CN117909928A

CN117909928A - Air conditioner load prediction method and system based on big data analysis

Info

Publication number: CN117909928A
Application number: CN202410310985.1A
Authority: CN
Inventors: 何景彦; 陈泫光; 唐殊; 余晋; 喻华; 籍雁南
Original assignee: Chengdu Power Supply Co Of State Grid Sichuan Electric Power Corp
Current assignee: Chengdu Power Supply Co Of State Grid Sichuan Electric Power Corp
Priority date: 2024-03-19
Filing date: 2024-03-19
Publication date: 2024-04-19
Anticipated expiration: 2044-03-19
Also published as: CN117909928B

Abstract

The application provides an air conditioner load prediction method and an air conditioner load prediction system based on big data analysis. In the training process, the feature fusion calibration network performs feature fusion calibration on the energy consumption point state description of the expansion template and the sample learning data to generate optimized energy consumption point state description of the expansion template and the sample learning data, so that the prediction accuracy of the model is improved. Finally, the target air conditioner load prediction model obtained through training can be used for predicting load prediction thermodynamic diagrams of candidate air conditioner energy consumption big data, and powerful support is provided for optimizing operation and energy management of an air conditioner system.

Description

Air conditioner load prediction method and system based on big data analysis

Technical Field

The application relates to the technical field of artificial intelligence, in particular to an air conditioner load prediction method and system based on big data analysis.

Background

With the rapid development of society and the advancement of technology, air conditioning systems have been widely used in various buildings to provide comfortable indoor environments for people. However, the energy consumption of the air conditioning system occupies a considerable proportion of the total energy consumption of the building, and in the existing air conditioning system, the air conditioning load prediction is a key technology. The air conditioning load prediction can help optimize energy management of the air conditioning system, thereby achieving energy conservation and improving energy efficiency. However, since the air conditioning load is affected by many factors such as weather conditions (e.g., temperature, humidity, wind speed, etc.), building characteristics (e.g., building materials, designs, etc.), and usage patterns (e.g., number of people, usage time, etc.), it is very challenging to make accurate air conditioning load predictions.

Conventional methods typically rely on empirical formulas or simplified models, which often fail to take into account all influencing factors, and thus have limited prediction accuracy. In recent years, with the development of big data and machine learning techniques, more and more researches have begun to attempt air conditioning load prediction using these techniques. However, existing prediction methods typically only focus on a single influencing factor, such as considering only past energy consumption data or only weather data, and neglecting interactions between these factors. This may result in a large error in the predicted outcome. Therefore, how to effectively integrate and utilize various data to improve the accuracy of air conditioner load prediction is still a problem to be solved.

In general, existing air conditioning load prediction techniques have some limitations and require further research and improvement.

Disclosure of Invention

In view of the above, the present application aims to provide an air conditioner load prediction method and system based on big data analysis.

According to a first aspect of the present application, there is provided an air conditioner load prediction method based on big data analysis, the method comprising:

in the construction task of an air conditioner load prediction model of each stage, an extended sample learning data sequence and a multi-sample learning data sequence are obtained based on a past energy consumption data set and an extended meteorological data set, wherein the air conditioner load prediction model comprises an energy consumption vector extraction network for initializing weight parameters, a meteorological influence factor extraction network and a characteristic fusion calibration network;

For each sample learning data in the multi-sample learning data sequence, taking the sample learning data as the model loading data of the air conditioner load prediction model, generating a load prediction thermodynamic diagram of the sample learning data, wherein the characteristic fusion calibration network is used for carrying out characteristic fusion calibration on an expansion template and the energy consumption point state description of the sample learning data, generating an optimized expansion template and the optimized energy consumption point state description of the sample learning data, and the expansion template is a generation result of the expansion sample learning data sequence through the weather influence factor extraction network;

Based on the load prediction thermodynamic diagram of each sample learning data in the multi-sample learning data sequence and the air conditioner load labeling data of each sample learning data, performing model training on the air conditioner load prediction model until the model convergence requirement is met;

And determining the air conditioner load prediction model which meets the model convergence requirement as a target air conditioner load prediction model, wherein the target air conditioner load prediction model is used for predicting a load prediction thermodynamic diagram of candidate air conditioner energy consumption big data.

In some design ideas of the first aspect, the generating a load prediction thermodynamic diagram of the sample learning data by using the sample learning data as model loading data of the air conditioner load prediction model includes:

Loading the extended sample learning data sequence to the meteorological influence factor extraction network to generate the extended template;

Taking the sample learning data as the model loading data of the energy consumption vector extraction network, and generating an energy consumption characteristic vector of the sample learning data;

taking the sample learning data as model loading data of the meteorological influence factor extraction network, and generating an extended meteorological influence vector of the sample learning data;

Generating an energy consumption point state description of the sample learning data based on an energy consumption reference template obtained by initializing model learning, the expansion template, a fusion function, an energy consumption characteristic vector of the sample learning data and an expansion weather influence vector of the sample learning data, wherein the fusion function is determined by focusing a focusing function;

Performing feature fusion calibration on the energy consumption point state description of the expansion template and the sample learning data to generate an optimized energy consumption point state description of the optimized expansion template and the sample learning data;

And predicting and outputting a load prediction thermodynamic diagram of the sample learning data based on the energy consumption reference template, the optimized expansion template and the optimized energy consumption point state description of the sample learning data.

In some design ideas of the first aspect, the performing feature fusion calibration on the energy consumption point state descriptions of the expansion template and the sample learning data to generate optimized energy consumption point state descriptions of the expansion template and the sample learning data includes:

calculating a focusing attention coefficient between the energy consumption point state description of the sample learning data and each of the extended sample learning data in the extended sample learning data sequence;

According to the focusing attention coefficient between the energy consumption point state description of the sample learning data and each expansion sample learning data in the expansion sample learning data sequence, fusing the expansion sample learning data in the expansion sample learning data sequence through the characteristic vector of the meteorological influence factor extraction network, and generating the optimized expansion template;

and generating an optimized energy consumption point state description of the sample learning data based on the energy consumption reference template, the optimized expansion template, a fusion function, an energy consumption characteristic vector of the sample learning data and an expanded weather influence vector of the sample learning data.

And carrying out feature fusion calibration optimization on the energy consumption point state description of the expansion template and the sample learning data based on preset feature fusion calibration parameters, and generating optimized energy consumption point state description of the expansion template and the sample learning data.

In some design ideas of the first aspect, the performing feature fusion calibration optimization on the energy consumption point state description of the expansion template and the sample learning data based on the preset feature fusion calibration parameters, and generating an optimized energy consumption point state description of the expansion template and the sample learning data includes:

Calculating a focusing attention coefficient between a kth-1 round of optimized energy consumption point state description of the sample learning data and each extended sample learning data in the extended sample learning data sequence, wherein k is greater than 0, and the 0 th round of optimized energy consumption point state description of the sample learning data is the energy consumption point state description of the sample learning data;

According to the k-1 th round of optimized energy consumption point state description of the sample learning data and the focusing attention coefficient between each expansion sample learning data in the expansion sample learning data sequence, fusing the expansion sample learning data in the expansion sample learning data sequence through the characteristic vector of the meteorological influence factor extraction network, and generating a k round of optimized expansion template;

Generating a kth round of optimized energy consumption point state description of the sample learning data based on the energy consumption reference template, the kth round of optimized expansion template, a fusion function, an energy consumption characteristic vector of the sample learning data and an expanded weather influence vector of the sample learning data;

And iteratively executing the step of calculating the focusing attention coefficient between the K-1 th round of optimized energy consumption point state description of the sample learning data and each extended sample learning data in the extended sample learning data sequence until K is equal to K, wherein K is the characteristic fusion calibration parameter.

In some design ideas of the first aspect, the performing model training on the air-conditioning load prediction model based on the load prediction thermodynamic diagram of each sample learning data in the multiple sample learning data sequence and the air-conditioning load labeling data of each sample learning data until the model convergence requirement is met includes:

Determining training cost parameters based on load prediction thermodynamic diagrams of each sample learning data in the multi-sample learning data sequence and air conditioner load labeling data of each sample learning data;

and updating network parameter definition information of the meteorological influence factor extraction network by using a gradient descent algorithm based on the training cost parameter until the model convergence requirement is met.

