Baseline evaluation method for air conditioner load under various meteorological conditions

Abstract

The invention discloses a baseline evaluation method for air conditioner loads under various meteorological conditions. The baseline evaluation method comprises the steps that S1, air conditioner load data are marked, and a data set is constructed based on the air conditioner load data; s2, an evaluation model is constructed, the model comprises a self-encoder and a classifier, the air conditioner load data are imported into the self-encoder, and self-encoding features are obtained; s3, importing the self-encoding features into a classifier, and obtaining a predicted load type; and respectively constructing loss functions in the processes of the step S2 and the step S3, and minimizing the loss functions to optimize the evaluation model. According to the method, real-time monitoring and self-adaptive adjustment of air conditioner load data distribution drift caused by meteorological condition changes are achieved, and it is ensured that feature expression is stable after dimensionality reduction; gradient updating fluctuation in the training process is reduced, the model convergence time is shortened, the classification accuracy of all categories in an air conditioner load classification task is improved, and a quantitative standard is provided for energy consumption monitoring and abnormal load recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence model training methods, and specifically relates to a baseline assessment method for air conditioning load under various meteorological conditions. Background Art

[0002] In recent years, the popularity of Internet of Things technology has prompted meteorological monitoring stations and air conditioning operation monitoring systems in various regions to continuously collect a large amount of real-time and historical data. The data content covers multiple dimensions such as ambient temperature, humidity, wind speed, air pressure, and air conditioning energy consumption, current, power, and indoor temperature. The diversity of meteorological conditions and regional characteristics make the data collected by different monitoring stations have significant differences in format, number of attributes, and time series association. In addition, the inevitable noise, missing and abnormal conditions in the data collection process bring many challenges to the subsequent data preprocessing, feature extraction, and model training. Especially when facing the distribution drift of air conditioning load data caused by changes in meteorological conditions, traditional data processing and dimensionality reduction methods are difficult to ensure the stability of features after dimensionality reduction and accurate prediction of air conditioning load, affecting the accuracy of baseline assessment and abnormal load identification. In view of the above problems, it is urgent to establish a set of overall technical solutions from data collection, preprocessing, feature extraction to machine learning modeling and load baseline construction to achieve accurate monitoring and evaluation of air conditioning load changes under various meteorological conditions, and ensure that the monitoring system can still maintain high processing efficiency and prediction accuracy when facing practical problems such as data heterogeneity, noise, and distribution drift.

[0003] The Chinese invention patent with publication number CN114636228A proposes a method for predicting the dispatchable load of a central air-conditioning system based on a machine learning algorithm. With the help of the historical operation data of the central air-conditioning system, a dispatchable load prediction model of the central air-conditioning system is constructed using a machine learning algorithm. The entire prediction model establishment process is clear and simple, with strong operability and high prediction accuracy, which facilitates the implementation of the central air-conditioning system participating in the demand-side response.

[0004] The Chinese invention patent with the publication number CN115186965A proposes a method for uncertainty analysis of the photovoltaic air conditioner life cycle based on machine learning. The method includes the following steps: 1) selecting photovoltaic air conditioner equipment; 2) collecting the resource and energy consumption of photovoltaic air conditioners during their entire life cycle; 3) determining the environmental impact category and environmental impact factor; 4) calculating different environmental impact potentials; 5) calculating the total environmental load of photovoltaic air conditioners; 6) conducting environmental impact assessment of photovoltaic air conditioners; 7) conducting uncertainty analysis on the environmental impact indicators of photovoltaic air conditioners; 8) establishing a machine learning model for variable importance analysis. This method can accurately evaluate the impact of photovoltaic air conditioners on the environment, provide more reliable suggestions for designers, and also has certain positive significance for protecting the environment.

[0005] The Chinese invention patent with publication number CN119227285A proposes an air conditioning load prediction method, device, computer equipment and medium, including: obtaining environmental factor data of the air-conditioned area within a preset period, the environmental factor data including: current radiation intensity, current temperature and current relative humidity; based on the environmental factor data and the initial load prediction model, determining the load category of the air-conditioned area within the preset period, the initial load prediction model is used to make an initial prediction of the air-conditioning load through the environmental factor data; based on the load category of the air-conditioned area within the preset period, determining the load prediction result of the air-conditioned area within the preset period. Thus, by introducing machine learning methods, a new technical path is provided for air conditioning load prediction, which facilitates accurate prediction of air conditioning load.

