CN116235188A

CN116235188A - Method, apparatus and storage medium for generating predictive model of analysis object

Info

Publication number: CN116235188A
Application number: CN202080105837.6A
Authority: CN
Inventors: 曲颖; 白新; 丹尼尔·施尼盖斯; 管金艳; 王焦剑; 刘晓南
Original assignee: Siemens Ltd China
Current assignee: Siemens Ltd China
Priority date: 2020-11-30
Filing date: 2020-11-30
Publication date: 2023-06-06
Also published as: WO2022110213A1

Abstract

A method, apparatus and storage medium for generating a predictive model of an analysis object. The method (100) comprises: acquiring (101) first data (201) of an analysis object; acquiring (102) second data (204) of the analysis object, wherein the second data (204) and the first data (201) represent the same physical quantity of the analysis object, and the data type of the second data (204) is different from the data type of the first data (201); training (103) a neural network model (202) comprising N hidden layers as a predictive model (203) of an analysis object using the first data (201), wherein N is a positive integer of at least 2; at least one hidden layer of the N hidden layers contained in the predictive model (203) is updated (104) with the second data (204). The method adopts a plurality of data training models, can reduce the dependence of the models on single data, improves the model performance, realizes the transfer learning and accelerates the model training process.

Description

Method, apparatus and storage medium for generating predictive model of analysis object

Technical Field

The present invention relates to the field of artificial intelligence (Artificial Intelligence, AI) technology, and more particularly, to a method, apparatus, and storage medium for generating a predictive model of an analysis object.

Background

Machine Learning (ML) is a method of implementing AI. Machine learning has close relation with fields such as pattern recognition, calculation statistics, artificial intelligence and the like. Machine learning can use machines (computers and software) to mine meaning from known data, thereby giving the machine learning environment the ability. Machine learning algorithms may include supervised learning (e.g., classification problems), unsupervised learning (e.g., clustering problems), semi-supervised learning, ensemble learning, deep learning, and reinforcement learning, among others.

Predictive analysis is a widespread use of machine learning in business problems. The process of generating the predictive model using the training data is the model training process. The course of the output when the input changes can be predicted using a predictive model.

The accuracy of the predictive model is often severely dependent on data availability. Currently, a single training data (e.g., historical data) is typically used to train the predictive model. However, the single training data is usually insufficient in data amount, so that it is difficult to train a predictive model with good performance.

Disclosure of Invention

The embodiment of the invention provides a method, a device and a storage medium for generating a prediction model of an analysis object.

In a first aspect, a method of generating a predictive model of an analysis object, comprises:

Acquiring first data of an analysis object;

acquiring second data of the analysis object, wherein the second data and the first data represent the same physical quantity of the analysis object, and the data type of the second data is different from the data type of the first data;

training a neural network model comprising N hidden layers into a predictive model of the analysis object using the first data, wherein N is a positive integer of at least 2;

updating at least one hidden layer of the N hidden layers included in the prediction model with the second data.

In a second aspect, there is provided an apparatus for generating a predictive model of an analysis object, comprising:

the first data acquisition module is used for acquiring first data of an analysis object;

a second data acquisition module, configured to acquire second data of the analysis object, where the second data and the first data represent the same physical quantity of the analysis object, and a data type of the second data is different from a data type of the first data;

the training module is used for training a neural network model containing N hidden layers into a prediction model of the analysis object by utilizing the first data, wherein N is a positive integer which is at least 2;

An updating module for updating at least one hidden layer of the N hidden layers included in the prediction model with the second data.

In a third aspect, an apparatus for generating a predictive model of an analysis object is provided, comprising a processor and a memory;

the memory has stored therein an application executable by the processor for causing the processor to perform the method of generating a predictive model of an analysis object as claimed in any one of the preceding claims.

In a fourth aspect, a computer readable storage medium is provided, in which computer readable instructions are stored for performing the method of generating a predictive model of an analysis object as claimed in any one of the preceding claims.

Therefore, the embodiment of the invention can generate the training model of the analysis object by adopting a plurality of data of a plurality of data types, and can reduce the dependence of the model on single type data. Moreover, as the data characteristics of the second data are the same as the physical quantity of the analysis object represented by the first data, partial hidden layers in the prediction model can be updated by using the second data instead of retraining the whole prediction model, so that the model training process is accelerated, and the migration learning of the model is realized.

For any of the above aspects, preferably, the data type of the second data is different from the data type of the first data, including:

the first data are actual data in a first time interval, and the second data are simulation data in a second time interval; or (b)

The first data is actual data in a third time interval, and the second data is actual data in a fourth time interval; or (b)

The first data are simulation data in a fifth time interval, and the second data are actual data in a sixth time interval; or (b)

The first data are simulation data in a seventh time interval, and the second data are simulation data in an eighth time interval; or (b)

The first data is combined data comprising simulation data in a ninth time interval and actual data in a tenth time interval, and the second data is actual data in an eleventh time interval; or (b)

The first data is combined data comprising simulation data in a twelfth time interval and actual data in a thirteenth time interval, and the second data is simulation data in a fourteenth time interval; or (b)

The first data is actual data in a fifteenth time interval, and the second data is combined data comprising simulation data in a sixteenth time interval and actual data in a seventeenth time interval; or (b)

The first data is simulation data in an eighteenth time interval, and the second data is combined data comprising simulation data in a nineteenth time interval and actual data in a twentieth time interval.

Preferably, the actual data and the simulation data in the combined data have data indexes that are superimposable on a time attribute or data indexes that are not superimposable on a time attribute.

Therefore, the first data and the second data in the embodiment of the invention have various types, enrich training data and improve model accuracy.

For any of the above aspects, preferably, the method further comprises:

establishing a simulation model of the analysis object based on predetermined analysis object metadata;

the simulation data is generated based on the simulation model.

Therefore, simulation data can be quickly acquired through the simulation model, and the data acquisition efficiency is improved.

For any of the above aspects, preferably, the analysis object is a Heating Ventilation and Air Conditioning (HVAC) system, and the predictive model is a power consumption predictive model; further comprises: receiving a predicted time; and generating a power consumption predicted value corresponding to the predicted time based on the updated prediction model.

Thus, the updated predictive model may be applied to the HVAC system.

For any of the above aspects, preferably, said updating at least one hidden layer of the N hidden layers included in the prediction model with the second data comprises:

and training the prediction model by using the second data, wherein a preset M hidden layers in the prediction model are fixed, and updating the rest hidden layers except the M hidden layers in the prediction model, wherein M is a positive integer which is at least 2, and M is less than or equal to N.

Therefore, the M hidden layers in the prediction model are fixed, and can be reserved as mature knowledge, so that knowledge migration is realized, and the training workload after migration is reduced.

Drawings

Fig. 1 is an exemplary flowchart of a method of generating a predictive model of an analysis object according to an embodiment of the invention.

Fig. 2 is a first exemplary flowchart of a method of generating a predictive model of an analysis object according to an embodiment of the invention.

Fig. 3 is a second exemplary flowchart of a method of generating a predictive model of an analysis object according to an embodiment of the invention.

Fig. 4 is a third exemplary flowchart of a method of generating a predictive model of an analysis object according to an embodiment of the invention.

Fig. 5 is a fourth exemplary flowchart of a method of generating a predictive model of an analysis object according to an embodiment of the invention.

Fig. 6 is a fifth exemplary flowchart of a method of generating a predictive model of an analysis object according to an embodiment of the invention.

Fig. 7 is a sixth exemplary flowchart of a method of generating a predictive model of an analysis object according to an embodiment of the invention.

Fig. 8 is a seventh exemplary flowchart of a method of generating a predictive model of an analysis object according to an embodiment of the invention.

Fig. 9 is an eighth exemplary flowchart of a method of generating a predictive model of an analysis object according to an embodiment of the invention.

FIG. 10 is a flow chart of a method of predicting power usage of an HVAC system according to an embodiment of the present invention.

