WO2022110213A1 - 生成分析对象的预测模型的方法、装置和存储介质 - Google Patents

生成分析对象的预测模型的方法、装置和存储介质 Download PDF

Info

Publication number
WO2022110213A1
WO2022110213A1 PCT/CN2020/132933 CN2020132933W WO2022110213A1 WO 2022110213 A1 WO2022110213 A1 WO 2022110213A1 CN 2020132933 W CN2020132933 W CN 2020132933W WO 2022110213 A1 WO2022110213 A1 WO 2022110213A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
time interval
analysis object
prediction model
simulation
Prior art date
Application number
PCT/CN2020/132933
Other languages
English (en)
French (fr)
Inventor
曲颖
白新
施尼盖斯·丹尼尔
管金艳
王焦剑
刘晓南
Original Assignee
西门子(中国)有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 西门子(中国)有限公司 filed Critical 西门子(中国)有限公司
Priority to CN202080105837.6A priority Critical patent/CN116235188A/zh
Priority to PCT/CN2020/132933 priority patent/WO2022110213A1/zh
Publication of WO2022110213A1 publication Critical patent/WO2022110213A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods

Definitions

  • the present invention relates to the field of artificial intelligence (Artificial Intelligence, AI) technology, and in particular, to a method, a device and a storage medium for generating a prediction model of an analysis object.
  • AI Artificial Intelligence
  • Machine Learning is an approach to implementing AI.
  • Machine learning is closely related to pattern recognition, computational statistics, artificial intelligence and other fields.
  • Machine learning can use machines (computers and software) to mine meaning from known data, giving the machine learning environment the power.
  • Machine learning algorithms can include supervised learning (such as classification problems), unsupervised learning (such as clustering problems), semi-supervised learning, ensemble learning, deep learning, and reinforcement learning, among others.
  • Predictive analytics is a broad application of machine learning to business problems.
  • the process of using training data to generate a predictive model is the model training process.
  • the embodiments of the present invention propose a method, an apparatus, and a storage medium for generating a prediction model of an analysis object.
  • a method for generating a predictive model of an object of analysis includes:
  • N is a positive integer at least 2;
  • At least one hidden layer of the N hidden layers included in the prediction model is updated using the second data.
  • an apparatus for generating a predictive model of an analysis object including:
  • a first data acquisition module used for acquiring the first data of the analysis object
  • a second data acquisition module configured to acquire 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 the same as that of the analysis object.
  • the data types of the first data are different;
  • a training module configured to use the first data to train a neural network model comprising N hidden layers as a prediction model for the analysis object, where N is a positive integer that is at least 2;
  • An update module for updating at least one hidden layer of the N hidden layers included in the prediction model by using the second data.
  • an apparatus for generating a predictive model of an object of analysis including a processor and a memory;
  • An application program executable by the processor is stored in the memory for causing the processor to execute the method for generating a predictive model of an analysis object as described in any of the above.
  • a computer-readable storage medium in which computer-readable instructions are stored, the computer-readable instructions for performing the method of generating a predictive model of an object of analysis as described in any of the above.
  • the embodiments of the present invention can use multiple data of multiple data types to generate the training model of the analysis object, which can reduce the dependence of the model on a single type of data. Moreover, since the data characteristics of the second data and the first data represent the same physical quantity of the analysis object, part of the hidden layers in the prediction model can be updated by using the second data, instead of retraining the entire prediction model, thus speeding up the model training. The process realizes the transfer learning of the model.
  • the data type of the second data is different from the data type of the first data, including:
  • the first data is actual data in a first time interval
  • the second data is simulated data in a second time interval
  • the first data is the actual data in the third time interval
  • the second data is the actual data in the fourth time interval
  • the first data is simulated data in a fifth time interval, and the second data is actual data in a sixth time interval;
  • the first data is the simulation data in the seventh time interval
  • the second data is the simulation data in the eighth time interval
  • the first data is the combined data including the simulation data in the ninth time interval and the actual data in the tenth time interval, and the second data is the actual data in the eleventh time interval;
  • the first data is the combined data including the simulation data in the twelfth time interval and the actual data in the thirteenth time interval, and the second data is the simulation data in the fourteenth time interval;
  • the first data is the actual data in the fifteenth time interval
  • the second data is the combined data including the simulation data in the sixteenth time interval and the actual data in the seventeenth time interval;
  • the first data is simulation data in the eighteenth time interval
  • the second data is combined data including the simulation data in the nineteenth time interval and the actual data in the twentieth time interval.