In some design ideas of the first aspect, the multiple sample learning data sequence includes extended sample learning data and energy consumption reference sample learning data, and the training the air conditioner load prediction model based on load prediction thermodynamic diagrams of each sample learning data in the multiple sample learning data sequence and air conditioner load labeling data of each sample learning data until the model convergence requirement is met includes:

calculating blending training cost parameters based on the various expansion sample learning data in the multi-sample learning data sequence, the optimized expansion templates and the energy consumption reference templates obtained by initializing parameter learning;

constructing training cost parameters based on load prediction thermodynamic diagrams of each sample learning data in the multi-sample learning data sequence and air conditioner load labeling data of each sample learning data; and updating the network parameter definition information of the meteorological influence factor extraction network by using a gradient descent algorithm based on the blending training cost parameter and the training cost parameter until the model convergence requirement is met.

In some design ideas of the first aspect, the calculating the blending training cost parameter based on the respective expansion sample learning data in the multiple sample learning data sequence, the optimized expansion template, and the energy consumption reference template obtained by initializing parameter learning includes:

Calculating target feature deviation parameters of the expansion sample learning data according to the expansion sample learning data, wherein the target feature deviation parameters are the sum of first target feature deviation parameters and second target feature deviation parameters, the first target feature deviation parameters are determined based on a first definition coefficient, the first feature deviation parameters and the second feature deviation parameters, and the second target feature deviation parameters are determined based on a second definition coefficient, the first feature deviation parameters and the third feature deviation parameters; the first characteristic deviation parameter is a characteristic distance between a characteristic vector of the extended sample learning data and a template of a target air conditioner load prediction tag corresponding to the extended sample learning data; the second characteristic deviation parameter is: the feature distance between the feature vector of the extended sample learning data and a template of an air conditioner load prediction tag except the target air conditioner load prediction tag in the optimized extended template; or the second characteristic deviation parameter is: the characteristic distance between the characteristic vector of the extended sample learning data and a template of an air conditioner load prediction tag in the energy consumption reference template; the third characteristic deviation parameter is a characteristic distance between a template of an air conditioner load prediction tag except the target air conditioner load prediction tag in the optimized extension template and a template of an air conditioner load prediction tag in the energy consumption reference template;

And determining the sum of target characteristic deviation parameters of all the expansion sample learning data in the multi-sample learning data sequence as the blending training cost parameter.

In some design considerations of the first aspect, the method further comprises:

Receiving candidate air conditioner energy consumption big data;

Loading the candidate air conditioner energy consumption big data to a target air conditioner load prediction model, and generating a load prediction thermodynamic diagram of the candidate air conditioner energy consumption big data;

Determining a load parameter interval corresponding to a maximum predicted thermodynamic value in a load prediction thermodynamic diagram of the candidate air conditioner energy consumption big data as the load parameter interval of the candidate air conditioner energy consumption big data;

The candidate air conditioner energy consumption big data comprise an energy consumption record, a first energy consumption point, a second energy consumption point, a recording area of the first energy consumption point in the energy consumption record and a recording area of the second energy consumption point in the energy consumption record;

the loading the candidate air conditioner energy consumption big data to a target air conditioner load prediction model, and generating a load prediction thermodynamic diagram of the candidate air conditioner energy consumption big data comprises the following steps:

Loading the candidate air conditioner energy consumption big data to the target air conditioner load prediction model, and generating a load prediction thermodynamic diagram between the first energy consumption point and the second energy consumption point;

The outputting the load parameter interval corresponding to the maximum predicted thermodynamic value in the load prediction thermodynamic diagram of the candidate air conditioner energy consumption big data as the load parameter interval of the candidate air conditioner energy consumption big data includes:

And outputting a load parameter interval corresponding to the maximum predicted thermodynamic value in the load prediction thermodynamic diagram between the first energy consumption point and the second energy consumption point as the load parameter interval between the first energy consumption point and the second energy consumption point.

According to a second aspect of the present application, there is provided an air conditioning load prediction system based on big data analysis, the air conditioning load prediction system based on big data analysis includes a machine-readable storage medium storing machine executable instructions and a processor, the processor implementing the aforementioned air conditioning load prediction method based on big data analysis when executing the machine executable instructions.

According to a third aspect of the present application, there is provided a computer-readable storage medium having stored therein computer-executable instructions that, when executed, implement the aforementioned big data analysis-based air conditioner load prediction method.

According to any one of the aspects, the application has the technical effects that:

According to the embodiment of the application, the air conditioner load prediction model comprising the energy consumption vector extraction network, the meteorological influence factor extraction network and the characteristic fusion calibration network is constructed, so that the efficient and accurate prediction of the air conditioner energy consumption is realized, an extended sample learning data sequence and a multiple sample learning data sequence are generated based on past energy consumption data and extended meteorological data, and then the model training is performed by utilizing the data. In the training process, the feature fusion calibration network performs feature fusion calibration on the energy consumption point state description of the expansion template and the sample learning data to generate optimized energy consumption point state description of the expansion template and the sample learning data, so that the prediction accuracy of the model is improved. Finally, the target air conditioner load prediction model obtained through training can be used for predicting load prediction thermodynamic diagrams of candidate air conditioner energy consumption big data, and powerful support is provided for optimizing operation and energy management of an air conditioner system.

That is, a learning data sequence is generated by using past energy consumption data and meteorological data, and then learning is performed through a constructed prediction model and a prediction thermodynamic diagram is generated, including deep analysis and processing of the data using an energy consumption vector extraction network, a meteorological influence factor extraction network and a feature fusion calibration network. The design makes the air-conditioner load prediction more accurate and has higher reliability. In the feature fusion calibration stage, feature fusion calibration can be performed on the energy consumption point state description of the expansion template and the sample learning data, so that the optimized expansion template and the optimized energy consumption point state description are generated. The method can improve the accuracy of data and effectively reduce prediction errors caused by factors such as weather changes. Therefore, by comprehensively considering the past energy consumption data and the weather influencing factors, accurate and reliable air conditioner load prediction can be provided.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings that are needed in the embodiments will be briefly described below, it being understood that the following drawings only illustrate some embodiments of the present application and therefore should not be considered as limiting the scope, and other related drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a schematic flow chart of an air conditioner load prediction method based on big data analysis according to an embodiment of the present application;

fig. 2 is a schematic component structure diagram of an air conditioner load prediction system based on big data analysis for implementing the air conditioner load prediction method based on big data analysis according to an embodiment of the present application.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions of the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application, and it should be understood that the accompanying drawings in the present application are for the purpose of illustration and description only, and are not intended to limit the scope of the present application. In addition, it should be understood that the schematic drawings are not drawn to scale. A flowchart, as used in this disclosure, illustrates operations implemented in accordance with some embodiments of the present application. It should be understood that the operations of the flow diagrams may be implemented out of order and that steps without logical context may be performed in reverse order or concurrently. Furthermore, one skilled in the art, under the direction of this disclosure, may add at least one other operation to the flowchart and may destroy at least one operation from the flowchart.

Fig. 1 is a schematic flow chart illustrating an air conditioner load prediction method and a system based on big data analysis according to an embodiment of the present application, and it should be understood that in other embodiments, the order of part of the steps in the air conditioner load prediction method based on big data analysis according to the present application may be shared with each other according to actual needs, or part of the steps may be omitted or maintained. The air conditioner load prediction method based on big data analysis comprises the following detailed steps:

Step S110, in the construction task of the air conditioner load prediction model of each stage, an extended sample learning data sequence and a multi-sample learning data sequence are obtained based on a past energy consumption data set and an extended meteorological data set, and the air conditioner load prediction model comprises an energy consumption vector extraction network for initializing weight parameters, a meteorological influence factor extraction network and a characteristic fusion calibration network.

The past energy consumption data set refers to historical air conditioner energy consumption data, and generally comprises energy consumption of different time periods, such as energy consumption data of each hour, each day, each week or each month. For example, suppose an office building records energy consumption data of its air conditioning system every hour over the past year, which data constitutes a past energy consumption data set. These data are valuable for understanding the energy consumption pattern of the air conditioning system and predicting future energy consumption.