Fig. 11 is a block diagram of an apparatus for predicting power consumption of an HVAC system according to an embodiment of the present invention.

FIG. 12 is a schematic diagram of the prediction of HVAC power usage according to an embodiment of the present invention.

Fig. 13 is a block diagram of an HVAC electricity consumption prediction apparatus according to an embodiment of the present invention.

Fig. 14 is an exemplary configuration diagram of an apparatus for generating a prediction model of an analysis object according to an embodiment of the present invention.

Fig. 15 is an exemplary block diagram of an apparatus for generating a predictive model of an analysis object having a memory-processor architecture according to an embodiment of the present invention.

Wherein, the reference numerals are as follows:

reference numerals	Meaning of
100	Method for generating predictive model of analysis object
201,301,401,501,601,701,801,901	First data
202,302,402,502,602,702,802,902	Neural network model
203,303,403,503,603,703,803,903	Predictive model
204,304,404,504,604,704,804,904	Second data
205,305,405,505,605,705,805,905	Updated predictive model
1000	Method for predicting power consumption of HVAC system
1001～1002	Step (a)
80	Power consumption prediction device of HVAC system
81	Receiving module
82	Prediction module
61	Actual power consumption in winter
62	Simulated electricity consumption in summer
63	Practical power consumption in summer
64	The predicted electricity consumption of the embodiment of the invention
65	Predictive power usage for standard methods
30	Prediction device for HVAC (heating, ventilation and air conditioning) power consumption
31	Interface
32	Memory device

33	Processor and method for controlling the same
34	Bus line
50	Device for analyzing predictive model of object
51	First data acquisition module
52	Second data acquisition module
53	Training module
54	Update module
55	Simulation data acquisition module
56	Receiving module
57	Prediction module
70	Device for analyzing predictive model of object
71	Processor and method for controlling the same
72	Memory device

Detailed Description

In order to make the technical scheme and advantages of the present invention more apparent, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the detailed description is intended by way of illustration only and is not intended to limit the scope of the invention.

For simplicity and clarity of description, the following description sets forth aspects of the invention by describing several exemplary embodiments. Numerous details in the embodiments are merely configured to provide an understanding of aspects of the invention. It will be apparent, however, that the embodiments of the invention may be practiced without limitation to these specific details. Some embodiments are not described in detail in order to avoid unnecessarily obscuring aspects of the present invention, but rather only to present a framework. Hereinafter, "comprising" means "including but not limited to", "according to … …" means "according to at least … …, but not limited to only … …". The term "a" or "an" is used herein to refer to a number of components, either one or more, or at least one, unless otherwise specified.

In consideration of the defect that a single training data is adopted to train a prediction model, the embodiment of the invention trains the prediction model comprising hidden layers by using the first data, and then updates part of hidden layers in the prediction model by using the second data, so that training data can be enriched and model performance can be improved. Moreover, the problem of mobility of the prediction model can be solved, and the model training process is quickened.

As shown in fig. 1, the method includes:

step 101: first data of an analysis object is acquired.

The analysis object is here an object that needs to be analyzed, which may be implemented as a thermodynamic system or an electrical system, etc. Preferably, the analysis object is implemented as a Heating Ventilation and Air Conditioning (HVAC) system.

The data type of the first data may be determined by the data acquisition mode or the data source of the first data. For example, the data types of the first data may include:

type (1): analyzing actual data (e.g., historical data) of the object;

type (2): analyzing simulation data of the object;

type (3): the combined data of the object, including the actual data and the simulation data, is analyzed.

The first data characterizes a predetermined physical quantity of the analysis object. For example, the first data may be used to characterize the power usage of the analysis object, the fan usage time, the refrigerant usage, and so on.

Step 102: and acquiring second data of the analysis object, wherein the second data and the first data represent the same physical quantity of the analysis object, and the data type of the second data is different from that of the first data.

Similarly, the data type of the second data may be determined by the data acquisition mode or the data source of the second data. For example, the data types of the second data may include:

type (1): analyzing actual data (e.g., historical data) of the object;

type (2): analyzing simulation data of the object;

The data type of the second data is different from the data type of the first data. Thus, when the data type of the first data is type (1), then the data type of the second data is type (2) or type (3); when the data type of the first data is the type (2), the data type of the second data is the type (1) or the type (3); when the data type of the first data is type (3), then the data type of the second data is type (1) or type (2).

The second data characterizes the same physical quantity of the analysis object as the first data. For example, when the first data characterizes the electricity consumption of the analysis object, the second data likewise characterizes the electricity consumption of the analysis object; when the first data characterizes a fan usage time of the analysis object, the second data likewise characterizes a fan usage time of the analysis object, and so on.

Preferably, the second data is more relevant to the prediction task of the resulting prediction model than the first data.

In one embodiment, the first data is actual data in a first time interval and the second data is simulation data in a second time interval. Both the first time interval and the second time interval may be described by a time of year, quarter, month, week, or day, etc. unit of measure. For example, the first data is real electricity usage data for plant a containing an analysis object (such as HVAC system) at 3 months 2012; the second data is simulated electricity usage data for plant a containing the analysis object (e.g., HVAC system) at 4 months 2012 or simulated electricity usage data for plant B containing the analysis object (e.g., HVAC system) at 4 months 2012.

In one embodiment, the first data is actual data in a third time interval, and the second data is actual data in a fourth time interval. Both the third time interval and the fourth time interval may be described by a time of year, quarter, month, week, or day unit of measure. For example, the first data is real electricity usage data for plant a containing an analysis object (such as HVAC system) at 3 months 2012; the second data is real electricity usage data for plant a containing an analysis object (such as HVAC system) at 4 months 2012, or real electricity usage data for plant B containing an analysis object (such as HVAC system) at 4 months 2012.

In one embodiment, the first data is simulation data in a fifth time interval, and the second data is actual data in a sixth time interval. Both the fifth time interval and the sixth time interval may be described by a unit of time measurement such as year, quarter, month, week, or day. For example, the first data is simulated electricity usage data for plant a containing an analysis object (such as an HVAC system) at 3 months 2012; the second data is real electricity usage data for plant a containing an analysis object (such as HVAC system) at 4 months 2012, or real electricity usage data for plant B containing an analysis object (such as HVAC system) at 4 months 2012.

In one embodiment, the first data is simulation data in a seventh time interval, and the second data is simulation data in an eighth time interval. Both the seventh time interval and the eighth time interval may be described by a time of year, quarter, month, week, or day unit of measure. For example, the first data is simulated electricity usage data for plant a containing an analysis object (such as an HVAC system) at 3 months 2012; the second data is simulated electricity usage data for plant a containing an analysis object (such as an HVAC system) at 4 months 2012, or simulated electricity usage data for plant B containing an analysis object (such as an HVAC system) at 4 months 2012.

In one embodiment, the first data is combined data including simulation data in a ninth time interval and actual data in a tenth time interval, and the second data is actual data in an eleventh time interval. The ninth time interval, tenth time interval, and eleventh time interval may be described by units of time of year, quarter, month, week, or day, etc. For example, the first data is a combination of simulated electricity usage data for plant a at 4 months 2012 and real electricity usage data for 3 months 2012 containing an analysis object (such as an HVAC system); the second data is real electricity usage data for plant a containing an analysis object (such as an HVAC system) at 2 months in 2012, or real electricity usage data for plant B containing an analysis object (such as an HVAC system) at 2 months in 2012.

In one embodiment, the first data is a combination data including simulation data in a twelfth time interval and actual data in a thirteenth time interval, and the second data is simulation data in a fourteenth time interval. The twelfth time interval, thirteenth time interval, and fourteenth time interval may be described by a unit of time such as a year, quarter, month, week, or day. For example, the first data is a combination of simulated electricity usage data for plant a of an analysis object (such as an HVAC system) at 4 months 2012 and real electricity usage data for 3 months 2012; the second data is simulated electricity usage data for plant a containing an analysis object (such as an HVAC system) at 5 months in 2012, or simulated electricity usage data for plant B containing an analysis object (such as an HVAC system) at 5 months in 2012.