  • the actual data and the simulated data in the combined data have data indicators that can be superimposed on time attributes or data indicators that cannot be superimposed on time attributes.
  • first data and the second data in the embodiment of the present invention have various types, which enriches the training data and also improves the accuracy of the model.
  • the simulation data is generated based on the simulation model.
  • the simulation data can be quickly acquired through the simulation model, which improves the data acquisition efficiency.
  • the analysis object is a heating, ventilation and air conditioning (HVAC) system
  • the prediction model is a power consumption prediction model; further comprising: receiving a prediction time; generating based on the updated prediction model The predicted value of electricity consumption corresponding to the predicted time.
  • HVAC heating, ventilation and air conditioning
  • the updated prediction model can be applied to the HVAC system.
  • the updating at least one hidden layer of the N hidden layers included in the prediction model using the second data includes:
  • the prediction model is trained using the second data, wherein the predetermined M hidden layers in the prediction model are fixed, and the remaining hidden layers in the prediction model except the M hidden layers are updated A containing layer, wherein M is a positive integer of at least 2, and M is less than or equal to N.
  • the M hidden layers can be retained as mature knowledge, thereby realizing knowledge transfer and reducing the training workload after transfer.
  • FIG. 1 is an exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • FIG. 2 is a first exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • FIG. 3 is a second exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • FIG. 4 is a third exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • FIG. 5 is a fourth exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • FIG. 6 is a fifth exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • FIG. 7 is a sixth exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • FIG. 8 is a seventh exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • FIG. 9 is an eighth exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • FIG. 10 is a flowchart of a method for predicting power consumption of an HVAC system according to an embodiment of the present invention.
  • FIG. 11 is a configuration diagram of a power consumption prediction device of an HVAC system according to an embodiment of the present invention.
  • FIG. 12 is a schematic diagram of prediction of HVAC power consumption according to an embodiment of the present invention.
  • FIG. 13 is a configuration diagram of an HVAC power consumption prediction device 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 structural diagram of an apparatus for generating a prediction model of an analysis object with a memory-processor architecture according to an embodiment of the present invention.
  • processor 34 bus 50 Apparatus for analyzing predictive models of objects 51
  • Simulation data acquisition module 56 receiving module 57 prediction module 70
  • the embodiment of the present invention uses the first data to train the prediction model including the hidden layer, and then uses the second data to update part of the hidden layers in the prediction model, which can enrich the training data, and improve model performance. Moreover, it also addresses the transferability of predictive models and speeds up the model training process.
  • FIG. 1 is an exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • the method includes:
  • Step 101 Acquire the first data of the analysis object.
  • the analysis object is the object that needs to be analyzed, for example, it can be implemented as a thermal system or a power system, and so on.
  • the object of analysis is implemented as a heating ventilation and air conditioning (HVAC) system.
  • HVAC heating ventilation and air conditioning
  • the data type of the first data may be determined by the data acquisition method or data source of the first data.
  • the data type of the first data may include:
  • Type (1) the actual data of the analysis object (for example, historical data);
  • Type (2) Simulation data of the analysis object
  • Type (3) Combined data of the analysis object, including actual data and simulated data.
  • the first data represents a predetermined physical quantity of the object of analysis.
  • the first data may be used to characterize the power consumption of the analysis object, the fan usage time, the refrigerant usage, and so on.
  • Step 102 Acquire 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.
  • the data type of the second data may be determined by the data acquisition method or data source of the second data.
  • the data type of the second data may include:
  • Type (1) the actual data of the analysis object (for example, historical data);
  • Type (2) Simulation data of the analysis object
  • Type (3) Combined data of the analysis object, including actual data and simulated data.
  • the data type of the second data is different from the data type of the first data. Therefore, 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 type (2), then the first data type is type (2).
  • the data type of the second data is type (1) or type (3); when the data type of the first data is type (3), the data type of the second data is type (1) or type (2).
  • the second data and the first data represent the same physical quantity of the analysis object.