The extended meteorological data set refers to meteorological data related to air conditioner energy consumption, and the meteorological data comprise temperature, humidity, wind speed, solar radiation and the like. These data may affect the operation and energy consumption of the air conditioning system. For example, if the air temperature in a region is very high, the air conditioning system may require more energy to cool the indoor space. Therefore, the future energy consumption of the air conditioning system can be predicted more accurately by combining the meteorological data with the energy consumption data.

The extended sample learning data sequence is a data sequence formed by combining meteorological data with energy consumption data and is used for training a meteorological influence factor extraction network. It helps the model learn how to extract features from the meteorological data that affect air conditioning energy consumption. For example, assuming that there is a weekly hour meteorological data and corresponding air conditioning energy consumption data, which are arranged in time order, an extended sample learning data sequence is formed. This sequence can be used to train a model that can identify the relationship between the air-bearing data and the energy consumption data.

The multi-sample learning data sequence is a data sequence which simultaneously contains a plurality of characteristics (energy consumption data and meteorological data) and is used for training the whole air conditioner load prediction model. These data sequences contain all the information required for model learning. Continuing with the above example, if the energy consumption data per hour and the corresponding weather data (e.g., temperature, humidity, etc.) are combined, a multiple sample learning data sequence is formed. This sequence contains all the input information needed by the model to predict future energy consumption.

The energy consumption vector extraction network is one part of an air conditioner load prediction model and is responsible for extracting a characteristic vector from energy consumption data. The feature vectors contain key information of the energy consumption data, and are helpful for accurate prediction of the model. For example, assuming that the energy consumption data includes energy consumption information of peak and off-peak periods, the energy consumption vector extraction network may learn to identify these periods and extract corresponding feature vectors, such as average energy consumption of peak periods, energy consumption fluctuation of off-peak periods, and the like.

The meteorological influence factor extraction network is another part of the air conditioner load prediction model and is specially used for extracting characteristics influencing the energy consumption of the air conditioner from meteorological data. These features may help the model better understand the relationship between meteorological data and energy consumption. For example, the weather-influencing factor extraction network may learn to identify severe weather conditions such as high temperature, high humidity, etc., and extract corresponding features such as duration of the high temperature period, correlation between humidity and energy consumption, etc. This information is very helpful for predicting future energy consumption.

The characteristic fusion calibration network is the last part of the air conditioner load prediction model and is responsible for fusing and calibrating the characteristics extracted by the energy consumption vector extraction network and the weather influence factor extraction network. Through this process, the air conditioner load prediction model can generate a more accurate prediction result. For example, assume that the energy consumption vector extraction network extracts the average energy consumption characteristics of the peak period, while the weather-influencing factor extraction network extracts the duration characteristics of the high-temperature period. The feature fusion calibration network may combine these two features to generate a more comprehensive representation of the features for predicting future air conditioning energy consumption during high temperature peak periods.

Thus, in the present embodiment, the air conditioning load prediction system based on the big data analysis starts to perform the task of constructing the air conditioning load prediction model as a server. First, a past energy consumption data set can be obtained from a database, and at the same time, the server also obtains an extended meteorological data set. The server combines the two types of data to form an extended sample learning data sequence and a multiple sample learning data sequence. The extended sample learning data sequence mainly comprises meteorological data and is used for training a meteorological influence factor extraction network to generate an extended template. The multi-sample learning data sequence comprises energy consumption data and corresponding meteorological data and is used for training the whole air conditioner load prediction model.

In the process of constructing the model, the server needs to initialize the weight parameters of the energy consumption vector extraction network, the meteorological influence factor extraction network and the characteristic fusion calibration network. These weight parameters are key to model learning and they are continually adjusted during the training process to enable the model to better fit the data.

Step S120, for each sample learning data in the multi-element sample learning data sequence, taking the sample learning data as model loading data of the air conditioner load prediction model, generating a load prediction thermodynamic diagram of the sample learning data, wherein the feature fusion calibration network is used for performing feature fusion calibration on an expansion template and an energy consumption point state description of the sample learning data, generating an optimized expansion template and an optimized energy consumption point state description of the sample learning data, and the expansion template is a generation result of the expansion sample learning data sequence through the weather influence factor extraction network.

In this embodiment, the sample learning data refers to a single data point or a piece of data selected from a multi-sample learning data sequence, which is used as an input of a model for learning and prediction. Each sample learning data includes energy consumption data and corresponding meteorological data. For example, suppose that one day of data including 24 hours of energy consumption data and corresponding weather data (such as temperature, humidity, etc.) is randomly selected from a multi-sample learning data sequence of one year as sample learning data.

The model loading data refers to that sample learning data is input into an air conditioner load prediction model and used as a basis for calculation and prediction of the model. The model adjusts its internal parameters based on the data and generates a prediction result. Continuing with the above example, the data of the selected day is input as model loading data into the model, and the model computes and predicts based on the data to generate the air conditioning load prediction result of the day.

The load prediction thermodynamic diagram is a visual representation of the predicted outcome of the model generated from the input data. The thermodynamic diagram represents the air conditioner load in different time periods through the darkness of the color, so that the prediction result can be conveniently and intuitively checked and analyzed. For example, assuming that the model generates a load prediction result that is a 24 hour load curve, the curve may be converted into a thermodynamic diagram in which a darker color indicates a greater load and a lighter color indicates a lesser load. Thus, the air conditioning load conditions of different time periods in one day can be quickly known by looking at the thermodynamic diagrams.

The expansion template is a result generated after the expansion sample learning data sequence is processed through the meteorological influence factor extraction network. The method comprises the step of extracting characteristics affecting the energy consumption of the air conditioner from meteorological data and is used for a subsequent characteristic fusion calibration process. For example, assuming that there is a week of extended sample learning data sequence (including only meteorological data), meteorological features such as average temperature, maximum temperature, minimum temperature, humidity change, etc. within the week can be extracted by the process of the meteorological influence factor extraction network, and these features are combined into one extended template.

The energy consumption point state description refers to information for describing and representing the energy consumption data in the sample learning data. It may include characteristics of energy consumption, trend of change, fluctuation, etc. for reflecting the state and characteristics of the energy consumption data. For example, assuming that one sample learning data contains energy consumption data within one hour of a certain building, average energy consumption, maximum energy consumption, minimum energy consumption, fluctuation of energy consumption, and the like within the hour can be calculated, and these information are described as energy consumption point states. This information helps the model to better understand the nature and regularity of the energy consumption data.

The feature fusion calibration refers to the process of fusing and calibrating the expansion template and the energy consumption point state description. Through the process, the information of the meteorological data and the energy consumption data can be combined to generate a more accurate prediction result. The feature fusion calibration network is responsible for accomplishing this task. For example, assume that an extended template describes meteorological features for a week, while an energy consumption point state description describes energy consumption for an hour. The feature fusion calibration network can fuse the two information, consider the influence of meteorological features on energy consumption, and generate a more accurate prediction result. For example, in high temperature and humidity situations, the air conditioning system may need to consume more energy to maintain indoor comfort temperatures, which information may be captured and utilized by the model through a feature fusion calibration process.

Thus, in the present embodiment, the server starts processing each sample learning data in the multiple sample learning data sequence. For each sample learning data, the server inputs it into the air conditioning load prediction model. The energy consumption vector extraction network in the air conditioner load prediction model can extract the characteristics of energy consumption data, the weather influence factor extraction network can extract the characteristics of air conditioner data, and then the characteristic fusion calibration network can carry out fusion calibration on the two types of characteristics.

The server generates a load prediction thermodynamic diagram for each sample of learning data by calculation of the model. This load prediction thermodynamic diagram presents the air conditioning load prediction results of the air conditioning load prediction model for different time periods in an intuitive manner.

And step S130, performing model training on the air conditioner load prediction model based on the load prediction thermodynamic diagram of each sample learning data in the multi-sample learning data sequence and the air conditioner load labeling data of each sample learning data until the model convergence requirement is met.