[0006] The prior art has the following deficiencies:

[0007] 1. The present invention uses a variational autoencoder model based on air conditioning load data distribution drift detection, and monitors data distribution changes by real-time calculation of skewness and kurtosis to achieve adaptive adjustment of data drift caused by changes in meteorological conditions.

[0008] 2. The present invention adopts a gradient phase alignment mechanism to balance the gradient direction during the variational autoencoder training process, reduce gradient update fluctuations, and improve the stability of the training process.

[0009] 3. The present invention adopts manifold constraint technology to constrain the local geometric structure between data in the low-dimensional latent space, ensuring that the key features of the air-conditioning load data are retained after dimensionality reduction.

[0010] 4. The present invention adopts a sparsity-based cascade-correlation neural network classification algorithm, which reduces the risk of model overfitting through sparsity masking and L1 regularization constraints, and ensures accurate discrimination of air-conditioning load in the low load, medium load and high load divisions. Summary of the invention

[0011] In view of the technical problems in the prior art, a baseline assessment method for air conditioning load under various meteorological conditions is provided, including: Step S1: label the air conditioning load data and construct a data set based on it; Step S2: construct an evaluation model, which includes an autoencoder and a classifier, import air-conditioning load data into the autoencoder, and obtain autoencoder features; in this process, air-conditioning load data distribution drift detection and adaptive adjustment are performed, and the skewness and kurtosis of the air-conditioning load data are calculated by performing real-time statistical analysis on the input air-conditioning load data to achieve dynamic compensation for the drift of the air-conditioning load data, thereby dynamically adjusting the distribution hypothesis of the autoencoder; Step S3: import the self-encoding features into the classifier to obtain the predicted load type; in this process, the sparsity of the classifier is controlled by using the cross entropy loss function based on sparsity constraints; In the process of step S2 and step S3, loss functions are respectively constructed, and the loss functions are minimized to optimize the evaluation model.

[0012] Furthermore, the air conditioning load data in step S1 includes: low load, medium load and high load.

[0013] Furthermore, in step S2, the air conditioning load data is imported into the autoencoder to obtain the autoencoder features, specifically: Step S21: Initialize the parameters of the autoencoder, specifically initialize the weights and biases of the autoencoder by random initialization. The initialization method is expressed as: ; ; in, and are the weight matrices of the encoder and decoder of the autoencoder respectively; and are the bias vectors of the encoder and decoder of the autoencoder respectively; is the standard deviation of the autoencoder initialization, represents normal distribution; Step S22: inputting the air-conditioning load data into the autoencoder, performing air-conditioning load data distribution drift detection and adaptive adjustment, then performing real-time statistical analysis on the input air-conditioning load data, calculating the skewness and kurtosis of the air-conditioning load data, thereby capturing the change of the air-conditioning load data distribution, and realizing dynamic compensation of the air-conditioning load data drift, thereby dynamically adjusting the distribution assumption of the autoencoder, and obtaining the dynamically adjusted air-conditioning load data; Step S23: The input dynamically adjusted air conditioning load data is mapped to a low-dimensional latent space representation through an encoder to achieve effective dimensionality reduction processing. The dynamically adjusted air conditioning load data is masked to simulate the situation where part of the air conditioning load data is lost. The calculation method of the latent space representation is: ; In the formula, The first potential representation; represents the element-wise product, represents the activation function; For the The mask of the air conditioning load data after dynamic adjustment of the autoencoder, Indicates Dynamically adjusted air conditioning load data; Step S24: After the representation learning of the latent space is completed, the dynamically adjusted air-conditioning load data is reconstructed through the decoder, the reconstructed air-conditioning load data points are obtained and the reconstruction error is calculated, the autoencoder loss function is constructed through the reconstruction error, and a stable training effect is obtained under the condition of air-conditioning load data distribution drift. The autoencoder loss function is expressed as: ; in, The decoder output reconstructed air conditioning load data, i.e., the self-encoding feature, represents the total number of reconstructed air conditioning load data, represents the regularization parameter of the manifold constraint, represents the popular constraint loss function, represents the phase alignment loss function; Step S25: Use the gradient descent method to optimize the network parameters of the autoencoder, gradually reduce the autoencoder loss function, obtain the optimal parameters of the autoencoder through iterative optimization, and complete the training process. The parameter update rules are as follows: ; ; ; ; in, represents the learning rate of the autoencoder; Indicates parameter update operation; Step S26: Repeat the iteration of steps S21 to S25 until a preset stop iteration condition is met.

