculation formula is as follows:
[0098]
[0099] Where X={x1,x2,...x n} and Y={y1,y2,...,y m} represents the source domain and target domain data, k(·) is the Gaussian radial basis function, and σ is the width of the Gaussian kernel function.
[0100] Furthermore, the calculation process of the loss function is as follows:
[0101] First, define the loss function to optimize the mapping process; for the feature space X on the target domain Dt t and label space Y t , define the loss function L LOSS as follows:
[0102] L LOSS =L TASK +αR1+βR2 (3)
[0104] Among them, L TASK is the model prediction consistency error. Here, the mean square error is used to measure the difference between the predicted value and the true value obtained by weighting the prediction model; α and β are the weights of the regularization terms, which control the strength of different regularizers and the model complexity respectively;
[0105] R1 is the distribution consistency error regularization term, which is used to measure the label distribution difference between source domain models. Its calculation formula is:
[0106]
[0107] R2 is a regularization term that controls the complexity of the model and is used to prevent the model from overfitting. Its calculation formula is as follows:
[0108]
[0109] Embodiment 3:
[0110] The present invention provides a transfer learning model of a multi-source transfer learning method for scarce sample air conditioning energy consumption prediction, and the construction process thereof is as follows:
[0111] During the entire training process, a gradient descent-based optimization algorithm is used for optimization;
[0112] In each round of iteration, the model parameters are updated by optimizing the loss function;
[0113] Use the cross-validation method to adjust the values of the regularization term weights α and β;
[0114] The loss on the validation set is continuously monitored. Once the loss stops decreasing or starts to increase after a certain number of times, the iteration is stopped. The model weights and mapping functions at the time of stopping the iteration are used as the final model results.
[0115] It should be noted that the preferred implementation scheme involved in the above embodiments of the present application is the same as the scheme provided in Example 1, as well as the application scenario and implementation process, but is not limited to the scheme provided in Example 1.
[0116] A prediction device for a multi-source transfer learning method for scarce sample air conditioning energy consumption prediction, the prediction device comprising:
[0117] A data preprocessing module, used for preprocessing data;
[0118] The loss function module is used to introduce a model-level regularizer to quantitatively evaluate the differences between multiple source domains and target domains based on the objective function of the multi-source base prediction model;
[0119] The transfer learning model module is used to predict the air conditioning energy consumption in the target domain and obtain the energy consumption prediction value.
[0120] The data preprocessing module preprocesses the data including:
[0121] Collect multiple source domain data sets and target domain data sets, including different types of air conditioning energy consumption data, building parameter data, and auxiliary meteorological data;
[0122] The previous values of adjacent features are used to interpolate missing values in energy consumption and meteorological time series data, and a method based on quantile statistics is used to remove noise data and outliers;
[0123] Transform all feature values in the source and target domain datasets to a common time scale and value range;
[0124] Sequence segments are selected with the required sliding window length as the input and output data sequences of the experimental data set, and are divided into training sets, validation sets, and test sets according to a certain ratio.
[0125] The loss function module is used to introduce a model-level regularizer to quantitatively evaluate the differences between multiple source domains and target domains based on the objective function of the multi-source base prediction model;
[0126] The multi-source prediction model is:
[0127] For a sufficient sample source domain D s1,D s2 ,...,D sm , train a base prediction model on each source domain, denoted as f1(x)...,f i (x),...,f m (x). Build a new migration model f on the target domain t (x), its prediction function is recorded as:
[0128] f t (x) = ∑ω i f i (x) (1)
[0129] Among them, ω i is the weight associated with source domain i.
[0130] The dynamically adjusting the parameters of the transfer learning model according to the distribution difference between the source domain and the target domain includes:
[0131] During the entire training process, a gradient descent-based optimization algorithm is used for optimization;
[0132] In each round of iteration, the model parameters are updated by optimizing the loss function;
[0133] Use the cross-validation method to adjust the values of the regularization term weights α and β;
[0134] The loss on the validation set is continuously monitored. Once the loss stops decreasing or starts to increase after a certain number of times, the iteration is stopped. The model weights and mapping functions at the time of stopping the iteration are used as the final model results.
[0135] Embodiment 4:
[0136] According to an embodiment of the present invention, there is also provided an electronic device, comprising one or more processors for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors execute the multi-source transfer learning method for scarce sample air conditioning energy consumption prediction.
[0137] Embodiment 5:
[0138] According to an embodiment of the present invention, a storage medium is further provided, in which a computer program is stored, wherein the computer program is configured to execute the multi-source transfer learning method for scarce sample air conditioning energy consumption prediction when running.