In this embodiment, the air conditioner load labeling data refers to real labeled air conditioner load data, and is used as a reference standard for model training. These data are typically obtained by monitoring and recording the energy consumption of the actual air conditioning system when it is operating. For example, suppose an office building is monitored for energy consumption by an air conditioning system for one month, and energy consumption data per hour is recorded. The data can be used as air conditioning load labeling data for training the model and evaluating the prediction accuracy of the model.

Thus, the air conditioning load prediction model can be trained using the load prediction thermodynamic diagram and the air conditioning load annotation data. By comparing the difference between the prediction result of the model and the labeling data, the parameters of the model are continuously adjusted, so that the model can predict the air conditioner load more accurately. For example, in the model training process, each sample learning data in the multi-sample learning data sequence is input into the model to obtain a corresponding load prediction thermodynamic diagram. Then, these thermodynamic diagrams are compared with corresponding air conditioning load label data to calculate a prediction error. According to the size and direction of the error, parameters of the model are adjusted, so that the model can generate a more accurate prediction result in the next iteration.

The model convergence requirement refers to a standard or condition set in the model training process, and is used for judging whether the model has reached sufficient prediction accuracy or stability. When the model meets this requirement, the training process is stopped and the current model is determined as the target air conditioning load prediction model. Assume that a model convergence requirement is set for a prediction error of less than 5%. In the model training process, the prediction error is continuously calculated and judged whether the prediction error is less than 5%. If the prediction error is still greater than 5% after several continuous iterations, the model is not converged, and training needs to be continued; if the prediction error is less than 5%, indicating that the model has converged, training may be stopped and the current model may be determined as the target air conditioning load prediction model.

That is, in this embodiment, the server performs model training using the load prediction thermodynamic diagram and the corresponding air conditioner load annotation data. For example, the loss function value of the model is calculated by comparing the difference between the load prediction thermodynamic diagram and the corresponding air conditioning load signature data. The loss function value reflects the magnitude of the error between the model predicted result and the true result.

The server uses an optimization algorithm to adjust the weight parameters of the air conditioning load prediction model to minimize the loss function value. This process is repeated until the predictive performance of the air conditioning load predictive model reaches the convergence requirement. The convergence requirement may be determined by setting a certain threshold, for example, when the change in the loss function value of several consecutive iterations is smaller than a certain threshold, the air conditioning load prediction model is considered to have converged.

During the training process, the server may also use the validation data set to validate the performance of the air conditioning load prediction model. The verification data set is a part of data which is divided from the original data, does not participate in training of the air conditioner load prediction model, but is used for evaluating the prediction capability of the air conditioner load prediction model. The server may evaluate the generalization performance of the air conditioner load prediction model by comparing the prediction result of the air conditioner load prediction model on the validation data set with the real result.

And step S140, determining the air conditioner load prediction model which meets the model convergence requirement as a target air conditioner load prediction model, wherein the target air conditioner load prediction model is used for predicting a load prediction thermodynamic diagram of candidate air conditioner energy consumption big data.

The target air conditioner load prediction model is an air conditioner load prediction model which is determined after training and meeting the model convergence requirement. The model has better prediction capability and can be used for predicting the load condition of new and unknown air conditioner energy consumption data. For example, assuming an air conditioning load prediction model that is trained and meets convergence requirements, this model can now be used to predict the energy consumption of an air conditioning system of a newly built office building for one month in the future. The corresponding load prediction thermodynamic diagram can be obtained only by inputting relevant data (such as building area, using function, meteorological data and the like) of the newly built office building into the model, so that the energy consumption change trend and peak value condition of the air conditioning system in the future month are known.

The candidate air conditioner energy consumption big data refers to a new air conditioner energy consumption data set to be predicted. Such data may be from different buildings, different time periods, or different weather conditions for testing the predictive capabilities of the target air conditioning load predictive model. For example, assuming a target air conditioning load prediction model, it is now necessary to test the predictive capability of this model under different meteorological conditions. The air conditioner energy consumption data under different seasons and different weather conditions can be collected to serve as candidate air conditioner energy consumption big data, and the data are input into the model to be predicted. By comparing the difference between the prediction result and the actual energy consumption data, the prediction accuracy and generalization capability of the model can be evaluated.

That is, when the model meets the convergence requirement, the server determines it as the target air conditioning load prediction model. The target air conditioner load prediction model is fully trained and optimized, and has good prediction capability. The server can input the new air conditioner energy consumption big data into the target air conditioner load prediction model to generate a corresponding load prediction thermodynamic diagram. These load prediction thermodynamic diagrams may provide an important reference basis for optimal operation of the air conditioning system. For example, the operation strategy of the air conditioner is adjusted according to the load prediction thermodynamic diagram, the energy consumption distribution is optimized, and the like, so that the aims of energy conservation and emission reduction are achieved.

Based on the steps, the embodiment of the application realizes the efficient and accurate prediction of the air conditioner energy consumption by constructing the air conditioner load prediction model comprising the energy consumption vector extraction network, the weather influence factor extraction network and the characteristic fusion calibration network, firstly generates an extended sample learning data sequence and a multi-sample learning data sequence based on past energy consumption data and extended weather data, and then utilizes the data to carry out model training. In the training process, the feature fusion calibration network performs feature fusion calibration on the energy consumption point state description of the expansion template and the sample learning data to generate optimized energy consumption point state description of the expansion template and the sample learning data, so that the prediction accuracy of the model is improved. Finally, the target air conditioner load prediction model obtained through training can be used for predicting load prediction thermodynamic diagrams of candidate air conditioner energy consumption big data, and powerful support is provided for optimizing operation and energy management of an air conditioner system.

In one possible implementation, step S120 may include:

Step S121, loading the extended sample learning data sequence to the weather-influencing factor extraction network, and generating the extended template.

In this embodiment, the server first obtains a one week extended sample learning data sequence that includes meteorological information (e.g., temperature, humidity, wind speed, etc.) for each hour. The server loads these data into the weather-influencing factor extraction network. The weather influencing factor extraction network is processed and analyzed to extract key weather features in the week, such as average temperature, highest temperature, lowest temperature, humidity change trend and the like, and the features are combined into an expansion template. This extended template provides important weather information for subsequent feature fusion calibration.

And step S122, taking the sample learning data as the model loading data of the energy consumption vector extraction network, and generating the energy consumption characteristic vector of the sample learning data.

The server then obtains a specific sample learning data that includes air conditioning energy consumption information for a given building during a day. The server loads this sample learning data into the energy consumption vector extraction network. The network extracts the energy consumption characteristic vector in the day through learning and analysis, such as average energy consumption, peak energy consumption, energy consumption fluctuation condition and the like. These feature vectors reflect the energy consumption characteristics and laws of the building air conditioning system during the day.

And step S123, taking the sample learning data as the model loading data of the meteorological influence factor extraction network, and generating an extended meteorological influence vector of the sample learning data.

The server again uses the weather-influencing factor extraction network, but this time processes the weather portions of the sample learning data. In this way, the server can extract the weather-influencing vector directly related to the specific energy consumption data, such as the air temperature, humidity during energy consumption peaks, etc. These weather effect vectors will provide more accurate weather information for subsequent feature fusion calibration.

Step S124, generating an energy consumption point state description of the sample learning data based on the energy consumption reference template obtained by initializing model learning, the expansion template, a fusion function, an energy consumption characteristic vector of the sample learning data and an expansion weather influence vector of the sample learning data, wherein the fusion function is determined by focusing a focusing attention function.

The server generates an energy consumption point state description based on an energy consumption reference template, an expansion template, a fusion function and an energy consumption characteristic vector and an meteorological influence vector extracted from sample learning data, which are obtained by initializing model learning. The description integrates the information of energy consumption data and air condition data, and provides comprehensive characteristic representation for predicting air conditioner load. The fusion function is determined by focusing on a function that helps the server focus on key features in generating the energy consumption point state description.

And step S125, performing feature fusion calibration on the energy consumption point state description of the expansion template and the sample learning data to generate an optimized energy consumption point state description of the expansion template and the sample learning data.

And the server performs feature fusion calibration on the expansion template and the energy consumption point state description. In the process, the server fuses and calibrates the meteorological features and the energy consumption features, and the influence of meteorological conditions on the energy consumption is considered. Through this process, the server generates an optimized expansion template and an optimized energy consumption point state description, and these optimized feature representations will more accurately reflect the energy consumption conditions of the air conditioning system.