In one embodiment, the first data is actual data in a fifteenth time interval, and the second data is combined data including simulation data in a sixteenth time interval and actual data in a seventeenth time interval. The fifteenth time interval, the sixteenth time interval, and the seventeenth time interval may be described by units of time of year, quarter, month, week, or day, etc. For example, the first data is real electricity usage data of plant a in 5 months 2012 including an analysis object (such as an HVAC system), the second data is a combination of simulated electricity usage data of plant a in 4 months 2012 and real electricity usage data of 3 months 2012 including an analysis object (such as an HVAC system), or a combination of simulated electricity usage data of plant B in 4 months 2012 and real electricity usage data of 3 months 2012 including an analysis object (such as an HVAC system).

In one embodiment, the first data is simulation data in an eighteenth time interval, and the second data is combined data including simulation data in a nineteenth time interval and actual data in a twentieth time interval. The eighteenth time interval, the nineteenth time interval, and the twentieth time interval may all be described by units of time of year, quarter, month, week, or day, etc. For example, the first data is simulated electricity usage data for plant a containing an analysis object (such as an HVAC system) at 5 months 2012; the second data is a combination of simulated electricity usage data of plant a of an analysis object (such as HVAC system) at 4 months 2012 and real electricity usage data of 3 months 2012, or a combination of simulated electricity usage data of plant B of an analysis object (such as HVAC system) at 4 months 2012 and real electricity usage data of 3 months 2012.

In one embodiment, the actual data and the simulated data in the combined data have data metrics that are superimposable on a temporal attribute.

For example, the actual data and the simulation data in the combined data are respectively implemented as the electricity consumption amount data on the stackable month. For example, the actual data includes the electricity consumption data of 1 month in 2011 and the electricity consumption data of 2 months in 2011, the simulation data is the electricity consumption data of 3 months in 2011, and the combined data is the electricity consumption data of the first quarter (including 1 month to 3 months) in 2011.

In one embodiment, the actual data and the simulated data in the combined data have data metrics that are non-superimposable on a temporal attribute.

For example, the actual data is electricity consumption data of 1 month in 2011, and the simulation data is a room temperature value of 1 month in 2011. For another example, the actual data is electricity consumption data of 1 month in 2011, and the simulation data is fan use time of 2 months in 2011.

While the foregoing exemplary descriptions of typical examples of first data and second data, those skilled in the art will recognize that such descriptions are merely exemplary and are not intended to limit the scope of embodiments of the present invention.

Simulation data may be obtained using a simulation model of the analysis object. The method specifically comprises the following steps: establishing a simulation model of the analysis object based on predetermined analysis object metadata (such as a building information model, a design drawing and the like); simulation data is generated based on the simulation model.

Step 103: a neural network model comprising N hidden layers is trained as a predictive model of an analysis object using first data, where N is a positive integer of at least 2.

Specifically, the neural network model may be implemented as: a feed forward neural network model, a radial basis neural network model, a Long Short Term Memory (LSTM) network model, an Echo State Network (ESN), a gate loop unit (GRU) network model, or a deep residual network model, etc. Preferably, the neural network model is implemented as an LSTM network model.

Here, a neural network model including N hidden layers is trained as a predictive model of an analysis object using the first data. Based on the training step of step 103, parameters (e.g., weights) for each hidden layer may be determined.

Step 104: at least one hidden layer among the N hidden layers included in the prediction model is updated with the second data.

The prediction model of the analysis object trained by the first data in step 103 may be used to perform prediction analysis on the analysis object applied to each scene.

In one embodiment, when the application scenario of the second data is the same as the application scenario of the first data, the second data is preferably used to update a predetermined part of hidden layers in the prediction model of the analysis object trained by the first data, so as to improve the accuracy of the prediction model. Here, it may be determined whether a specific hidden layer is to be an hidden layer that needs to be updated based on a comparison of the prediction effect of the updated and the non-updated specific hidden layer. For example, when the prediction effect of a specific hidden layer is significantly improved (for example, the improvement amplitude of the prediction accuracy is greater than a predetermined threshold value) when the specific hidden layer is updated or not updated, the specific hidden layer is taken as the hidden layer to be updated; when the prediction effect of a specific hidden layer is not significantly improved (for example, the improvement amplitude of the prediction accuracy is less than or equal to a predetermined threshold value) when the prediction effect of the specific hidden layer is updated or not updated, the specific hidden layer is not used as the hidden layer which needs to be updated.

In fact, since the first data generally originates from a specific scene, and the analysis object may be applied to another application scene related to the specific scene, it is preferable to update an hidden layer having a relevance to the application scene of the second data in the prediction model of the analysis object trained by the first data by using the second data originating from the another application scene, so that the updated prediction model is more suitable for the another application scene than the prediction model before the update, thereby implementing knowledge migration. Here, the determining manner of the hidden layer having the association with the application scenario of the second data includes:

(1) And deriving an implicit layer having relevance to the application scene of the second data theoretically based on model analysis of the application scene of the second data.

(2) And determining whether the specific hidden layer has relevance with the application scene of the second data based on the comparison of the prediction effect of the updated specific hidden layer and the prediction effect of the non-updated specific hidden layer. For example, when the prediction effect of a specific hidden layer is significantly improved (for example, the improvement amplitude of the prediction accuracy is greater than a predetermined threshold value) when the specific hidden layer is updated or not updated, the specific hidden layer has relevance to the application scene of the second data, so that the specific hidden layer is used as the hidden layer to be updated; when the prediction effect of a specific hidden layer is not significantly improved (for example, the improvement amplitude of the prediction accuracy is smaller than or equal to a predetermined threshold value) when the prediction effect of the specific hidden layer is updated or not updated, the specific hidden layer has no relevance to the application scene of the second data, and therefore the specific hidden layer is not used as the hidden layer which needs to be updated.

In one embodiment, updating at least one hidden layer of the N hidden layers included in the prediction model with the second data includes: and training a prediction model by using the second data, wherein M hidden layers in the prediction model are fixed, and the rest hidden layers except the M hidden layers in the prediction model are updated, wherein M is a positive integer which is at least 2, and M is less than or equal to N.

A typical process of generating a predictive model of an analysis object is described below.

In one embodiment of fig. 2, the first data 201 of the analysis object is: actual data of the analysis object in a first time interval, which is collected in a first application scene containing the analysis object; the second data 204 of the analysis object is: simulation data of the analysis object in a second time interval, which are acquired in a first application scene containing the analysis object. The first time interval and the second time interval may be the same or different, and preferably are different.

For example, the predictive task is a prediction of power usage in a first application scenario (such as plant a) that includes an analysis object (such as an HVAC system); the first data 201 is the actual usage time of the fan in 2011 month 1 in factory a; the second data 204 is the fan simulation usage time in plant a, 2011, 2 months.

First, the first data 201 is input as training data into a neural network model 202 including N hidden layers, and the neural network model 202 can be trained as a prediction model 203 of an analysis object through a training process. In the predictive model 203, parameters of N hidden layers are all determined.

Then, the second data 204 is input as training data into the predictive model 203 to perform training again. During this retraining process: the parameters of the remaining (N-M) hidden layers are updated with this retraining process, keeping the predetermined M hidden layers of the N hidden layers of the predictive model 203 fixed (i.e. the parameters of the M hidden layers are not updated). After training is performed again, an updated predictive model 205 may be obtained. At this time, the prediction of the amount of electricity used by the analysis object in the plant a may be performed using the updated prediction model 205.

In another embodiment of fig. 2, the first data 201 of the analysis object is: actual data of the analysis object in a first time interval, which is collected in a first application scene containing the analysis object; the second data 204 of the analysis object is: simulation data of the analysis object in a second time interval, which are acquired in a second application scenario containing the analysis object. The first time interval and the second time interval may be the same or different, and preferably are different. The second data 204 is more relevant to the predicted task in the second application scenario than the first data 201.