[0139] According to the multi-source transfer learning method for scarce sample air conditioning energy consumption prediction proposed in this application, the Figure 1 The flowchart of the multi-source transfer learning method for scarce sample air conditioning energy consumption prediction is shown in FIG. Figure 2Schematic diagram of the multi-source domain adaptive transfer learning framework structure shown.
[0140] in Figure 1 The process in the paper is as follows: data preprocessing, multi-source domain adaptive transfer learning process and result prediction.
[0141] The data preprocessing includes: obtaining N air-conditioning source domains and target domain data sets to be predicted, filling missing values, cleaning outliers, normalizing data, and generating partition data sliding window sequences.
[0142] The multi-source domain adaptive transfer learning process includes: training the base prediction model for each source domain data set, introducing distribution consistency error and complexity regularization terms, designing the loss function, and using adaptive parameters to adjust the training and update the model parameters.
[0143] The result prediction includes: inputting the target domain sample feature sequence and outputting the air conditioning energy consumption prediction result.
[0144] Figure 2 In the figure, multiple source domains including source domain 1 and source domain 2 are included until the target domain; multiple source domains enter the feature extraction module through the pre-training stage, the target domain enters the transfer module through the transfer learning stage, and the feature extraction module is integrated into the transfer module after weighted learning; the feature extraction module enters the fine-tuning module, and the transfer module can also enter the fine-tuning module; the fine-tuning module can generate a pre-training model through the fully connected layer, and can also enter the fully connected layer through the hidden variables to generate the target model.
[0145] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A multi-source transfer learning method for air conditioning energy consumption prediction with scarce samples, characterized by: Use the following steps; Preprocess the data; Determine a loss function for a multi-source transfer learning model based on the preprocessed data; Dynamically adjust the parameters of the transfer learning model according to the distribution difference between the source domain and the target domain; The transfer learning model is used to predict the air conditioning energy consumption in the target domain and obtain the predicted energy consumption value.
2. According to claim 1, a multi-source transfer learning method for scarce sample air conditioning energy consumption prediction is characterized by: The data preprocessing includes: Collect multiple source domain data sets and target domain data sets, including different types of air conditioning energy consumption data, building parameter data, and auxiliary meteorological data; The previous values of adjacent features are used to interpolate missing values in energy consumption and meteorological time series data, and a method based on quantile statistics is used to remove noise data and outliers; Transform all feature values in the source and target domain datasets to a common time scale and value range; Sequence segments are selected with the required sliding window length as the input and output data sequences of the experimental data set, and are divided into training sets, validation sets, and test sets according to a certain ratio.
3. The multi-source transfer learning method for scarce sample air conditioning energy consumption prediction according to claim 1 is characterized by: The determination of the loss function includes introducing a model-level regularizer to quantitatively evaluate the differences between multiple source domains and target domains based on the objective function of the multi-source base prediction model.
4. The multi-source transfer learning method for scarce sample air conditioning energy consumption prediction according to claim 1 is characterized by: Determining a loss function for a multi-source transfer learning model based on the preprocessed data includes: A neural network model is used to train each source domain data set to obtain the corresponding base prediction model as the main part of the loss function; A consistency error regularization term is introduced into the loss function to measure the distribution difference between the source domain model and the target domain model; A regularization term is introduced into the loss function to control the complexity of the model and prevent overfitting.
5. The multi-source transfer learning method for scarce sample air conditioning energy consumption prediction according to claim 3 is characterized by: The multi-source base prediction model is: For a sufficient sample source domain D s1 ,D s2 ,...,D sm , train a base prediction model on each source domain, denoted as f1(x)...,f i (x),...,f m (x); Build a new migration model f on the target domain t (x), its prediction function is recorded as: f t (x)=∑ω i f i (x) (1) Among them, ω i is the weight associated with source domain i.
6. The multi-source transfer learning method for scarce sample air conditioning energy consumption prediction according to claim 3 is characterized by: Using a model-level regularizer to quantitatively evaluate the difference between the source domain and the target domain includes: introducing a distribution mapping based on the maximum mean difference, where the distribution mapping of the maximum mean difference is used to measure the distribution difference between the source domain and the target domain. The specific calculation formula is as follows: Where X={x1,x2,...x n } and Y={y1,y2,...,y m } represents the source domain and target domain data, k(·) is the Gaussian radial basis function, and σ is the width of the Gaussian kernel function.