Step S126, predicting and outputting a load prediction thermodynamic diagram of the sample learning data based on the energy consumption reference template, the optimized expansion template and the optimized energy consumption point state description of the sample learning data.

Finally, the server predicts a load prediction thermodynamic diagram outputting sample learning data based on the energy consumption reference template, the optimized extension template and the optimized energy consumption point state description. This thermodynamic diagram visually shows the load of the building air conditioning system in different time periods by the shade of the color. The server may provide this thermodynamic diagram to the user or for other subsequent analysis to help the user better understand the energy consumption of the air conditioning system and to formulate energy conservation measures.

In one possible implementation, step S125 may include:

Step S1251, calculating a focusing attention coefficient between the energy consumption point state description of the sample learning data and each of the extended sample learning data in the extended sample learning data sequence.

The server starts to perform the first step of feature fusion calibration, namely calculating a focusing attention coefficient between the energy consumption point state description of the sample learning data and each of the extended sample learning data in the extended sample learning data sequence. These coefficients reflect the correlation and importance between the energy consumption point state description and each expansion sample learning data. The server determines these coefficients by comparing the similarity of the energy consumption point profile to the characteristics of the data learned for each of the expansion samples, such as the meteorological conditions of temperature, humidity, etc. A higher coefficient indicates a stronger correlation between the energy consumption point state description and the corresponding expansion sample learning data, and therefore more attention should be paid in the subsequent feature fusion.

Step S1252, according to the energy consumption point state description of the sample learning data and the focusing attention coefficient between each extended sample learning data in the extended sample learning data sequence, fusing the extended sample learning data in the extended sample learning data sequence through the feature vector of the meteorological influence factor extraction network, and generating the optimized extended template.

And the server fuses the extended sample learning data in the extended sample learning data sequence according to the calculated focusing attention coefficient by extracting the characteristic vector of the network through the weather influence factor. In this process, the server will consider the feature vector of each expansion sample learning data weighted, where the weights are determined by the focus attention coefficients. In this way, the server can generate an optimized extended template that integrates the weather information most relevant to the description of the energy consumption point states, thereby improving the accuracy of the subsequent load predictions.

Step S1253, generating an optimized energy consumption point state description of the sample learning data based on the energy consumption reference template, the optimized expansion template, the fusion function, the energy consumption feature vector of the sample learning data, and the expanded weather influence vector of the sample learning data.

After the optimized expansion template is generated, the server continues to generate optimized energy consumption point state description of the sample learning data based on the energy consumption reference template, the optimized expansion template, the fusion function, the energy consumption characteristic vector of the sample learning data and the extended weather influence vector of the sample learning data. In the process, the server fuses the energy consumption characteristic vector, the extended weather influence vector and the optimized extended template, and the interaction and influence among all factors are comprehensively considered through a fusion function. The resulting optimized energy consumption point state description is a more comprehensive, accurate representation of the features that will be used in subsequent load prediction thermodynamic diagram generation.

In one possible implementation, step S125 further includes: and carrying out feature fusion calibration optimization on the energy consumption point state description of the expansion template and the sample learning data based on preset feature fusion calibration parameters, and generating optimized energy consumption point state description of the expansion template and the sample learning data.

In this embodiment, the server starts the step of performing feature fusion calibration. Previously, a set of preset feature fusion calibration parameters are prepared, which are set based on historical data, model training experience, and domain knowledge, to guide the process of feature fusion.

The server first loads the energy consumption point state description of the expansion template and the sample learning data. The expansion template contains characteristic information of meteorological factors, and the energy consumption point state description captures the characteristics of energy consumption data. Both pieces of information are critical for subsequent load prediction.

And then, the server fuses the two parts of information by using a preset characteristic fusion calibration parameter. This process may involve weight adjustment, feature combination, and screening operations. For example, the server may increase the weight of some important features or delete some redundant or less relevant features as directed by the parameters.

The server may also take into account interactions and effects between the different features in this fusion process. Through comprehensive adjustment and processing, the server generates an optimized energy consumption point state description of the optimized expansion template and sample learning data. The optimized characteristic representations not only contain key information of the original data, but also are more suitable for subsequent load prediction tasks.

Finally, the server stores the optimized expansion template and the optimized energy consumption point state description for the subsequent load prediction model. These optimized features will help to improve the accuracy and stability of load prediction.

In a possible implementation manner, the performing feature fusion calibration optimization on the energy consumption point state description of the expansion template and the sample learning data based on the preset feature fusion calibration parameters, and generating an optimized energy consumption point state description of the expansion template and the sample learning data includes:

Step A110, calculating focusing attention coefficients between the k-1 th round of optimized energy consumption point state description of the sample learning data and each extended sample learning data in the extended sample learning data sequence, wherein k is greater than 0, and the 0 th round of optimized energy consumption point state description of the sample learning data is the energy consumption point state description of the sample learning data.

And step A120, according to the k-1 th round of optimized energy consumption point state description of the sample learning data and the focusing attention coefficient between each expansion sample learning data in the expansion sample learning data sequence, fusing the expansion sample learning data in the expansion sample learning data sequence through the characteristic vector of the meteorological influence factor extraction network, and generating a k round of optimized expansion template.

And step A130, generating a kth round of optimized energy consumption point state description of the sample learning data based on the energy consumption reference template, the kth round of optimized expansion template, the fusion function, the energy consumption characteristic vector of the sample learning data and the expanded weather influence vector of the sample learning data.

In detail, the following is detailed.

Initial state (k=0);

The server first has an energy consumption point state description of the sample learning data (as a round 0 optimization energy consumption point state description) and expands the sample learning data sequence. In addition, the server is also provided with a meteorological influence factor extraction network for extracting feature vectors from the extended sample learning data. The maximum iteration number K of the energy consumption reference template, the fusion function and the feature fusion calibration is also preset.

Iterative step (k > 0);

After the k-1 th round of optimization (k is greater than 0), the server calculates a focus attention coefficient between the k-1 th round of optimized energy consumption point state description of the sample learning data and each of the extended sample learning data in the extended sample learning data sequence. These coefficients reflect the similarity or correlation between the current optimized energy consumption point state description and the various expansion sample learning data. The calculation of the focus attention coefficient may be based on similarity measurement between features, statistical correlation analysis, and the like.

The server uses the focusing attention coefficient obtained by the previous calculation to guide the fusion of the expansion sample learning data in the expansion sample learning data sequence. Specifically, the server extracts feature vectors of the learning data of each expansion sample through the meteorological influence factor extraction network, and performs weighted fusion on the feature vectors according to the focusing attention coefficient. The weighted fusion may involve a linear combination or a non-linear transformation of the different feature vectors, where the weights are determined by the focus attention coefficients. In this way, the server generates an extended template for the kth round of optimization that focuses more on the weather information related to the current optimized energy consumption point state description.

After the k-th round of optimized expansion template is obtained, the server combines the energy consumption reference template, the optimized expansion template, the fusion function, the energy consumption characteristic vector of the sample learning data and the expanded weather influence vector to generate a k-th round of optimized energy consumption point state description of the sample learning data. This process may involve further fusion and transformation of the plurality of feature vectors to generate a more comprehensive, finer representation of the features to describe the state of the energy consumption point. The fusion function plays a key role here, which defines how these different feature vectors are effectively combined together.

The server will repeat the above steps (calculating focusing attention coefficient, generating optimizing expansion template, generating optimizing energy consumption point state description) until reaching the preset maximum iteration number K. Each iteration is adjusted and improved based on the optimization result of the previous round so as to gradually approach the optimal feature fusion calibration result.

After K rounds of iterative optimization, the server obtains a final optimization expansion template and an optimization energy consumption point state description. These optimized feature representations will be used in subsequent load prediction tasks to improve the accuracy and reliability of the predictions. The overall feature fusion calibration optimization process aims to capture more complex nonlinear relationships and useful information hidden in the data by iteratively adjusting and refining the feature representation.

In one possible implementation, step S130 may include:

Step S131, determining a training cost parameter based on the load prediction thermodynamic diagram of each sample learning data in the multi-sample learning data sequence and the air conditioner load labeling data of each sample learning data.