[0014] Furthermore, the calculation method of the skewness and kurtosis in step S22 is expressed as: ; ; In the formula, For the The skewness of air conditioning load data, For the The kurtosis of air conditioning load data is and Air conditioning load data The mean and standard deviation of Represents the dimension of the air conditioning load data input to the autoencoder; is the first input to the autoencoder The first Features.

[0015] Furthermore, in step S24, the phase alignment loss function is used to utilize the gradient phase alignment mechanism to align the gradient direction and reduce the directional inconsistency in the gradient update process, so that the autoencoder can perform smooth learning. During the training process, by minimizing the phase alignment loss function, the oscillation in the gradient update process can be effectively reduced, and the stability of the autoencoder training can be improved. The gradient phase alignment alleviates the training oscillation caused by the inconsistent distribution of the air-conditioning load data, thereby ensuring that the key features captured in the dimensionality reduction process are stable and have high robustness, providing a more accurate low-dimensional representation for the baseline evaluation. The phase alignment loss function is expressed as: ; in, is the phase alignment loss function; and are the gradients of the encoder and decoder of the autoencoder, respectively. represents the inner product of gradient; is the gradient of the loss function of the autoencoder with respect to the encoder weights; is the gradient of the loss function of the autoencoder with respect to the decoder weights; The number of air conditioning load data input to the autoencoder for the current batch.

[0016] Furthermore, step S24 uses manifold constraints to constrain the geometric distance between low-dimensional representations to ensure the coherence of the air-conditioning load data in the latent space. By minimizing the manifold constraint loss function, the local structure and geometric characteristics of the air-conditioning load data in the latent space can be effectively maintained. The manifold constraint loss function The calculation formula is expressed as: ; in, and They represent the first potential representation and the jth potential representation; Indicates air conditioning load data of nearest neighbors; is a positive integer; represents the L2 norm.

[0017] Furthermore, step S3 is specifically as follows: Step S31: Initialize the parameters of the classifier, which include the weight and bias of the classifier. Small random numbers are used for initialization to break the symmetry. Specifically, the initialization method of the weight and bias of the classifier is expressed as: ; ; In the formula, and Represent the initialized weights and biases respectively; and The numbers of neurons representing input and output respectively; represents a normal distribution with a mean of 0 and a variance of 1; Step S32: Input the feature vector of the air conditioning load data after dimensionality reduction, and the classifier processes it through a series of cascaded neural network units. Each unit will perform sparsity constraints and correlation analysis. Specifically, in the forward propagation process, the activation of the neuron based on sparsity is expressed as: ; ; in, Represents the activation value of the Lth layer of the classifier; represents the activation function; and Respectively represent the weight and bias of the Lth layer of the classifier; is the sparse mask, represents the random dropout function, represents the probability of retention; represents element-wise product; Step S33: Calculate the loss function based on the output of the classifier and the true label, using the cross entropy loss function based on sparsity constraints, and the calculation method is expressed as: ; in, is the cross entropy loss function based on sparsity constraints; and denote the true label and the predicted label respectively, It is a hyperparameter that controls the strength of regularization and is used to adjust the strength of sparsity; is the sum of the absolute values ​​of all weights of the classifier; Step S34: Calculate the gradient of the L1 regularization term for each parameter through the back propagation algorithm, and update the model parameters. Use the gradient descent method with momentum to accelerate convergence and reduce oscillation. Specifically, perform gradient descent optimization on the loss function, and update the weights and biases as follows: ; ; in, and Represents the updated weights and biases of the classifier; is the learning rate of the classifier; and Represent the partial derivatives of the loss function with respect to weight and bias respectively; Step S35: Repeat the iteration of steps S31 to S34 until a preset stop iteration condition is met.