  • the second data also represents the power consumption of the analysis object;
  • the first data represents the fan usage time of the analysis object, the second data also represents the fan usage of the analysis object time, wait.
  • the second data is more relevant to the prediction task of the finally obtained prediction model.
  • the first data is actual data in a first time interval
  • the second data is simulation data in a second time interval. Both the first time interval and the second time interval can be described by units of time such as years, quarters, months, weeks, or days.
  • the first data is the real electricity consumption data of factory A including the analysis object (such as HVAC system) in March 2012
  • the second data is the factory A that includes the analysis object (such as HVAC system) in April 2012
  • the first data is actual data in a third time interval
  • the second data is actual data in a fourth time interval.
  • Both the third time interval and the fourth time interval can be described by units of time such as years, quarters, months, weeks, or days.
  • the first data is the real electricity consumption data of factory A including the analysis object (such as HVAC system) in March 2012
  • the second data is the actual electricity consumption data of factory A including the analysis object (such as HVAC system) in April 2012 Actual electricity consumption data, or the actual electricity consumption data of plant B in April 2012 that contains the object of analysis (such as HVAC system).
  • the first data is simulated data in a fifth time interval
  • the second data is actual data in a sixth time interval. Both the fifth time interval and the sixth time interval can be described by time measurement units such as year, quarter, month, week, or day.
  • the first data is the simulated electricity consumption data of factory A including the analysis object (such as HVAC system) in March 2012
  • the second data is the data of factory A including the analysis object (such as HVAC system) in April 2012 Actual electricity consumption data, or the actual electricity consumption data of plant B in April 2012 that contains the object of analysis (such as HVAC system).
  • the first data is simulation data in a seventh time interval
  • the second data is simulation data in an eighth time interval. Both the seventh time interval and the eighth time interval can be described by time measurement units such as year, quarter, month, week, or day.
  • the first data is the simulated electricity consumption data of factory A including the analysis object (such as HVAC system) in March 2012
  • the second data is the data of factory A including the analysis object (such as HVAC system) in April 2012
  • the first data is combined data including simulation data in the ninth time interval and actual data in the tenth time interval
  • the second data is actual data in the eleventh time interval.
  • the ninth time interval, the tenth time interval, and the eleventh time interval can all be described by time measurement units such as year, quarter, month, week, or day.
  • the first data is the combined data of the simulated electricity consumption data in April 2012 and the real electricity consumption data in March 2012 of Factory A including the analysis object (such as the HVAC system);
  • the second data is the combination data including the analysis object
  • the first data is combined data including simulation data in the twelfth time interval and actual data in the thirteenth time interval
  • the second data is simulation data in the fourteenth time interval.
  • the twelfth time interval, the thirteenth time interval, and the fourteenth time interval can all be described by time measurement units such as year, quarter, month, week, or day.
  • the first data is the combined data of the simulated electricity consumption data in April 2012 and the real electricity consumption data in March 2012 of Factory A including the analysis object (such as the HVAC system);
  • the second data is the combination data including the analysis object
  • the first data is actual data in the fifteenth time interval
  • the second data is combined data including simulation data in the sixteenth time interval and actual data in the seventeenth time interval.
  • the fifteenth time interval, the sixteenth time interval, and the seventeenth time interval can all be described by time measurement units such as year, quarter, month, week, or day.
  • the first data is the actual electricity consumption data of factory A including the analysis object (such as HVAC system) in May 2012
  • the second data is the actual electricity consumption data of factory A including the analysis object (such as HVAC system) in April 2012
  • the first data is simulation data in the eighteenth time interval
  • the second data is combined data including the simulation data in the nineteenth time interval and the actual data in the twentieth time interval.
  • the eighteenth time interval, the nineteenth time interval, and the twentieth time interval can all be described by time measurement units such as year, quarter, month, week, or day.
  • the first data is the simulated electricity consumption data of factory A including the analysis object (such as HVAC system) in May 2012
  • the second data is the factory A including the analysis object (such as the HVAC system) in April 2012.
  • the actual data and simulated data in the combined data have data metrics that can be superimposed over time.
  • the actual data and the simulated data in the combined data are respectively implemented as electricity consumption data on the superimposable months.
  • the actual data includes the electricity consumption data in January 2011 and the electricity consumption data in February 2011, the simulated data is the electricity consumption data in March 2011, and the combined data is the first quarter of 2011 (including 1 Month to March) electricity consumption data.