For example, the predictive task is a prediction of power usage in a second application scenario (e.g., plant B) that includes an analysis object (e.g., HVAC system); the first data 201 is the fan usage time in factory a, 2011, 2 months; the second data 204 is the fan simulation usage time in factory B, 1 month 2011.

Then, the second data 204 is input as training data into the predictive model 203 to perform training again. During this retraining process: the predetermined M hidden layers of the N hidden layers of the predictive model 203 that are not related to the application scenario of the second data 204 are kept fixed (i.e. the parameters of the M hidden layers are not updated), while the remaining (N-M) hidden layer parameters related to the application scenario of the second data 204 are updated with the retraining process. After training is performed again, an updated predictive model 205 may be obtained. At this time, the prediction of the amount of electricity consumption of the analysis object in the plant B may be performed using the updated prediction model 205.

In one embodiment of fig. 3, the first data 301 of the analysis object is: the method comprises the steps of acquiring actual data of an analysis object in a third time interval in a first application scene containing the analysis object; the second data 304 of the analysis object is: the actual data of the analysis object in the fourth time interval, which is acquired in the first application scene containing the analysis object. Wherein the third time interval and the fourth time interval may be the same or different.

For example, the predictive task is a prediction of power usage in a first application scenario (such as plant a) that includes an analysis object (such as an HVAC system); the first data 301 is actual data of electricity consumption in 2011 in 2 months in the factory a; the second data 304 is actual electricity consumption data of 2011 month in factory a.

First, the first data 301 is input as training data into a neural network model 302 including N hidden layers, and the neural network model 302 can be trained as a prediction model 303 of an analysis object through a training process. In this predictive model 303, parameters of all N hidden layers are determined.

Then, the second data 304 is input as training data into the predictive model 303 to perform training again. During this retraining process: the parameters of the remaining (N-M) hidden layers are updated with this retraining process, keeping the predetermined M hidden layers of the N hidden layers of the predictive model 303 fixed (i.e. the parameters of the M hidden layers are not updated). After training is performed again, an updated predictive model 305 may be obtained. At this time, the prediction of the amount of electricity used by the analysis object in the plant a may be performed using the updated prediction model 305.

In another embodiment of fig. 3, the first data 301 of the analysis object is: the method comprises the steps of acquiring actual data of an analysis object in a third time interval in a first application scene containing the analysis object; the second data 304 of the analysis object is: the actual data of the analysis object in the fourth time interval, which is acquired in the second application scene containing the analysis object. Wherein the third time interval and the fourth time interval may be the same or different. The second data 304 is more relevant to the predicted task in the second application scenario than the first data 301.

For example, the predictive task is a prediction of power usage in a second application scenario (e.g., plant B) that includes an analysis object (e.g., HVAC system); the first data 301 is actual data of electricity consumption in 2011 in 2 months in the factory a; the second data 304 is actual electricity consumption data of 2011 month in factory B.

Then, the second data 304 is input as training data into the predictive model 303 to perform training again. During this retraining process: the predetermined M hidden layers of the N hidden layers of the predictive model 303 that are not related to the application scenario of the second data 304 are kept fixed (i.e. the parameters of the M hidden layers are not updated), while the parameters of the remaining (N-M) hidden layers that are related to the application scenario of the second data 304 are updated with the retraining process. After training is performed again, an updated predictive model 305 may be obtained. At this time, the prediction of the amount of electricity consumption of the analysis object in the plant B may be performed using the updated prediction model 305.

In fig. 4, first data 401 of an analysis object is: simulation data of the analysis object in a fifth time interval, which are collected in a first application scene containing the analysis object; the second data 404 of the analysis object is: the actual data of the analysis object in the sixth time interval, which is acquired in the first application scenario containing the analysis object. The fifth time interval and the sixth time interval may be the same or different, and preferably are different.

For example, the predictive task is a prediction of power usage in a first application scenario (such as plant a) that includes an analysis object (such as an HVAC system); the first data 401 is the fan simulation usage time of 2011, 2 months in factory a; the second data 404 is the actual fan usage time in factory a, 1 month 2011.

First, the first data 401 is input as training data into a neural network model 402 including N hidden layers, and the neural network model 402 can be trained as a prediction model 403 of an analysis object through a training process. In this predictive model 403, parameters for all N hidden layers are determined.

Then, the second data 404 is input as training data into the predictive model 403 to perform training again. During this retraining process: the parameters of the remaining (N-M) hidden layers are updated with this retraining process, keeping the predetermined M hidden layers of the N hidden layers of the predictive model 403 fixed (i.e. the parameters of the M hidden layers are not updated). After training is performed again, an updated predictive model 405 may be obtained. At this time, the prediction of the amount of electricity consumption of the analysis object in the plant a may be performed using the updated prediction model 405.

In another embodiment of fig. 4, the first data 401 of the analysis object is: simulation data of the analysis object in a fifth time interval, which are collected in a first application scene containing the analysis object; the second data 404 of the analysis object is: the actual data of the analysis object in the sixth time interval, which is acquired in the second application scenario containing the analysis object. Wherein the fifth time interval and the sixth time interval may be the same or different. The second data 404 is more relevant to the predicted task in the second application scenario than the first data 401.

For example, the predictive task is a prediction of power usage in a second application scenario (e.g., plant B) that includes an analysis object (e.g., HVAC system); the first data 401 is the fan simulation usage time of 2011, 2 months in factory a; the second data 404 is the actual fan usage time of 2011 month in factory B.

Then, the second data 404 is input as training data into the predictive model 403 to perform training again. During this retraining process: the M hidden layers of the N hidden layers of the predictive model 403 that are not related to the application scenario of the second data 404 are kept fixed (i.e. the parameters of the M hidden layers are not updated), whereas the remaining (N-M) hidden layer parameters that are related to the application scenario of the second data 404 are updated with the retraining process. After training is performed again, an updated predictive model 405 may be obtained. At this time, the prediction of the amount of electricity used by the analysis object in the plant B may be performed using the updated prediction model 405.

In fig. 5, first data 501 of an analysis object is: simulation data of the analysis object in a seventh time interval, which are collected in a first application scene containing the analysis object; the second data 504 of the analysis object is: simulation data of the analysis object in an eighth time interval, which are acquired in a first application scenario containing the analysis object. Wherein the seventh time interval and the eighth time interval may be the same or different.

For example, the predictive task is a prediction of power usage in a first application scenario (such as plant a) that includes an analysis object (such as an HVAC system); the first data 501 is electricity consumption simulation data of 2011 in 2 months in the factory a; the second data 404 is electricity consumption simulation data of 2011 month 1 in the factory a.

First, the first data 501 is input as training data into a neural network model 502 including N hidden layers, and the neural network model 502 can be trained as a prediction model 503 of an analysis object through a training process. In this predictive model 503, parameters for all N hidden layers are determined.

Then, the second data 504 is input as training data into the predictive model 503 to perform training again. During this retraining process: the parameters of the remaining (N-M) hidden layers are updated with this retraining process, keeping the predetermined M hidden layers of the N hidden layers of the predictive model 503 fixed (i.e., the parameters of the M hidden layers are not updated). After training is performed again, an updated predictive model 505 may be obtained. At this time, the prediction of the amount of electricity used by the analysis object in the plant a may be performed using the updated prediction model 505.

In another embodiment of fig. 5, the first data 501 of the analysis object is: simulation data of the analysis object in a seventh time interval, which are collected in a first application scene containing the analysis object; the second data 504 of the analysis object is: simulation data of the analysis object in an eighth time interval acquired in a second application scenario containing the analysis object. Wherein the seventh time interval and the eighth time interval may be the same or different. The second data 504 is more relevant to the predicted task in the second application scenario than the first data 501.