[0018] Furthermore, the calculation method of the L1 regularization term weight in step S3 is expressed as: ; in, represents the number of neurons in the hidden layer of the classifier; and They represent the classifier The absolute values ​​of the weights and biases of each neuron.

[0019] The positive and progressive effects of the present invention are:

[0020] 1. The present invention realizes real-time monitoring and adaptive adjustment of the distribution drift of air-conditioning load data caused by changes in meteorological conditions, ensuring stable feature expression after dimensionality reduction.

[0021] 2. Reduce the fluctuation of gradient updates during training, shorten the model convergence time, and improve the training efficiency of the autoencoder.

[0022] 3. Maintain the local geometric structure of air conditioning load data in the low-dimensional latent space so that the reduced-dimensional features can truly reflect the relationship between air conditioning operation and meteorological conditions.

[0023] 4. Improve the accuracy of classification of each category in air-conditioning load classification tasks, and provide quantitative standards for energy consumption monitoring and abnormal load identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flowchart of the steps of a baseline assessment method for air conditioning loads under various meteorological conditions;

[0025] Figure 2 It is a comparison diagram of the convergence speed of the evaluation model based on the autoencoder of the present invention and the autoencoder in the prior art;

[0026] Figure 3 It is a comparison chart of reconstruction error under different data drift conditions;

[0027] Figure 4 It is a comparison chart of the gradient variance of the evaluation model based on the autoencoder of the present invention and the autoencoder in the prior art. DETAILED DESCRIPTION

[0028] The following describes the embodiments of the present invention through specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention.

[0029] Reference Figure 1 , a baseline assessment method for air conditioning loads under various meteorological conditions, including: Step S1: label the air conditioning load data and construct a data set based on it; Step S2: construct an evaluation model, which includes an autoencoder and a classifier, import air-conditioning load data into the autoencoder, and obtain autoencoder features; in this process, air-conditioning load data distribution drift detection and adaptive adjustment are performed, and the skewness and kurtosis of the air-conditioning load data are calculated by performing real-time statistical analysis on the input air-conditioning load data to achieve dynamic compensation for the drift of the air-conditioning load data, thereby dynamically adjusting the distribution hypothesis of the autoencoder; Step S3: import the self-encoding features into the classifier to obtain the predicted load type; in this process, the sparsity of the classifier is controlled by using the cross entropy loss function based on sparsity constraints; In the process of step S2 and step S3, loss functions are respectively constructed, and the loss functions are minimized to optimize the evaluation model.

[0030] Furthermore, the air conditioning load data in step S1 includes: low load, medium load and high load.