And step S132, based on the training cost parameters, updating the network parameter definition information of the meteorological influence factor extraction network by using a gradient descent algorithm until the model convergence requirement is met.

The server firstly receives a plurality of sample learning data sequences, wherein the data sequences comprise load prediction thermodynamic diagrams of the sample learning data and corresponding air conditioner load marking data. The load prediction thermodynamic diagram is a preliminary prediction result of the model for each sample learning data, and the air conditioner load labeling data is a real load value.

The server next needs to determine training cost parameters based on these data. The training cost parameter is an index for measuring the difference between the model predicted result and the real result, and is also a key parameter in the model optimization process. Typically, the calculation of the training cost parameter involves an error calculation between the predicted value and the true value, such as a mean square error, cross entropy loss, etc.

In this scenario, the server would traverse each sample of learning data, calculate the error between its load prediction thermodynamic diagram and the corresponding air conditioning load annotation data, and accumulate these errors to get the total training cost. The server may also normalize or weight the errors as needed to better reflect the model's behavior on different instances.

Once the training cost parameters are determined, the server may begin updating the network parameter definition information of the weather-influencing factor-extracting network with a gradient descent algorithm. The gradient descent algorithm is a commonly used optimization algorithm that directs the direction of updating of the parameters by calculating the gradient of the cost function with respect to the model parameters, thereby minimizing the cost function.

In this scenario, the server first calculates a gradient of the cost function based on the current network parameters. This gradient indicates the direction and magnitude in which the network parameters should be adjusted in order for the cost function to decrease. The server will then update the network parameters in the direction of this gradient. The magnitude of the update is typically controlled by a learning rate, which is a super-parameter that determines the step size of the parameter adjustment for each update of the model.

The server repeats the above process until the cost function reaches a preset threshold or the updated amplitude is less than a small value, at which point the model may be considered to have converged. The converged model is a trained air conditioner load prediction model, which can be used for load prediction of new meteorological data.

In one possible implementation, the multiple sample learning data sequence includes extended sample learning data and energy consumption reference sample learning data, and step S130 may include:

And step S133, calculating blending training cost parameters based on the various expansion sample learning data in the multi-sample learning data sequence, the optimized expansion templates and the energy consumption reference templates obtained by initializing parameter learning.

Step S134, constructing training cost parameters based on the load prediction thermodynamic diagram of each sample learning data in the multi-sample learning data sequence and the air conditioner load labeling data of each sample learning data. And updating the network parameter definition information of the meteorological influence factor extraction network by using a gradient descent algorithm based on the blending training cost parameter and the training cost parameter until the model convergence requirement is met.

In this embodiment, the server first processes a sequence of multivariate sample learning data including extended sample learning data and energy consumption reference sample learning data. The server has generated an optimized extension template through feature fusion calibration optimization and has an initialized energy consumption reference template.

When processing the extended sample learning data, the server calculates blending training cost parameters in combination with the optimized extended templates. The blending training cost parameter reflects the degree of difference or matching between the expansion sample learning data and the optimized expansion template. For calculating the blending training cost parameter, the server may use various measurement methods, such as mean square error, cosine similarity, etc., to measure the similarity or distance between the features of the extended sample learning data and the optimized extended template.

For the energy consumption reference sample learning data, the server can also compare the energy consumption reference sample learning data with the energy consumption reference template, and calculate corresponding blending training cost parameters. These parameters will be used in the subsequent model training process to help the model learn and adapt better to different data distributions and feature patterns.

In addition to blending the training cost parameters, the server also needs to construct the training cost parameters based on the load prediction thermodynamic diagram of each sample learning data in the multi-sample learning data sequence and the corresponding air conditioner load annotation data. The load prediction thermodynamic diagram is a preliminary prediction result of the model for each sample learning data, and the air conditioner load labeling data is a real load value.

The server traverses each sample learning data and calculates the error between the load prediction thermodynamic diagram and the corresponding air conditioner load labeling data. These errors may be calculated by means of a metric such as mean square error, absolute error, etc. And the server accumulates errors of the learning data of all the samples to obtain the total training cost parameter. This parameter reflects the overall predictive performance of the model in the current state.

After the blended training cost parameters and the training cost parameters are obtained, the server starts to update the network parameter definition information of the meteorological influence factor extraction network by using a gradient descent algorithm. The gradient descent algorithm is an optimization algorithm that directs parameter updates by calculating the gradient of the cost function with respect to the model parameters.

The server first calculates the gradient of the cost function based on the current network parameters. This gradient indicates the direction and magnitude in which the network parameters should be adjusted in order for the cost function to decrease. The server will then update the network parameters in the direction of this gradient. The magnitude of the update is typically controlled by a learning rate, which is a super-parameter that determines the step size of the parameter adjustment for each update of the model.

In the process of updating the network parameters, the server can consider the influence of the blended training cost parameters and the training cost parameters at the same time. Blending training cost parameters helps the model to learn and adapt to the characteristic mode of the extended sample learning data better, and training cost parameters ensure that the model can accurately predict the air conditioner load annotation data. By comprehensively considering the influence of the two parameters, the server can more comprehensively optimize the model performance.

In one possible embodiment, step S133 may include:

Step S1331, calculating, for each expansion sample learning data, a target feature deviation parameter of the expansion sample learning data, the target feature deviation parameter being a sum of a first target feature deviation parameter determined based on a first definition coefficient, the first feature deviation parameter, and a second feature deviation parameter determined based on a second definition coefficient, the first feature deviation parameter, and a third feature deviation parameter. The first characteristic deviation parameter is a characteristic distance between a characteristic vector of the extended sample learning data and a template of a target air conditioner load prediction tag corresponding to the extended sample learning data. The second characteristic deviation parameter is: and the characteristic distance between the characteristic vector of the extended sample learning data and a template of an air conditioner load prediction tag except the target air conditioner load prediction tag in the optimized extended template. Or the second characteristic deviation parameter is: and the characteristic distance between the characteristic vector of the extended sample learning data and a template of an air conditioner load prediction tag in the energy consumption reference template. And the third characteristic deviation parameter is a characteristic distance between a template of an air conditioner load prediction tag except the target air conditioner load prediction tag in the optimized extension template and a template of an air conditioner load prediction tag in the energy consumption reference template.

And S1332, determining the sum of target characteristic deviation parameters of all the expansion sample learning data in the multi-sample learning data sequence as the blending training cost parameter.

In this embodiment, for each extended sample learning data, the server extracts its feature vector and compares it with the template of the target air conditioning load prediction tag corresponding to the extended sample learning data. The "target air conditioning load prediction tag" herein refers to a true air conditioning load value or category corresponding to the expansion sample learning data currently being processed.

And the server calculates the characteristic distance between the characteristic vector of the extended sample learning data and the template of the corresponding target air conditioner load prediction label to obtain a first characteristic deviation parameter. The feature distance can be calculated by Euclidean distance, cosine similarity and other methods, and the similarity or difference between the feature vector and the template is measured.

And then, the server performs weighting processing on the first characteristic deviation parameter according to a preset first definition coefficient to obtain a weighted first characteristic deviation parameter. The first defining coefficient is a weight value for adjusting the importance of the first feature deviation parameter in calculating the target feature deviation parameter.

Next, the server continues to process the extended sample learning data, calculating a second target feature deviation parameter. The server compares the feature vector of the expansion sample learning data with templates of other air conditioner load prediction tags except the target air conditioner load prediction tag in the optimized expansion template. The optimized expansion template is generated in the previous characteristic fusion calibration optimization process and comprises template information of a plurality of air conditioner load prediction labels.

And the server calculates feature distances between feature vectors of the extended sample learning data and other label templates in the optimized extended templates to obtain second feature deviation parameters. Likewise, this feature distance may be calculated by different metrics.

In addition, the server can also select to compare the characteristic vector of the extended sample learning data with a template of an air conditioner load prediction label in the energy consumption reference template, and calculate to obtain another second characteristic deviation parameter. The energy consumption reference template is obtained by initializing parameter learning and represents a characteristic mode of the energy consumption reference sample learning data.