  • the actual data and simulated data in the combined data have data metrics that are not superimposable in temporal attributes.
  • the actual data is the electricity consumption data in January 2011, and the simulated data is the room temperature value in January 2011.
  • the actual data is the electricity consumption data in January 2011, and the simulated data is the fan usage time in February 2011.
  • Simulation data can be obtained using the simulation model of the analysis object. Specifically, it includes: establishing a simulation model of the analysis object based on predetermined analysis object metadata (eg, building information model, design drawings, etc.); and generating simulation data based on the simulation model.
  • predetermined analysis object metadata eg, building information model, design drawings, etc.
  • Step 103 using the first data to train a neural network model including N hidden layers as a prediction model of the analysis object, where N is a positive integer of at least 2.
  • the neural network model can be implemented as: a feedforward neural network model, a radial basis neural network model, a long short-term memory (LSTM) network model, an echo state network (ESN), a gated recurrent unit (GRU) network model, or a deep residual Poor network model, etc.
  • the neural network model is implemented as an LSTM network model.
  • the first data is used to train a neural network model including N hidden layers as a prediction model for the analysis object.
  • parameters eg, weights
  • Step 104 Using the second data to update at least one hidden layer of the N hidden layers included in the prediction model.
  • the prediction model of the analysis object trained by using the first data in step 103 may be used to perform prediction analysis on the analysis object applied to each scene.
  • the second data when the application scenario of the second data is the same as the application scenario of the first data, it is preferable to use the second data to perform a predetermined part of the hidden layer in the prediction model of the analysis object trained from the first data. update, thereby improving the accuracy of the predictive model.
  • it can be determined whether to use the specific hidden layer as the hidden layer that needs to be updated.
  • the specific hidden layer is regarded as the hidden layer that needs to be updated;
  • the prediction effect of updating and not updating a specific hidden layer is not significantly improved (for example, the improvement of the prediction accuracy is less than or equal to a predetermined threshold value)
  • the specific hidden layer is not used as the hidden layer that needs to be updated.
  • the first data is usually derived from a specific scene, and the analysis object may be applied to another application scene related to the specific scene, it is preferable to use the second data derived from the other application scene to compare the first data with the second data.
  • the hidden layer that is related to the application scenario of the second data is updated, so that the updated prediction model is more suitable for the other application than the prediction model before the update. scenarios, so as to realize knowledge transfer.
  • the manner of determining the hidden layer related to the application scenario of the second data includes:
  • the specific hidden layer determines whether the specific hidden layer is related to the application scenario of the second data. For example, when the prediction effect of updating and not updating a specific hidden layer is significantly improved (for example, the improvement of the prediction accuracy is greater than a predetermined threshold), then the specific hidden layer is associated with the application scenario of the second data Therefore, it is the hidden layer that needs to be updated; when the prediction effect of updating and not updating a specific hidden layer is not significantly improved (for example, the improvement of the prediction accuracy is less than or equal to a predetermined threshold), then the specific hidden layer The application scenario of the containing layer and the second data is not related, so it is not used as the hidden layer that needs to be updated.
  • using the second data to update at least one hidden layer of the N hidden layers included in the prediction model includes: using the second data to train the prediction model, wherein a predetermined M of the prediction models are fixed Hidden layer, update the remaining hidden layers in the prediction model except M hidden layers, where M is a positive integer of at least 2, and M is less than or equal to N.
  • FIG. 2 is a first exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • the first data 201 of the analysis object is: the actual data of the analysis object in the first time interval collected in the first application scene including the analysis object;
  • the second data of the analysis object 204 is: simulation data of the analysis object in the second time interval collected in the first application scenario including the analysis object.
  • the first time interval and the second time interval may be the same or different, preferably different.
  • the prediction task is the electricity consumption prediction in the first application scenario (such as Factory A) including the analysis object (such as the HVAC system);
  • the first data 201 is the actual usage time of the fans in Factory A in January 2011;
  • the second Data 204 is the simulated usage time of the fans in Factory A in February 2011.
  • the first data 201 is input into the neural network model 202 including N hidden layers as training data.
  • the neural network model 202 can be trained as the prediction model 203 of the analysis object.
  • the prediction model 203 the parameters of N hidden layers are all determined.