[0031] Furthermore, in step S2, the air conditioning load data is imported into the autoencoder to obtain the autoencoder features, specifically: Step S21: Initialize the parameters of the autoencoder, specifically initialize the weights and biases of the autoencoder by random initialization. The initialization method is expressed as: ; ; in, and are the weight matrices of the encoder and decoder of the autoencoder respectively; and are the bias vectors of the encoder and decoder of the autoencoder respectively; is the standard deviation of the autoencoder initialization, represents normal distribution; Step S22: inputting the air-conditioning load data into the autoencoder, performing air-conditioning load data distribution drift detection and adaptive adjustment, then performing real-time statistical analysis on the input air-conditioning load data, calculating the skewness and kurtosis of the air-conditioning load data, thereby capturing the change of the air-conditioning load data distribution, and realizing dynamic compensation of the air-conditioning load data drift, thereby dynamically adjusting the distribution assumption of the autoencoder, and obtaining the dynamically adjusted air-conditioning load data; Step S23: The input dynamically adjusted air conditioning load data is mapped to a low-dimensional latent space representation through an encoder to achieve effective dimensionality reduction processing. The dynamically adjusted air conditioning load data is masked to simulate the situation where part of the air conditioning load data is lost. The calculation method of the latent space representation is: ; In the formula, The first potential representation; represents the element-wise product, represents the activation function; For the The mask of the air conditioning load data after dynamic adjustment of the autoencoder, Indicates Dynamically adjusted air conditioning load data; Step S24: After the representation learning of the latent space is completed, the dynamically adjusted air-conditioning load data is reconstructed through the decoder, the reconstructed air-conditioning load data points are obtained and the reconstruction error is calculated, the autoencoder loss function is constructed through the reconstruction error, and a stable training effect is obtained under the condition of air-conditioning load data distribution drift. The autoencoder loss function is expressed as: ; in, The decoder output reconstructed air conditioning load data, i.e., the self-encoding feature, represents the total number of reconstructed air conditioning load data, represents the regularization parameter of the manifold constraint, represents the popular constraint loss function, represents the phase alignment loss function; Step S25: Use the gradient descent method to optimize the network parameters of the autoencoder, gradually reduce the autoencoder loss function, obtain the optimal parameters of the autoencoder through iterative optimization, and complete the training process. The parameter update rules are as follows: ; ; ; ; in, represents the learning rate of the autoencoder; Indicates parameter update operation; Step S26: Repeat the iteration of steps S21 to S25 until a preset stop iteration condition is met.

[0032] Furthermore, the calculation method of the skewness and kurtosis in step S22 is expressed as: ; ; In the formula, For the The skewness of air conditioning load data, For the The kurtosis of air conditioning load data is and Air conditioning load data The mean and standard deviation of Represents the dimension of the air conditioning load data input to the autoencoder; is the first input to the autoencoder The first Features.

[0033] Furthermore, in step S24, the phase alignment loss function is used to utilize the gradient phase alignment mechanism to align the gradient direction and reduce the directional inconsistency in the gradient update process, so that the autoencoder can perform smooth learning. During the training process, by minimizing the phase alignment loss function, the oscillation in the gradient update process can be effectively reduced, and the stability of the autoencoder training can be improved. The gradient phase alignment alleviates the training oscillation caused by the inconsistent distribution of the air-conditioning load data, thereby ensuring that the key features captured in the dimensionality reduction process are stable and have high robustness, providing a more accurate low-dimensional representation for the baseline evaluation. The phase alignment loss function is expressed as: ; in, is the phase alignment loss function; and are the gradients of the encoder and decoder of the autoencoder, respectively. represents the inner product of gradient; is the gradient of the loss function of the autoencoder with respect to the encoder weights; is the gradient of the loss function of the autoencoder with respect to the decoder weights; The number of air conditioning load data input to the autoencoder for the current batch.

[0034] Furthermore, step S24 uses manifold constraints to constrain the geometric distance between low-dimensional representations to ensure the coherence of the air-conditioning load data in the latent space. By minimizing the manifold constraint loss function, the local structure and geometric characteristics of the air-conditioning load data in the latent space can be effectively maintained. The manifold constraint loss function The calculation formula is expressed as: ; in, and They represent the first potential representation and the jth potential representation; Indicates air conditioning load data of nearest neighbors; is a positive integer; represents the L2 norm.