For example, the predictive task is a prediction of power usage in a second application scenario (e.g., plant B) that includes an analysis object (e.g., HVAC system); the first data 501 is electricity consumption simulation data of 2011 in 2 months in the factory a; the second data 504 is electricity consumption simulation data of 2011 month 1 in factory B.

Then, the second data 504 is input as training data into the predictive model 503 to perform training again. During this retraining process: the predetermined M hidden layers of the N hidden layers of the predictive model 503 that are not related to the application environment of the second data 504 are kept fixed (i.e. the parameters of the M hidden layers are not updated), while the remaining (N-M) hidden layer parameters that are related to the application environment of the second data 504 are updated with the retraining process. After training is performed again, an updated predictive model 505 may be obtained. At this time, the prediction of the amount of electricity used by the analysis object in the plant B may be performed using the updated prediction model 505.

In fig. 6, first data 601 of an analysis object is: the method comprises the steps of collecting combined data comprising simulation data in a ninth time interval and actual data in a tenth time interval in a first application scene comprising an analysis object; the second data 604 of the analysis object is: the actual data of the analysis object in the eleventh time interval acquired in the first application scenario containing the analysis object. Wherein the ninth time interval and the tenth time interval may be the same or different. The second data 604 is more relevant to the predicted task in the first application scenario than the first data 601.

For example, the predictive task is a prediction of power usage in a first application scenario (such as plant a) that includes an analysis object (such as an HVAC system); the first data 601 is a combination data including a simulated fan usage time of 2011 month 2 in the factory a and an actual temperature value of 2011 month 3 in the factory a; the second data 604 is actual usage time of the fan in 2011 month in the factory a and actual power consumption data in 2011 month in the factory a.

First, the first data 601 is input as training data into a neural network model 602 including N hidden layers, and the neural network model 602 can be trained as a prediction model 603 of an analysis object through a training process. In this predictive model 603, parameters for all N hidden layers are determined.

Then, the second data 604 is input as training data into the predictive model 603 to perform training again. During this retraining process: the parameters of the remaining (N-M) hidden layers are updated with this retraining process, keeping the predetermined M hidden layers of the N hidden layers of the predictive model 603 fixed (i.e. the parameters of the M hidden layers are not updated). After training is performed again, an updated predictive model 605 may be obtained. At this time, the prediction of the amount of electricity consumption of the analysis object in the plant a may be performed using the updated prediction model 605.

In another embodiment of fig. 6, the first data 601 of the analysis object is: the method comprises the steps of collecting combined data comprising simulation data in a ninth time interval and actual data in a tenth time interval in a first application scene comprising an analysis object; the second data 604 of the analysis object is: the actual data of the analysis object in the eleventh time interval acquired in the second application scenario containing the analysis object. Wherein the ninth time interval and the tenth time interval may be the same or different. The second data 604 is more relevant to the predicted task in the second application scenario than the first data 601.

For example, the predictive task is a prediction of power usage in a second application scenario (e.g., plant B) that includes an analysis object (e.g., HVAC system); the first data 601 is a combination data including a simulated fan usage time of 2011 month 2 in the factory a and an actual temperature value of 2011 month 3 in the factory a; the second data 604 is actual usage time of the fan in 2011 month in the factory B and actual electricity consumption data in 2011 month in the factory a.

Then, the second data 604 is input as training data into the predictive model 603 to perform training again. During this retraining process: the predetermined M hidden layers of the N hidden layers of the predictive model 603 that are not related to the application environment of the second data 604 are kept fixed (i.e. the parameters of the M hidden layers are not updated), while the remaining (N-M) hidden layer parameters that are related to the application environment of the second data 604 are updated with the retraining process. After training is performed again, an updated predictive model 605 may be obtained. At this time, the prediction of the amount of electricity consumption of the analysis object in the plant B may be performed using the updated prediction model 605.

In fig. 7, the first data 701 of the analysis object is: the method comprises the steps of collecting combined data comprising simulation data in a twelfth time interval and actual data in a thirteenth time interval in a first application scene comprising an analysis object; the second data 704 of the analysis object is: simulation data within a fourteenth time interval acquired in a first application scenario containing an analysis object. Wherein the twelfth time interval and the thirteenth time interval may be the same or different, preferably different; the fourteenth time interval may be the same or different, preferably different, from the twelfth time interval. The second data 704 is more relevant to the predicted task in the first application scenario than the first data 701.

For example, the predictive task is a prediction of power usage in a first application scenario (such as plant a) that includes an analysis object (such as an HVAC system); the first data 701 is a combination data including a simulated fan usage time of 2011 month 2 in the factory a and an actual power consumption amount of 2011 month 3 in the factory a; the second data 704 is simulation data of the simulation fan usage time of 2011 month 1 in the factory a and the electricity consumption amount of 2011 month 1.

First, the first data 701 is input as training data to a neural network model 702 including N hidden layers, and the neural network model 702 can be trained as a prediction model 703 of an analysis object through a training process. In this predictive model 703, parameters for all N hidden layers are determined.

Then, the second data 704 is input as training data into the predictive model 703 to perform training again. During this retraining process: the parameters of the remaining (N-M) hidden layers are updated with this retraining process, keeping the predetermined M hidden layers of the N hidden layers of the predictive model 703 fixed (i.e. the parameters of the M hidden layers are not updated). After training is performed again, an updated predictive model 705 may be obtained. At this time, the prediction of the amount of electricity consumption of the analysis object in the plant a may be performed using the updated prediction model 705.

In another embodiment of fig. 7, the first data 701 of the analysis object is: the method comprises the steps of collecting combined data comprising simulation data in a twelfth time interval and actual data in a thirteenth time interval in a first application scene comprising an analysis object; the second data 704 of the analysis object is: the actual data of the analysis object in the fourteenth time interval acquired in the second application scenario containing the analysis object. Wherein the twelfth time interval and the thirteenth time interval may be the same or different, preferably different; the fourteenth time interval may be the same or different, preferably different, from the twelfth time interval. The second data 704 is more relevant to the predicted task in the second application scenario than the first data 701.

For example, the predictive task is a prediction of power usage in a second application scenario (e.g., plant B) that includes an analysis object (e.g., HVAC system); the first data 701 is a combination data including a simulated fan usage time of 2011 month 2 in the factory a and an actual power consumption amount of 2011 month 3 in the factory a; the second data 704 is simulation data of the simulation fan usage time of 2011 month 1 and the electricity consumption amount of 2011 month 1 in the factory B.

Then, the second data 704 is input as training data into the predictive model 703 to perform training again. During this retraining process: the predetermined M hidden layers of the N hidden layers of the predictive model 703 that are not related to the application environment of the second data 704 are kept fixed (i.e. the parameters of the M hidden layers are not updated), while the remaining (N-M) hidden layer parameters that are related to the application environment of the second data 704 are updated with the retraining process. After training is performed again, an updated predictive model 705 may be obtained. At this time, the prediction of the amount of electricity consumption of the analysis object in the plant B may be performed using the updated prediction model 705.

In fig. 8, first data 801 of an analysis object is: actual data in a fifteenth time interval acquired in a first application scenario containing an analysis object; the second data 804 of the analysis object is: the simulation data in the sixteenth time interval and the actual data in the seventeenth time interval acquired in the first application scenario containing the analysis object. Wherein the sixteenth time interval and the seventeenth time interval may be the same or different, preferably different; the fifteenth time interval and the seventeenth time interval may be the same or different, preferably different. Preferably different.

For example, the predictive task is a prediction of power usage in a first application scenario (such as plant a) that includes an analysis object (such as an HVAC system); the first data 801 includes actual data of electricity consumption in 2011, 2 months, and actual temperature value in 2011, 2 months in the factory a; the second data 804 is a combination data including electricity consumption simulation data of 2011 month in the plant a and an actual temperature value of 2011 month in the plant a.

First, the first data 801 is inputted as training data into a neural network model 802 including N hidden layers, and the neural network model 802 can be trained as a prediction model 803 of an analysis object through a training process. In this predictive model 803, parameters for all N hidden layers are determined.