  • the second data 204 is input into the prediction model 203 as training data to perform training again.
  • the predetermined M hidden layers among the N hidden layers of the prediction model 203 are kept fixed (that is, the parameters of the M hidden layers are not updated), and the remaining M hidden layers are not updated by the retraining process. , the parameters of (N-M) hidden layers.
  • the updated prediction model 205 can be obtained. At this time, prediction of the power consumption of the analysis object in Plant A can be performed using the updated prediction model 205 .
  • the first data 201 of the analysis object is: the actual data of the analysis object in the first time interval collected in the first application scene including the analysis object; the second data of the analysis object
  • the data 204 is: simulation data of the analysis object in the second time interval collected in the second application scenario including the analysis object.
  • the first time interval and the second time interval may be the same or different, preferably different.
  • the second data 204 is more relevant to the prediction task in the second application scenario.
  • the prediction task is the electricity consumption prediction in the second application scenario (such as factory B) including the analysis object (such as the HVAC system);
  • the first data 201 is the fan usage time in February 2011 in the factory A;
  • the second data 204 is the fan simulation usage time in January 2011 in factory B.
  • the first data 201 is input into the neural network model 202 including N hidden layers as training data.
  • the neural network model 202 can be trained as the prediction model 203 of the analysis object.
  • the prediction model 203 the parameters of N hidden layers are all determined.
  • the second data 204 is input into the prediction model 203 as training data to perform training again.
  • the predetermined M hidden layers in the N hidden layers of the prediction model 203 that are not related to the application scenario of the second data 204 are kept fixed (that is, the M hidden layers of the M hidden layers are not updated). parameters), and use the retraining process to update the parameters of the remaining (N-M) hidden layers related to the application scenario of the second data 204.
  • the updated prediction model 205 can be obtained.
  • prediction of the power consumption of the object of analysis in Plant B can be performed using the updated prediction model 205 .
  • FIG. 3 is a second exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • the first data 301 of the analysis object is: the actual data of the analysis object in the third time interval collected in the first application scene including the analysis object; the second data of the analysis object 304 is: the actual data of the analysis object in the fourth time interval collected in the first application scenario including the analysis object.
  • the third time interval and the fourth time interval may be the same or different.
  • the prediction task is the electricity consumption prediction in the first application scenario (eg Factory A) including the analysis object (eg HVAC system); the first data 301 is the actual electricity consumption data in February 2011 in Factory A; The second data 304 is the actual data of electricity consumption in factory A in January 2011.
  • the first application scenario eg Factory A
  • the analysis object eg HVAC system
  • the first data 301 is the actual electricity consumption data in February 2011 in Factory A
  • the second data 304 is the actual data of electricity consumption in factory A in January 2011.
  • the first data 301 is input into the neural network model 302 including N hidden layers as training data.
  • the neural network model 302 can be trained as the prediction model 303 of the analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 304 is input into the prediction model 303 as training data to perform training again.
  • the predetermined M hidden layers among the N hidden layers of the prediction model 303 are kept fixed (that is, the parameters of the M hidden layers are not updated), and the remaining M hidden layers are not updated by the retraining process. , the parameters of (N-M) hidden layers.
  • the updated prediction model 305 can be obtained. At this time, prediction of the power consumption of the object of analysis in the plant A can be performed using the updated prediction model 305 .
  • the first data 301 of the analysis object is: the actual data of the analysis object in the third time interval collected in the first application scene including the analysis object; the second data of the analysis object
  • the data 304 is: actual data of the analysis object in the fourth time interval collected in the second application scenario including the analysis object.
  • the third time interval and the fourth time interval may be the same or different.
  • the second data 304 is more relevant to the prediction task in the second application scenario.
  • the prediction task is the electricity consumption forecast in the second application scenario (eg factory B) including the analysis object (eg HVAC system);
  • the first data 301 is the actual electricity consumption data in February 2011 in factory A;
  • the second data 304 is the actual data of electricity consumption in factory B in January 2011.
  • the first data 301 is input into the neural network model 302 including N hidden layers as training data.
  • the neural network model 302 can be trained as the prediction model 303 of the analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 304 is input into the prediction model 303 as training data to perform training again.