[0035] Furthermore, step S3 is specifically as follows: Step S31: Initialize the parameters of the classifier, which include the weight and bias of the classifier. Small random numbers are used for initialization to break the symmetry. Specifically, the initialization method of the weight and bias of the classifier is expressed as: ; ; In the formula, and Represent the initialized weights and biases respectively; and The numbers of neurons representing input and output respectively; represents a normal distribution with a mean of 0 and a variance of 1; Step S32: Input the feature vector of the air conditioning load data after dimensionality reduction, and the classifier processes it through a series of cascaded neural network units. Each unit will perform sparsity constraints and correlation analysis. Specifically, in the forward propagation process, the activation of the neuron based on sparsity is expressed as: ; ; in, Represents the activation value of the Lth layer of the classifier; represents the activation function; and Respectively represent the weight and bias of the Lth layer of the classifier; is the sparse mask, represents the random dropout function, represents the probability of retention; represents element-wise product; Step S33: Calculate the loss function based on the output of the classifier and the true label, using the cross entropy loss function based on sparsity constraints, and the calculation method is expressed as: ; in, is the cross entropy loss function based on sparsity constraints; and denote the true label and the predicted label respectively, It is a hyperparameter that controls the strength of regularization and is used to adjust the strength of sparsity; is the sum of the absolute values ​​of all weights of the classifier; Step S34: Calculate the gradient of the L1 regularization term for each parameter through the back propagation algorithm, and update the model parameters. Use the gradient descent method with momentum to accelerate convergence and reduce oscillation. Specifically, perform gradient descent optimization on the loss function, and update the weights and biases as follows: ; ; in, and Represents the updated weights and biases of the classifier; is the learning rate of the classifier; and Represent the partial derivatives of the loss function with respect to weight and bias respectively; Step S35: Repeat the iteration of steps S31 to S34 until a preset stop iteration condition is met.

[0036] Furthermore, the calculation method of the L1 regularization term weight in step S3 is expressed as: ; in, represents the number of neurons in the hidden layer of the classifier; and They represent the classifier The absolute values ​​of the weights and biases of each neuron.

[0037] In order to verify the advantages of the technology of the present invention in the autoencoder structure, the following experiments were conducted:

[0038] In one example, refer to Figure 2The purpose of the experiment shown in the figure is to verify whether the autoencoder structure based on the air-conditioning load data distribution drift adopted by the present invention can converge to a lower loss value faster during the training process, thereby improving the training efficiency. The horizontal axis in the figure represents the number of training iterations, that is, the number of times the model updates the parameters during the training process; the vertical axis represents the loss function value calculated for each iteration, which reflects the error size of the model reconstructing the input data. The experiment uses simulated data to show the decrease in the loss value of the traditional autoencoder and the method of the present invention during the iteration process. It can be seen that the loss value of the method of the present invention decreases faster and reaches a lower final value under the same number of iterations, which shows that the method can more effectively capture the characteristics of the air-conditioning load data and dynamically adapt to the drift of the data distribution, thereby achieving faster convergence and better performance in practical applications.

[0039] In one example, refer to Figure 3 , Figure 3 The main purpose of the experiment shown is to examine how the model's ability to reconstruct data changes when facing different degrees of air-conditioning load data distribution drift (from no drift, slight drift to moderate and severe drift). The horizontal axis represents the degree of data drift, and different categories correspond to different drift conditions; the vertical axis represents the reconstruction error, that is, the difference between the model and the original data when reconstructing the input data. The experimental results show that under all degrees of drift, the method of the present invention can maintain a low reconstruction error, especially in the case of moderate and severe drift, the errors of traditional methods and other conventional preprocessing methods are significantly higher, which proves that the mechanism of dynamic compensation using real-time statistics (such as skewness and kurtosis) in the present invention enables the model to adjust internal parameters in a timely manner, thereby improving the robustness to data distribution drift and ensuring that key features can be accurately captured and reconstructed in various environments.

[0040] In one example, refer to Figure 4 , Figure 4 The purpose of the experiment shown is to evaluate the stability and adaptability of the two methods when the hyperparameter settings are different. The horizontal axis in the figure represents different learning rate values. The learning rate is the parameter of the step size in the gradient descent process; the vertical axis represents the accuracy finally achieved by the model, which reflects the ability of the model to capture the complex nonlinear relationship between air-conditioning load data and meteorological conditions. The experimental results show that at a lower learning rate, the accuracy of the two methods is relatively close, but as the learning rate increases, the accuracy of the traditional autoencoder decreases significantly, while the method of the present invention can still maintain a high and relatively stable accuracy, indicating that by dynamically adapting to data distribution drift and utilizing the gradient phase alignment mechanism, the method of the present invention has better robustness in hyperparameter selection and can maintain excellent performance in a wider range of learning rates, further verifying the adaptive advantage of the technology.