Then, the second data 804 is input as training data into the prediction model 803 to perform training again. During this retraining process: the parameters of the remaining (N-M) hidden layers are updated with this retraining process, keeping the predetermined M hidden layers of the N hidden layers of the predictive model 803 fixed (i.e., the parameters of the M hidden layers are not updated). After training is performed again, an updated predictive model 805 may be obtained. At this time, the prediction of the amount of electricity used by the analysis object in the plant a may be performed using the updated prediction model 805.

In another embodiment of fig. 8, the first data 801 of the analysis object is: actual data in a fifteenth time interval acquired in a first application scenario containing an analysis object; the second data 704 of the analysis object is: and the simulation data in the sixteenth time interval and the actual data in the seventeenth time interval are acquired in a second application scene containing the analysis object. The sixteenth time interval and the seventeenth time interval may be the same or different, preferably different; the fifteenth time interval and the seventeenth time interval may be the same or different, preferably different. Preferably different. The second data 804 is more relevant to the predicted task in the second application scenario than the first data 801.

For example, the predictive task is a prediction of power usage in a second application scenario (e.g., plant B) that includes an analysis object (e.g., HVAC system); the first data 801 includes actual data of electricity consumption in 2011, 2 months, and actual temperature value in 2011, 2 months in the factory a; the second data 804 is the combination data including the electricity consumption simulation data of 2011 month in the factory B and the actual temperature value of 2011 month in the factory B.

Then, the second data 804 is input as training data into the prediction model 803 to perform training again. During this retraining process: the predetermined M hidden layers of the N hidden layers of the predictive model 803 that are not related to the application environment of the second data 804 are kept fixed (i.e., the parameters of the M hidden layers are not updated), while the remaining (N-M) hidden layer parameters that are related to the application environment of the second data 804 are updated with the retraining process. After training is performed again, an updated predictive model 805 may be obtained. At this time, the prediction of the amount of electricity consumption of the analysis object in the plant B may be performed using the updated prediction model 805.

In fig. 9, first data 901 of an analysis object is: simulation data in an eighteenth time interval acquired in a first application scene containing an analysis object; the second data 904 of the analysis object is: and the combined data which is acquired in the first application scene containing the analysis object and contains the simulation data in the nineteenth time interval and the actual data in the twentieth time interval. Wherein the nineteenth time interval and the twentieth time interval may be the same or different, preferably different; the nineteenth time interval and the eighteenth time interval may be the same or different, preferably different. Preferably different.

For example, the predictive task is a prediction of power usage in a first application scenario (such as plant a) that includes an analysis object (such as an HVAC system); the first data 801 includes simulated fan usage time of 2011 month 2 in factory a and electricity consumption simulation data of 2011 month 2; the second data 804 is a combination data including the simulated fan usage time of 2011 month in the factory a and the electricity consumption actual data of 2011 month in the factory a.

First, the first data 901 is input as training data to a neural network model 902 including N hidden layers, and the neural network model 902 can be trained as a prediction model 903 of an analysis object through a training process. In this predictive model 903, parameters for all N hidden layers are determined.

Then, the second data 904 is input as training data into the predictive model 903 to perform training again. During this retraining process: the parameters of the remaining (N-M) hidden layers are updated with this retraining process, keeping the predetermined M hidden layers of the N hidden layers of the predictive model 903 fixed (i.e. the parameters of the M hidden layers are not updated). After training is performed again, an updated predictive model 905 may be obtained. At this time, the prediction of the amount of electricity used by the analysis object in the plant a may be performed using the updated prediction model 805.

In another embodiment of fig. 9, the first data 901 of the analysis object is: simulation data in an eighteenth time interval acquired in a first application scene containing an analysis object; the second data 904 of the analysis object is: and the combined data which is acquired in the second application scene containing the analysis object and contains the simulation data in the nineteenth time interval and the actual data in the twentieth time interval. Wherein the nineteenth time interval and the twentieth time interval may be the same or different, preferably different; the nineteenth time interval and the eighteenth time interval may be the same or different, preferably different. Preferably different. The second data 904 is more relevant to the predicted task in the second application scenario than the first data 901.

For example, the predictive task is a prediction of power usage in a second application scenario (e.g., plant B) that includes an analysis object (e.g., HVAC system); the first data 801 includes simulated fan usage time of 2011 month 2 in factory a and electricity consumption simulation data of 2011 month 2; the second data 804 is a combination data including the simulated fan usage time of 2011 month in the factory B and the actual power consumption data of 2011 month in the factory B.

In an embodiment of the present invention, a simulation model of an analysis object may be constructed using metadata, and simulation data (data set D) is generated using the simulation model _S ). Next, a process of training a predictive model of the analysis object is performed. In this process, the simulation data D may be used _S Actual data D of analysis object _A Combining to obtain combined data D _C D is _c＝ D _S +D _A And utilize the combined data D _C The neural network model, which is preferably implemented as an LSTM model, is trained to obtain a predictive model of the analysis object. In this process, it is also possible to first use the actual data D of the analysis object _A Training a neural network model, preferably implemented as an LSTM model, to obtain a predictive model M of an analysis object _A . The M is _A Can be used to perform predictions on analysis objects. Further, simulation data D of the analysis object can be utilized _S For predictive model M _A Training is performed. In using simulation data D of analysis object _S For predictive model M _A In the training process, the prediction model M _A The parameters of the predetermined hidden layer of (c) remain fixed while the parameters of other hidden layers than the predetermined hidden layer are updated. Updated prediction model M _A Comparing the pre-update prediction model M _A Has better accuracy.

In the following, exemplary comparisons are made between updating hidden layers and not updating hidden layers in an application environment to which the transfer learning is transferred. For an LSTM network model containing 2 hidden layers, the number of total trainable parameters is 508201 when the hidden layer updating approach is not employed in the application environment. When the application environment adopts the update hidden layer mode, the number of trainable parameters is 201. Therefore, after updating the hidden layer, the number of trainable parameters is significantly reduced, thereby improving the training speed.

In an embodiment of the invention, a method of predicting power usage of an HVAC system is also presented.

FIG. 10 is a flow chart of a method of predicting power usage of an HVAC system according to an embodiment of the present invention. As shown in fig. 10, the method includes:

step 1001: receiving a predicted time;

step 1002: generating a power usage prediction value corresponding to the prediction time based on a power usage prediction model of the HVAC system; the method for generating the electricity consumption prediction model comprises the following steps: acquiring first power data of an HVAC system; acquiring second power usage data for the HVAC system; training a neural network model comprising N hidden layers into the electricity consumption prediction model by utilizing the first electricity consumption data, wherein N is a positive integer which is at least 2; updating at least one hidden layer among the N hidden layers included in the power consumption prediction model with the second power consumption data.

In one embodiment, the first power consumption data is power consumption actual data in a first time interval, and the second power consumption data is power consumption simulation data in a second time interval; or the first electric quantity data is the actual electric quantity data in a third time interval, and the second electric quantity data is the actual electric quantity data in a fourth time interval; or the first electric quantity data is electric quantity simulation data in a fifth time interval, and the second electric quantity data is electric quantity actual data in a sixth time interval; or, the first electric quantity data is electric quantity simulation data in a seventh time interval, and the second electric quantity data is electric quantity simulation data in an eighth time interval; or, the first electricity consumption data is combined data comprising electricity consumption simulation data in a ninth time interval and electricity consumption actual data in a tenth time interval, and the second electricity consumption data is electricity consumption actual data in an eleventh time interval; or, the first electricity consumption data is combined data comprising electricity consumption simulation data in a twelfth time interval and electricity consumption actual data in a thirteenth time interval, and the second electricity consumption data is electricity consumption simulation data in a fourteenth time interval; or, the first electricity consumption data is actual electricity consumption data in a fifteenth time interval, and the second electricity consumption data is combined data comprising simulation electricity consumption data in a sixteenth time interval and actual data in a seventeenth time interval; or, the first power consumption data is simulation data in an eighteenth time interval, and the second power consumption data is combined data comprising power consumption simulation data in a nineteenth time interval and power consumption actual data in a twentieth time interval.