And the server performs weighting processing on the second characteristic deviation parameters according to a preset second definition coefficient to obtain weighted second characteristic deviation parameters. The second definition coefficient is also a weight value for adjusting the importance of the second feature deviation parameter in calculating the target feature deviation parameter.

Then, the server adds the weighted first feature deviation parameter and the weighted second feature deviation parameter to obtain the target feature deviation parameter of each expansion sample learning data. The target characteristic deviation parameter comprehensively reflects the difference between the extended sample learning data and the corresponding target air conditioner load prediction label template, the optimized extended template and the energy consumption reference template.

And finally, summing the target feature deviation parameters of all the extended sample learning data by the server to obtain blending training cost parameters. The blended training cost parameter measures the overall difference between all the extended sample learning data and the template, and is used in the subsequent model training process to help the model to learn better and adapt to the characteristic mode of the extended sample learning data.

Wherein for each expansion sample learning data (i), its first target feature deviation parameter (d_i ζ1) may be calculated as:

[ D_i^1 = \alpha \times \text{dist}(F_i, T_{\text{target}}^i) ]；

Wherein:

(\alpha) is a first coefficient of definition, a predetermined weight value.

(F_i) is a feature vector of the extension sample learning data (i).

(T_ { \text { target } ] pi) is a template for the target air conditioner load prediction tag corresponding to the extended sample learning data (i).

The (_text { dist } (F_i, T_text { target } ] and }. Pi.) is a characteristic distance between (F_i) and (T_text { target }. Pi.) and can be calculated by Euclidean distance, cosine similarity and other methods.

The calculation of the second target feature deviation parameter (D_i≡2) involves a comparison with the optimized expansion template and the energy consumption reference template. One possible way of calculation is:

[ D_i^2 = \beta \times \text{dist}(F_i, T_{\text{optimized}}^j) + \gamma \times \text{dist}(T_{\text{optimized}}^j, T_{\text{baseline}}^k) ];

Wherein:

The (\beta) and (\gamma) are second defined coefficients, weight values set in advance. Note that here, for simplicity of explanation, we consider the case of only one other label template, and in practice it may be necessary to calculate and average or maximum a plurality of templates, etc.

(T_ { \text { optimized } ] j) is the template of a certain air conditioner load prediction tag other than the target air conditioner load prediction tag in the optimized extension template.

(T_ { \text { baseine } k) is the template of a certain air conditioner load prediction tag in the energy consumption benchmark template.

(_Text { dist } (F_i, T_text { optimized } j)) is the feature distance between (F_i) and (T_text { optimized } ], j).

The ([ text ] { dist } (T_ { text { optimized } ], T_ { text { baseine } ] k)) is the feature distance between (T_ { text { optimized } ] j) and (T_ { text { baseine } ].

Finally, the target feature deviation parameter (d_i) of each of the expansion sample learning data (i) may be calculated as a sum of the first target feature deviation parameter and the second target feature deviation parameter:

[ D_i = D_i^1 + D_i^2 ]；

The blending training cost parameter (c_ { \text { end }) is the sum of the target feature deviation parameters of all the expansion sample learning data:

[ C_{\text{blend}} = \sum_{i=1}^{N} D_i ]；

where (N) is the number of expansion-sample learning data. This blended training cost parameter will be used in the subsequent model training process to help optimize the performance of the model.

In one possible embodiment, the method further comprises:

and step S150, receiving the candidate air conditioner energy consumption big data.

And step S160, loading the candidate air conditioner energy consumption big data to a target air conditioner load prediction model, and generating a load prediction thermodynamic diagram of the candidate air conditioner energy consumption big data.

Step S170, determining a load parameter interval corresponding to the maximum predicted thermodynamic value in the load prediction thermodynamic diagram of the candidate air conditioner energy consumption big data as the load parameter interval of the candidate air conditioner energy consumption big data.

In this embodiment, the server receives the candidate air conditioner energy consumption data transmitted from each air conditioning system or the energy management system. The candidate air conditioner energy consumption big data are collected in real time through a sensor, an intelligent ammeter and other devices, and comprise information such as an energy consumption record of an air conditioner, a first energy consumption point, a second energy consumption point and a recording area of the first energy consumption point and the second energy consumption point in the energy consumption record. The data has important significance in the aspects of analyzing the energy consumption mode of the air conditioner, predicting future load, optimizing energy use and the like.

And the server loads the received candidate air conditioner energy consumption big data into a pre-trained target air conditioner load prediction model. This model may be constructed based on machine learning or deep learning algorithms, and is capable of predicting future air conditioning loads from historical data and real-time data. By inputting the candidate air conditioner energy consumption big data, the model can generate a corresponding load prediction thermodynamic diagram, and the prediction thermodynamic values of different load parameter intervals can be intuitively displayed.

After the candidate air conditioner energy consumption big data are loaded, the target air conditioner load prediction model starts to operate, and a load prediction thermodynamic diagram is generated according to the input data. The thermodynamic diagram is a two-dimensional graph, the horizontal axis represents the load parameter interval, and the vertical axis represents the predicted thermodynamic value. Different colors or shades represent different predicted thermal values, so that it is possible to intuitively see which load parameter intervals have higher predicted thermal values, i.e. higher air conditioning loads may occur in the future.

And the server finds a load parameter interval corresponding to the maximum predicted thermodynamic value by analyzing the load prediction thermodynamic diagram, and determines the load parameter interval as the load parameter interval of the candidate air conditioner energy consumption big data. The load parameter interval represents the most probable air conditioner load range in a future period of time, and has important guiding significance in the aspects of energy management strategy preparation, air conditioner system operation optimization and the like. For example, according to the load parameter interval, parameters such as an operation mode, a set temperature and the like of the air conditioning system can be adjusted so as to realize more efficient energy utilization and more comfortable indoor environment.

The candidate air conditioner energy consumption big data comprise an energy consumption record, a first energy consumption point, a second energy consumption point, a recording area of the first energy consumption point in the energy consumption record and a recording area of the second energy consumption point in the energy consumption record.

Step S160 may include: and loading the candidate air conditioner energy consumption big data to the target air conditioner load prediction model, and generating a load prediction thermodynamic diagram between the first energy consumption point and the second energy consumption point.

Step S170 may include: and outputting a load parameter interval corresponding to the maximum predicted thermodynamic value in the load prediction thermodynamic diagram between the first energy consumption point and the second energy consumption point as the load parameter interval between the first energy consumption point and the second energy consumption point.

The candidate air conditioner energy consumption big data received by the server contains a plurality of key information. The energy consumption record is energy consumption data of the air conditioning system in a period of time, and may include power, electric quantity and the like. The first energy consumption point and the second energy consumption point are two specific points in the energy consumption record, which may be energy consumption peaks, energy consumption valleys or other important points specified by the user. The recording area of these points in the energy consumption record refers to their position and time range throughout the energy-consuming time sequence.

And the server loads the received candidate air conditioner energy consumption big data into the target air conditioner load prediction model. This model is pre-trained and can predict future air conditioning loads based on historical data and real-time data. The server is particularly concerned with the data between the first energy consumption point and the second energy consumption point when loading the data, as these two points may represent important features of the energy consumption or intervals of particular interest to the user. The model generates a load prediction thermodynamic diagram between the first energy consumption point and the second energy consumption point based on the data.

And after the server loads the candidate air conditioner energy consumption big data to the target air conditioner load prediction model, the model starts to operate and generates a load prediction thermodynamic diagram. This thermodynamic diagram is focused on showing the load prediction situation between the first energy consumption point and the second energy consumption point. The color or shade in the thermodynamic diagram represents the predicted thermodynamic value magnitudes for the different load parameter intervals so that it can be intuitively seen which load parameters have higher predicted thermodynamic values within this particular interval.

And the server finds a load parameter interval corresponding to the maximum predicted thermodynamic value by analyzing a load prediction thermodynamic diagram between the first energy consumption point and the second energy consumption point. This interval represents the range of air conditioning loads that are most likely to occur in the future between these two specific energy consumption points. And the server outputs the load parameter interval as a load prediction result of the candidate air conditioner energy consumption big data. This result may be used to guide optimal operation of the air conditioning system, formulate energy management strategies, and the like. For example, according to the load parameter interval, parameters such as an operation mode, a set temperature and the like of the air conditioning system can be adjusted so as to realize more efficient energy utilization and more comfortable indoor environment.