[0041] The present invention is described in detail above in conjunction with the embodiments of the accompanying drawings. A person skilled in the art can make various variations of the present invention according to the above description. Therefore, some details in the embodiments should not be construed as limiting the present invention, and the present invention shall be protected by the scope defined by the attached claims.

Claims

1. A baseline assessment method for air conditioning load under various meteorological conditions, characterized in that: include: Step S1: label the air conditioning load data and construct a data set based on it; Step S2: construct an evaluation model, which includes an autoencoder and a classifier, import air conditioning load data into the autoencoder, and obtain autoencoder features; In this process, the air conditioning load data distribution drift detection and adaptive adjustment are carried out. By performing real-time statistical analysis on the input air conditioning load data, the skewness and kurtosis of the air conditioning load data are calculated to achieve dynamic compensation for the air conditioning load data drift, thereby dynamically adjusting the distribution assumption of the self-editing device; Step S3: import the self-encoding features into the classifier to obtain the predicted load type; in this process, the sparsity of the classifier is controlled by using the cross entropy loss function based on sparsity constraints; In the process of step S2 and step S3, loss functions are respectively constructed, and the loss functions are minimized to optimize the evaluation model.

2. The method for evaluating the baseline of air conditioning load under various meteorological conditions according to claim 1, characterized in that: The air conditioning load data in step S1 includes: low load, medium load and high load.

3. The method for evaluating the baseline of air conditioning load under various meteorological conditions according to claim 1, characterized in that: In step S2, the air conditioning load data is imported into the autoencoder to obtain the autoencoder features, specifically: Step S21: Initialize the parameters of the autoencoder, specifically initialize the weights and biases of the autoencoder by random initialization. The initialization method is expressed as: ； ； in, and are the weight matrices of the encoder and decoder of the autoencoder respectively; and are the bias vectors of the encoder and decoder of the autoencoder respectively; is the standard deviation of the autoencoder initialization, represents normal distribution; Step S22: inputting the air-conditioning load data into the autoencoder, performing air-conditioning load data distribution drift detection and adaptive adjustment, then performing real-time statistical analysis on the input air-conditioning load data, calculating the skewness and kurtosis of the air-conditioning load data, thereby capturing the change of the air-conditioning load data distribution, and realizing dynamic compensation of the air-conditioning load data drift, thereby dynamically adjusting the distribution assumption of the autoencoder, and obtaining the dynamically adjusted air-conditioning load data; Step S23: The input dynamically adjusted air conditioning load data is mapped to a low-dimensional latent space representation through an encoder to achieve effective dimensionality reduction processing. The dynamically adjusted air conditioning load data is masked to simulate the situation where part of the air conditioning load data is lost. The calculation method of the latent space representation is: ； in, The first potential representation; represents the element-wise product, represents the activation function; For the The mask of the air conditioning load data after dynamic adjustment of the autoencoder, Indicates Dynamically adjusted air conditioning load data; Step S24: After the representation learning of the latent space is completed, the dynamically adjusted air-conditioning load data is reconstructed through the decoder, the reconstructed air-conditioning load data points are obtained and the reconstruction error is calculated, the autoencoder loss function is constructed through the reconstruction error, and a stable training effect is obtained under the condition of air-conditioning load data distribution drift. The autoencoder loss function It is expressed as: ； in, The decoder output reconstructed air conditioning load data, i.e., the self-encoding feature, represents the total number of reconstructed air conditioning load data, represents the regularization parameter of the manifold constraint, represents the popular constraint loss function, represents the phase alignment loss function; Step S25: Use the gradient descent method to optimize the network parameters of the autoencoder, gradually reduce the autoencoder loss function, obtain the optimal parameters of the autoencoder through iterative optimization, and complete the training process. The parameter update rules are as follows: ； ； ； ； in, represents the learning rate of the autoencoder; Indicates parameter update operation; Step S26: Repeat the iteration of steps S21 to S25 until a preset stop iteration condition is met.

4. The method for evaluating the baseline of air conditioning load under various meteorological conditions according to claim 3, characterized in that: The calculation method of the skewness and kurtosis in step S22 is expressed as: ； ； in, For the The skewness of air conditioning load data, For the The kurtosis of air conditioning load data is and Air conditioning load data The mean and standard deviation of Represents the dimension of the air conditioning load data input to the autoencoder; is the first input to the autoencoder The first Features.