In one embodiment, updating at least one hidden layer of the N hidden layers included in the power usage prediction model with second power usage data includes: and training the prediction model by using the second electricity consumption data, wherein a preset M hidden layers in the prediction model are fixed, and the rest hidden layers except the M hidden layers in the prediction model are updated, wherein M is a positive integer of at least 2, and M is less than or equal to N.

As shown in fig. 11, the power consumption amount prediction apparatus 80 of the HVAC system includes:

a receiving module 81 for receiving the predicted time;

a prediction module 82 for generating a predicted value of power usage corresponding to the predicted time based on a predicted model of power usage of the HVAC system; the method for generating the electricity consumption prediction model comprises the following steps: acquiring first power data of the HVAC system; acquiring second electricity consumption data of the HVAC system, wherein the data type of the second electricity consumption data is different from the data type of the first electricity consumption data; training a neural network model comprising N hidden layers into the electricity consumption prediction model by utilizing the first electricity consumption data, wherein N is a positive integer which is at least 2; updating at least one hidden layer among the N hidden layers included in the power consumption prediction model with the second power consumption data.

The implementation of embodiments of the present invention will be described below using HVAC in a building as an example.

The building is a large electricity consumer, and the HVAC system accounts for 30 to 40 percent of the total electricity consumption. Building owners or operators desire to be able to reduce energy costs while meeting comfortable room temperature, good air quality, etc. operating requirements. The AI technology can balance the energy waste between the demand party and the supply party, improve the energy utilization efficiency and reduce the energy consumption. However, current AI technology for HVAC presents challenges in application. These challenges include:

(1) There are data availability issues. High quality enough data is important for data driven models. Without the data, the model cannot be trained and tested, let alone providing insight or advice. While digital and internet of things technologies have not been widely used, a significant portion of existing buildings have been built. Some buildings have installed some sensors for system monitoring, but the data collected from these sensors is insufficient for model training and testing. In addition, installing new sensors in current systems is both difficult and expensive.

(2) There is a problem of mobility of the solution. If a solution can be easily transferred from one customer to another, even from one type of customer (e.g., an office building) to another type of customer (e.g., a business building), development costs will be significantly reduced.

In the embodiment of the invention, the realization data of the building with little data can be enriched. For example, the energy consumption of hvac systems has a seasonal law (e.g., high summer refrigeration demand and low winter refrigeration demand) how only available winter data is used to predict summer electricity consumption. In embodiments of the present invention, transfer learning may also be implemented between similar buildings, or from one type of building (e.g., an office building) to another type of building (e.g., a shopping mall), or from one season (e.g., winter) to another season (e.g., summer).

The method of predicting power usage of an HVAC system is described and evaluated by way of example. The data is from a cooling system of an LCD manufacturing plant. It has 33 data characteristics including outdoor temperature, humidity, secondary chilled water loop operating parameters (e.g., feed and return water temperatures) and pump flow, among others. The response variable is the total power consumption of the cooling system. Historical data for four months (2017, 3, 4, 6, and 8 months) may be provided.

Assume that: actual data for month 3 of 2017 (sample size 568) and simulation data for month 4 of 2017 (sample size 617) are provided.

The targets are as follows: the electricity consumption amounts of 2017, 6 (sample size 584) and 2017, 8 (sample size 528) were predicted. The adopted KPI is as follows: root mean square error (Mean Absolution Percentage Error, MAPE):

table 1 is a schematic representation of a predictive model trained in various ways.

In the table, the standard method of training does not adopt transfer learning, namely, training data is directly used for training a prediction model. In the transfer learning, a prediction model is first trained using pre-training data (data for pretraining), and then hidden layers in the prediction model are updated using the training data.

As can be seen from table 1, after the transfer learning is adopted, the running time of training is significantly reduced, and MAPE is also significantly reduced, so that the accuracy of the prediction model is also improved.

TABLE 1

In fig. 12, the abscissa indicates time and the ordinate indicates the amount of electricity consumption. Curve 61 is the actual power usage in winter; curve 62 is the simulated electricity consumption in summer; curve 63 is the actual power usage in summer; curve 64 is the predicted electricity usage for an embodiment of the present invention; curve 65 is the predicted power usage using standard methods.

In fig. 13, the HVAC electricity consumption prediction apparatus 30 includes:

an interface 31 for receiving a predicted time;

a memory 32 for storing a power usage prediction model of a heating HVAC system, wherein the method of generating the power usage prediction model comprises: acquiring first power data of the HVAC system; acquiring second electricity consumption data of the HVAC system, wherein the data type of the second electricity consumption data is different from the data type of the first electricity consumption data; training a neural network model comprising a plurality of hidden layers using the first power data to obtain a power usage prediction model of the HVAC system; training the power usage prediction model with the second power usage data to update at least one hidden layer of the HVAC system;

a processor 33 is coupled with the interface 31 and the memory 32, respectively, via a bus 34 for generating a predicted value of the amount of electricity usage corresponding to the predicted time based on the predicted model of the amount of electricity usage.

Fig. 14 is an exemplary configuration diagram of an apparatus for generating a prediction model of an analysis object according to an embodiment of the present invention. The apparatus 50 includes:

a first data acquisition module 51 for acquiring first data of an analysis object;

A second data obtaining module 52, configured to obtain second data of the analysis object, where the second data and the first data represent the same physical quantity of the analysis object, and a data type of the second data is different from a data type of the first data;

a training module 53, configured to train a neural network model including N hidden layers into a prediction model of the analysis object using the first data, where N is a positive integer of at least 2;

an updating module 54 for updating at least one hidden layer of the N hidden layers included in the prediction model with the second data.

In one embodiment, the first data is actual data in a first time interval, and the second data is simulation data in a second time interval; or, the first data is actual data in a third time interval, and the second data is actual data in a fourth time interval; or, the first data is simulation data in a fifth time interval, and the second data is actual data in a sixth time interval; or, the first data is simulation data in a seventh time interval, and the second data is simulation data in an eighth time interval; or, the first data is combined data comprising simulation data in a ninth time interval and actual data in a tenth time interval, and the second data is actual data in an eleventh time interval; or, the first data is combined data comprising simulation data in a twelfth time interval and actual data in a thirteenth time interval, and the second data is simulation data in a fourteenth time interval; or, the first data is actual data in a fifteenth time interval, and the second data is combined data comprising simulation data in a sixteenth time interval and actual data in a seventeenth time interval; or, the first data is simulation data in an eighteenth time interval, and the second data is combined data comprising simulation data in a nineteenth time interval and actual data in a twentieth time interval.

In one embodiment, the actual data and the simulation data in the combined data have data indicators that are superimposable on a temporal attribute or data indicators that are not superimposable on a temporal attribute.

In one embodiment, the apparatus 50 further comprises: a simulation data acquisition module 55 for establishing a simulation model of the analysis object based on predetermined analysis object metadata; the simulation data is generated based on the simulation model.

In one embodiment, the analysis object is an HVAC system and the predictive model is a power usage predictive model; the apparatus 50 further comprises: a receiving module 56 for receiving the predicted time; a prediction module 57 for generating a power consumption prediction value corresponding to the prediction time based on the power consumption prediction model.

In one embodiment, the updating module 54 is configured to train the prediction model using the second data, wherein a predetermined M hidden layers in the prediction model are fixed, and the remaining hidden layers in the prediction model except for the M hidden layers are updated, where M is a positive integer of at least 2, and M is less than or equal to N.

In fig. 15, the means 70 for generating a predictive model of an object of analysis comprises a memory 72 and a processor 71; the memory 72 has stored therein an application executable by the processor 71 for causing the processor 71 to perform the method of generating a predictive model of an analysis object as set forth in any one of the above.