It should be noted that, in other extended embodiments of the present application, the air conditioning load may be identified by using the user load (difference) in spring and summer, and the model may be trained by using a neural network (or other algorithms) to improve the prediction accuracy, where the following is a technical scheme implementation process:

First, user load data including power consumption records, temperature, humidity, and the like in spring and summer are collected, and the collected data are preprocessed, such as denoising, missing value filling, normalization, and the like.

Next, the preprocessed user load data is subjected to a difference process, that is, the load variation amount at the adjacent time point is calculated. By analyzing the relation between the load variation and meteorological data such as temperature, humidity and the like, the load variation caused by the use of the air conditioner is primarily identified.

On this basis, characteristics related to the air conditioning load, such as temperature difference, humidity difference, time stamp (taking seasonal, workday and day of rest into consideration, etc.), building type, user behavior pattern, etc., can be extracted. An extended feature set is then created, which may include, for example, historical load data, a sliding window average, a maximum, a minimum, etc. of the meteorological data.

Next, an appropriate neural network structure, such as a Deep Neural Network (DNN), a long short term memory network (LSTM), or a Convolutional Neural Network (CNN), is selected and tailored to the nature of the problem. The input layer, hidden layer and output layer of the model are defined at the same time, and the connection mode and activation function between each layer are defined.

Thus, the neural network model can be trained by using the processed characteristics and the corresponding air conditioner load labeling data. During the training process, cross-validation, regularization, early stop and other techniques can be used to prevent overfitting and optimize the performance of the model by adjusting the super parameters (e.g., learning rate, batch size, number of iterations, etc.).

The predictive performance of the model is then evaluated using separate test data sets, common evaluation metrics including Mean Square Error (MSE), root Mean Square Error (RMSE), mean Absolute Error (MAE), and the like. And selecting the model with the best performance as a final prediction model according to the evaluation result.

Finally, the trained model can be deployed into a production environment, real-time user load data and air-conditioning data are received as input, new data are predicted by using the deployed model, and the predicted air-conditioning load is output.

The air conditioner load prediction 100 based on big data analysis shown in fig. 2 includes: a processor 1001 and a memory 1003. The processor 1001 is coupled to the memory 1003, such as via a bus 1002. Optionally, the air conditioning load prediction 100 based on big data analysis may further include a transceiver 1004, and the transceiver 1004 may be used for data interaction between the server and other servers, such as transmission of data and/or reception of data, etc. It should be noted that, the transceiver 1004 is not limited to one in actual scheduling, and the structure of the air conditioner load prediction 100 based on big data analysis does not limit the embodiments of the present application.

The processor 1001 may be a CPU (Central Processing Unit ), general purpose processor, DSP (DIGITAL SIGNAL processor, data signal processor), ASIC (Application SpecificIntegrated Circuit ), FPGA (Field Programmable GATE ARRAY, field programmable gate array) or other programmable logic device, transistor logic device, hardware component, or any combination thereof. Which may implement or perform the various exemplary logic blocks, modules and circuits described in connection with this disclosure. The processor 1001 may also be a combination that implements computing functionality, such as a combination comprising one or more microprocessors, a combination of a DSP and a microprocessor, or the like.

Bus 1002 may include a path to transfer information between the components. Bus 1002 may be a PCI (PERIPHERAL COMPONENT INTERCONNECT, peripheral component interconnect standard) bus, or an EISA (ExtendedIndustry Standard Architecture ) bus, or the like. The bus 1002 may be divided into an address bus, a data bus, a control bus, and the like. For ease of illustration, only one thick line is shown in fig. 2, but not only one bus or one type of bus.

The memory 1003 may be, but is not limited to, ROM (read only memory) or other type of static storage device that can store static information and instructions, RAM (Random Access Memory ) or other type of dynamic storage device that can store information and instructions, EEPROM (ELECTRICALLY ERASABLEPROGRAMMABLE READ ONLY MEMORY ), CD-ROM (Compact DiscRead Only Memory, compact disc read only memory) or other optical disk storage, optical disk storage (including compact discs, laser discs, optical discs, digital versatile discs, blu-ray discs, etc.), magnetic disk storage media, other magnetic storage devices, or any other medium that can be used to carry or store program code and that can be read by a computer.

The memory 1003 is used for storing program codes for executing the embodiments of the present application and is controlled to be executed by the processor 1001. The processor 1001 is configured to execute the program code stored in the memory 1003 to implement the steps shown in the foregoing method embodiment.

Embodiments of the present application provide a computer readable storage medium having program code stored thereon, which when executed by a processor, implements the steps of the foregoing method embodiments and corresponding content.

It should be understood that, although various operation steps are indicated by arrows in the flowcharts of the embodiments of the present application, the order in which these steps are implemented is not limited to the order indicated by the arrows. In some implementations of embodiments of the application, the implementation steps in the flowcharts may be performed in other orders based on demand, unless explicitly stated herein. Furthermore, some or all of the steps in the flowcharts may include a plurality of sub-steps or a plurality of stages, depending on the actual implementation scenario. Some or all of these sub-steps or phases may be performed at the same time, or each of these sub-steps or phases may be performed at different times, respectively. In the case of different execution timings, the execution order of the sub-steps or stages may be flexibly configured based on requirements, which is not limited by the embodiment of the present application.

The foregoing is merely an optional implementation manner of some of the implementation scenarios of the present application, and it should be noted that, for those skilled in the art, other similar implementation manners according to the technical idea of the present application may be adopted without departing from the technical idea of the solution of the present application, which is also within the protection scope of the embodiments of the present application.

Claims

1. An air conditioner load prediction method based on big data analysis, which is characterized by comprising the following steps:

2. The method for predicting air conditioner load based on big data analysis according to claim 1, wherein the generating a load prediction thermodynamic diagram of the sample learning data by using the sample learning data as model loading data of the air conditioner load prediction model comprises:

3. The method for predicting air conditioner load based on big data analysis according to claim 2, wherein the performing feature fusion calibration on the energy consumption punctual description of the expansion template and the sample learning data to generate an optimized energy consumption punctual description of the optimized expansion template and the sample learning data includes:

4. The method for predicting air conditioner load based on big data analysis according to claim 2, wherein the performing feature fusion calibration on the energy consumption punctual description of the expansion template and the sample learning data to generate an optimized energy consumption punctual description of the optimized expansion template and the sample learning data includes:

5. The big data analysis-based air conditioner load prediction method according to claim 4, wherein the feature fusion calibration optimization is performed on the energy consumption point state description of the expansion template and the sample learning data based on the preset feature fusion calibration parameters, and the generation of the optimized energy consumption point state description of the expansion template and the sample learning data comprises the following steps:

6. The big data analysis-based air conditioner load prediction method according to claim 1, wherein the performing model training on the air conditioner load prediction model based on the load prediction thermodynamic diagram of each sample learning data and the air conditioner load labeling data of each sample learning data in the multi-sample learning data sequence until the model convergence requirement is met, comprises:

7. The big data analysis-based air conditioner load prediction method according to claim 1, wherein the multiple sample learning data sequence includes extended sample learning data and energy consumption reference sample learning data, the air conditioner load prediction model is model-trained based on load prediction thermodynamic diagrams of each sample learning data in the multiple sample learning data sequence and air conditioner load labeling data of each sample learning data until model convergence requirements are met, comprising:

8. The big data analysis-based air conditioner load prediction method according to claim 7, wherein the calculating a blending training cost parameter based on each of the expansion sample learning data in the multi-sample learning data sequence, the optimized expansion template, and the energy consumption reference template obtained by initializing parameter learning, comprises:

9. The big data analysis based air conditioner load prediction method according to any one of claims 1 to 8, further comprising:

Receiving candidate air conditioner energy consumption big data;

10. An air conditioning load prediction system based on big data analysis, characterized by comprising a processor and a computer readable storage medium storing machine executable instructions which when executed by the processor implement the big data analysis based air conditioning load prediction method of any of claims 1-9.