5. The method for evaluating the baseline of air conditioning load under various meteorological conditions according to claim 4, characterized in that: In step S24, the phase alignment loss function is used to utilize the gradient phase alignment mechanism to align the gradient direction and reduce the directional inconsistency in the gradient update process, so that the autoencoder can perform smooth learning. During the training process, by minimizing the phase alignment loss function, the oscillation in the gradient update process can be effectively reduced, and the stability of the autoencoder training can be improved. The gradient phase alignment alleviates the training oscillation caused by the inconsistent distribution of the air-conditioning load data, thereby ensuring that the key features captured in the dimensionality reduction process are stable and have high robustness, providing a more accurate low-dimensional representation for the baseline evaluation. The phase alignment loss function is expressed as: ； in, is the phase alignment loss function; and are the gradients of the encoder and decoder of the autoencoder, respectively. represents the inner product of gradient; is the gradient of the loss function of the autoencoder with respect to the encoder weights; is the gradient of the loss function of the autoencoder with respect to the decoder weights; The number of air conditioning load data input to the autoencoder for the current batch.

6. A baseline assessment method for air conditioning load under multiple meteorological conditions according to claim 5, characterized in that: Step S24 uses manifold constraints to constrain the geometric distance between low-dimensional representations to ensure the coherence of the air-conditioning load data in the latent space. By minimizing the manifold constraint loss function, the local structure and geometric characteristics of the air-conditioning load data in the latent space can be effectively maintained. The manifold constraint loss function The calculation formula is expressed as: ； in, and They represent the first potential representation and the jth potential representation; Indicates air conditioning load data of nearest neighbors; is a positive integer; represents the L2 norm.

7. The method for evaluating the baseline of air conditioning load under various meteorological conditions according to claim 1, characterized in that: Step S3 is specifically as follows: Step S31: Initialize the parameters of the classifier, which include the weight and bias of the classifier. Small random numbers are used for initialization to break the symmetry. Specifically, the initialization method of the weight and bias of the classifier is expressed as: ； ； in, and Represent the initialized weights and biases respectively; and The numbers of neurons representing input and output respectively; represents a normal distribution with a mean of 0 and a variance of 1; Step S32: Input the feature vector of the air conditioning load data after dimension reduction, and the classifier processes it through a series of cascaded neural network units. Each unit will perform sparsity constraint and correlation analysis. Specifically, in the forward propagation process, the activation of the neuron based on sparsity is expressed as: ； ； in, Represents the activation value of the Lth layer of the classifier; represents the activation function; and Respectively represent the weight and bias of the Lth layer of the classifier; is the sparse mask, represents the random dropout function, represents the probability of retention; represents element-wise product; Step S33: Calculate the loss function based on the output of the classifier and the true label, using the cross entropy loss function based on sparsity constraints, and the calculation method is expressed as: ； in, is the cross entropy loss function based on sparsity constraints; and denote the true label and the predicted label respectively, It is a hyperparameter that controls the strength of regularization and is used to adjust the strength of sparsity; is the sum of the absolute values ​​of all weights of the classifier; Step S34: Calculate the gradient of the L1 regularization term for each parameter through the back propagation algorithm, and update the model parameters. Use the gradient descent method with momentum to accelerate convergence and reduce oscillation. Specifically, perform gradient descent optimization on the loss function and update the weights. and bias The way is expressed as: ； ； in, and Represents the updated weights and biases of the classifier; is the learning rate of the classifier; and Represent the partial derivatives of the loss function with respect to weight and bias respectively; Step S35: Repeat the iteration of steps S31 to S34 until a preset stop iteration condition is met.

8. The method for evaluating the baseline of air conditioning load under various meteorological conditions according to claim 7, characterized in that: The calculation method of the L1 regularization term weight in step S3 is expressed as: ； in, represents the number of neurons in the hidden layer of the classifier; and They represent the classifier The absolute values ​​of the weights and biases of each neuron.

Images