7. The multi-source transfer learning method for scarce sample air conditioning energy consumption prediction according to claim 1 is characterized by: The calculation process of the loss function is as follows: First, define the loss function to optimize the mapping process; for the feature space Xt and label space Y on the target domain Dt t , define the loss function L LOSS as follows: L LOSS =L TASK +αR1+βR2 (3) Among them, L TASK is the model prediction consistency error. Here, the mean square error is used to measure the difference between the predicted value and the true value obtained by weighting the prediction model; α and β are the weights of the regularization terms, which control the strength of different regularizers and the model complexity respectively; R1 is the distribution consistency error regularization term, which is used to measure the label distribution difference between source domain models. Its calculation formula is: R2 is a regularization term that controls the complexity of the model and is used to prevent the model from overfitting. Its calculation formula is as follows:
8. The multi-source transfer learning method for scarce sample air conditioning energy consumption prediction according to claim 1, characterized in that: The dynamically adjusting the parameters of the transfer learning model according to the distribution difference between the source domain and the target domain includes: During the entire training process, a gradient descent-based optimization algorithm is used for optimization; In each round of iteration, the model parameters are updated by optimizing the loss function; Use the cross-validation method to adjust the values of the regularization term weights α and β; The loss on the validation set is continuously monitored. Once the loss stops decreasing or starts to increase after a certain number of times, the iteration is stopped. The model weights and mapping functions at the time of stopping the iteration are used as the final model results.
9. A prediction device for a multi-source transfer learning method for scarce sample air conditioning energy consumption prediction, characterized in that: The prediction device comprises: A data preprocessing module, used for preprocessing data; The loss function module is used to introduce a model-level regularizer to quantitatively evaluate the differences between multiple source domains and target domains based on the objective function of the multi-source base prediction model; The transfer learning model module is used to predict the air conditioning energy consumption in the target domain and obtain the energy consumption prediction value.
10. The prediction device of the multi-source transfer learning method for scarce sample air conditioning energy consumption prediction according to claim 9, characterized in that: The data preprocessing module preprocesses the data including: Collect multiple source domain data sets and target domain data sets, including different types of air conditioning energy consumption data, building parameter data, and auxiliary meteorological data; The previous values of adjacent features are used to interpolate missing values in energy consumption and meteorological time series data, and a method based on quantile statistics is used to remove noise data and outliers; Transform all feature values in the source and target domain datasets to a common time scale and value range; Sequence segments are selected with the required sliding window length as the input and output data sequences of the experimental data set, and are divided into training sets, validation sets, and test sets according to a certain ratio.
11. The prediction device of the multi-source transfer learning method for scarce sample air conditioning energy consumption prediction according to claim 9, characterized in that: The loss function module is used to introduce a model-level regularizer to quantitatively evaluate the differences between multiple source domains and target domains based on the objective function of the multi-source base prediction model; The multi-source prediction model is: For a sufficient sample source domain D s1 ,D s2 ,...,D sm , train a base prediction model on each source domain, denoted as f1(x)...,f i (x),...,f m (x); Build a new migration model f on the target domain t (x), its prediction function is recorded as: f t (x)=∑ω i f i (x) (1) Among them, ω i is the weight associated with source domain i.
12. The prediction device of the multi-source transfer learning method for scarce sample air conditioning energy consumption prediction according to claim 9, characterized in that: The dynamically adjusting the parameters of the transfer learning model according to the distribution difference between the source domain and the target domain includes: During the entire training process, a gradient descent-based optimization algorithm is used for optimization; In each round of iteration, the model parameters are updated by optimizing the loss function; Use the cross-validation method to adjust the values of the regularization term weights α and β; The loss on the validation set is continuously monitored. Once the loss stops decreasing or starts to increase after a certain number of times, the iteration is stopped. The model weights and mapping functions at the time of stopping the iteration are used as the final model results.
13. An electronic device, characterized in that: It includes one or more processors for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors execute the multi-source transfer learning method for scarce sample air conditioning energy consumption prediction as described in any one of claims 1-8.
14. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the multi-source transfer learning method for scarce sample air conditioning energy consumption prediction as described in any one of claims 1 to 8 when running.
Citation Information
Patent Citations
Task splitting and mitigating learning prediction method based on multi-source domains
CN107103364A
Rotating machine fault diagnosis method based on multi-source migration fusion contraction framework
CN114048568A
Multi-modal probability distribution adaptive primary liver cancer pathology grading prediction method
CN114898872A
Remaining service life prediction method of multi-source domain transfer learning based on dynamic distribution self-adaption
CN116415485A
Central air conditioner load prediction method and system based on model parameter transfer learning
CN117077845A
Cited By
Automatic testing system and method for industrial internet software
CN119988228A
Control method for combustion system of direct injection methanol generator based on multi-objective optimization
CN120968919A
Model migration-based airfoil profile gas-water dynamic coefficient prediction method
CN121902643A
A model migration-based airfoil gas-water dynamic coefficient prediction method
CN121902643B