It should be noted that not all the steps and modules in the above processes and the structure diagrams are necessary, and some steps or modules may be omitted according to actual needs. The execution sequence of the steps is not fixed and can be adjusted as required. The division of the modules is merely for convenience of description and the division of functions adopted in the embodiments, and in actual implementation, one module may be implemented by a plurality of modules, and functions of a plurality of modules may be implemented by the same module, and the modules may be located in the same device or different devices.

The hardware modules in the various embodiments may be implemented mechanically or electronically. For example, a hardware module may include specially designed permanent circuits or logic devices (e.g., special purpose processors such as FPGAs or ASICs) for performing certain operations. A hardware module may also include programmable logic devices or circuits (e.g., including a general purpose processor or other programmable processor) temporarily configured by software for performing particular operations. As regards implementation of the hardware modules in a mechanical manner, either by dedicated permanent circuits or by circuits that are temporarily configured (e.g. by software), this may be determined by cost and time considerations.

The present invention also provides a machine-readable storage medium storing instructions for causing a machine to perform a method as described herein. Specifically, a system or apparatus provided with a storage medium on which a software program code realizing the functions of any of the above embodiments is stored, and a computer (or CPU or MPU) of the system or apparatus may be caused to read out and execute the program code stored in the storage medium. Further, some or all of the actual operations may be performed by an operating system or the like operating on a computer based on instructions of the program code. The program code read out from the storage medium may also be written into a memory provided in an expansion board inserted into a computer or into a memory provided in an expansion unit connected to the computer, and then, based on instructions of the program code, a CPU or the like mounted on the expansion board or the expansion unit may be caused to perform part or all of actual operations, thereby realizing the functions of any of the above embodiments. Storage medium implementations for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (e.g., CD-ROMs, CD-R, CD-RWs, DVD-ROMs, DVD-RAMs, DVD-RWs, DVD+RWs), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, the program code may be downloaded from a server computer or cloud by a communications network.

The invention has been illustrated and described in detail in the drawings and preferred embodiments, but the invention is not limited to these disclosed embodiments. Based on the above embodiments, those skilled in the art can appreciate that the code auditing means in the above different embodiments can be combined to obtain further embodiments of the present invention, and these embodiments are also within the protection scope of the present invention.

Claims

A method (100) of generating a predictive model of an analysis object, comprising:

acquiring (101) first data of an analysis object;

acquiring (102) second data of the analysis object, wherein the second data and the first data characterize the same physical quantity of the analysis object, and the data type of the second data is different from the data type of the first data;

training (103) a neural network model comprising N hidden layers as a predictive model of the analysis object using the first data, wherein N is a positive integer of at least 2;

updating (104) at least one hidden layer of the N hidden layers contained in the predictive model with the second data.
The method (100) of claim 1, wherein the second data has a data type different from the data type of the first data, comprising:

The first data are actual data in a first time interval, and the second data are simulation data in a second time interval; or (b)

The first data is actual data in a third time interval, and the second data is actual data in a fourth time interval; or (b)

The first data are simulation data in a fifth time interval, and the second data are actual data in a sixth time interval; or (b)

The first data are simulation data in a seventh time interval, and the second data are simulation data in an eighth time interval; or (b)

The first data is combined data comprising simulation data in a ninth time interval and actual data in a tenth time interval, and the second data is actual data in an eleventh time interval; or (b)

The first data is combined data comprising simulation data in a twelfth time interval and actual data in a thirteenth time interval, and the second data is simulation data in a fourteenth time interval; or (b)

The first data is actual data in a fifteenth time interval, and the second data is combined data comprising simulation data in a sixteenth time interval and actual data in a seventeenth time interval; or (b)

The first data is simulation data in an eighteenth time interval, and the second data is combined data comprising simulation data in a nineteenth time interval and actual data in a twentieth time interval.
The method (100) according to claim 2, wherein the actual data and the simulated data in the combined data have data indicators that are superimposable on a temporal attribute or data indicators that are not superimposable on a temporal attribute.
The method (100) of claim 2, further comprising:

establishing a simulation model of the analysis object based on predetermined analysis object metadata;

the simulation data is generated based on the simulation model.
The method (100) of claim 1, wherein the analysis object is a heating ventilation and air conditioning, HVAC, system and the predictive model is a power usage predictive model; the method (100) further comprises:

receiving a predicted time;

and generating a power consumption predicted value corresponding to the predicted time based on the updated prediction model.
The method (100) according to any one of claims 1-5, wherein,

the updating (104) at least one hidden layer of the N hidden layers included in the predictive model with second data comprises:

And training the prediction model by using the second data, wherein a preset M hidden layers in the prediction model are fixed, and updating the rest hidden layers except the M hidden layers in the prediction model, wherein M is a positive integer which is at least 2, and M is less than or equal to N.
An apparatus (50) for generating a predictive model of an analysis object, comprising:

a first data acquisition module (51) for acquiring first data of an analysis object;

a second data acquisition module (52) for acquiring second data of the analysis object, wherein the second data and the first data characterize the same physical quantity of the analysis object, and a data type of the second data is different from a data type of the first data;

a training module (53) for training a neural network model comprising N hidden layers into a predictive model of the analysis object using the first data, wherein N is a positive integer of at least 2;

an updating module (54) for updating at least one hidden layer of the N hidden layers included in the predictive model with the second data.
The apparatus (50) of claim 7, wherein the second data is of a different data type than the first data, comprising:

The first data are actual data in a first time interval, and the second data are simulation data in a second time interval; or (b)

The first data is actual data in a third time interval, and the second data is actual data in a fourth time interval; or (b)

The first data are simulation data in a fifth time interval, and the second data are actual data in a sixth time interval; or (b)

The first data are simulation data in a seventh time interval, and the second data are simulation data in an eighth time interval; or (b)

The first data is combined data comprising simulation data in a ninth time interval and actual data in a tenth time interval, and the second data is actual data in an eleventh time interval; or (b)

The first data is combined data comprising simulation data in a twelfth time interval and actual data in a thirteenth time interval, and the second data is simulation data in a fourteenth time interval; or (b)

The first data is actual data in a fifteenth time interval, and the second data is combined data comprising simulation data in a sixteenth time interval and actual data in a seventeenth time interval; or (b)

The first data is simulation data in an eighteenth time interval, and the second data is combined data comprising simulation data in a nineteenth time interval and actual data in a twentieth time interval.
The apparatus (50) of claim 8, wherein,

the actual data and the simulation data in the combined data have data indexes which can be overlapped on the time attribute or data indexes which cannot be overlapped on the time attribute.
The apparatus (50) of claim 7, further comprising:

a simulation data acquisition module (55) for creating a simulation model of the analysis object based on predetermined analysis object metadata; the simulation data is generated based on the simulation model.
The apparatus (50) of claim 7, wherein,

the analysis object is a Heating Ventilation and Air Conditioning (HVAC) system, and the prediction model is a power consumption prediction model; the device (50) further comprises:

a receiving module (56) for receiving the predicted time;

a prediction module (57) for generating a power consumption prediction value corresponding to the prediction time based on the power consumption prediction model.
The device (50) according to any one of claims 7-11, wherein,

The updating module (54) is configured to train the prediction model using the second data, wherein a predetermined M hidden layers in the prediction model are fixed, and remaining hidden layers in the prediction model except for the M hidden layers are updated, where M is a positive integer of at least 2, and M is less than or equal to N.
Means (70) for generating a predictive model of an analysis object, characterized by comprising a processor (71) and a memory (72);

the memory (72) has stored therein an application executable by the processor (71) for causing the processor (71) to perform the method (100) of generating a predictive model of an analysis object as claimed in any one of claims 1 to 6.
Computer readable storage medium, having stored therein computer readable instructions for performing the method (100) of generating a predictive model of an analysis object according to any of claims 1 to 6.