  • the predetermined M hidden layers in the N hidden layers of the prediction model 303 that are not related to the application scenario of the second data 304 are kept fixed (that is, the M hidden layers of the M hidden layers are not updated). parameters), and the parameters of the remaining (N-M) hidden layers related to the application scenario of the second data 304 are updated by using the retraining process.
  • the updated prediction model 305 can be obtained.
  • prediction of the power consumption of the object of analysis in Plant B can be performed using the updated prediction model 305 .
  • FIG. 4 is a third exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • the first data 401 of the analysis object is: the simulation data of the analysis object in the fifth time interval collected in the first application scene including the analysis object;
  • the second data 404 of the analysis object is: The actual data of the analysis object in the sixth time interval collected in the first application scenario of the analysis object is included.
  • the fifth time interval and the sixth time interval may be the same or different, preferably different.
  • the prediction task is the electricity consumption prediction in the first application scenario (such as Factory A) including the analysis object (such as the HVAC system);
  • the first data 401 is the simulated usage time of fans in Factory A in February 2011;
  • the second Data 404 is the actual usage time of the fans in Factory A in January 2011.
  • the first data 401 is input into the neural network model 402 including N hidden layers as training data.
  • the neural network model 402 can be trained as the prediction model 403 of the analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 404 is input into the prediction model 403 as training data to perform training again.
  • the predetermined M hidden layers among the N hidden layers of the prediction model 403 are kept fixed (that is, the parameters of the M hidden layers are not updated), and the remaining M hidden layers are not updated by the retraining process. , the parameters of (N-M) hidden layers.
  • the updated prediction model 405 can be obtained. At this time, prediction of the power consumption of the object of analysis in the plant A can be performed using the updated prediction model 405 .
  • the first data 401 of the analysis object is: the simulation data of the analysis object in the fifth time interval collected in the first application scene including the analysis object; the second data of the analysis object
  • the data 404 is: actual data of the analysis object in the sixth time interval collected in the second application scenario including the analysis object.
  • the fifth time interval and the sixth time interval may be the same or different.
  • the second data 404 is more relevant to the prediction task in the second application scenario.
  • the prediction task is the electricity consumption prediction in the second application scenario (such as factory B) including the analysis object (such as the HVAC system);
  • the first data 401 is the simulated usage time of fans in factory A in February 2011;
  • the second Data 404 is the actual usage time of the fans in factory B in January 2011.
  • the first data 401 is input into the neural network model 402 including N hidden layers as training data.
  • the neural network model 402 can be trained as the prediction model 403 of the analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 404 is input into the prediction model 403 as training data to perform training again.
  • the retraining process keep the M hidden layers in the N hidden layers of the prediction model 403 that are not related to the application scenario of the second data 404 fixed (that is, do not update the parameters of the M hidden layers) ), and the parameters of the remaining (N-M) hidden layers related to the application scenario of the second data 404 are updated using the retraining process.
  • the updated prediction model 405 can be obtained.
  • prediction of the power consumption of the object of analysis in Plant B can be performed using the updated prediction model 405 .
  • FIG. 5 is a fourth exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • the first data 501 of the analysis object is: the simulation data of the analysis object in the seventh time interval collected in the first application scene including the analysis object;
  • the second data 504 of the analysis object is: The simulation data in the eighth time interval of the analysis object collected in the first application scenario of the analysis object is included.
  • the seventh time interval and the eighth time interval may be the same or different.
  • the prediction task is power consumption prediction in a first application scenario (such as Factory A) including an analysis object (such as an HVAC system);
  • the first data 501 is the power consumption simulation data in Factory A in February 2011;
  • the second data 404 is the power consumption simulation data of factory A in January 2011.
  • the first data 501 is input into the neural network model 502 including N hidden layers as training data.
  • the neural network model 502 can be trained as the prediction model 503 of the analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 504 is input into the prediction model 503 as training data to perform training again.
  • the predetermined M hidden layers among the N hidden layers of the prediction model 503 are kept fixed (that is, the parameters of the M hidden layers are not updated), and the remaining M hidden layers are not updated by the retraining process. , the parameters of (N-M) hidden layers.
  • the updated prediction model 505 can be obtained. At this time, prediction of the power consumption of the object of analysis in the plant A can be performed using the updated prediction model 505 .
  • the first data 501 of the analysis object is: the simulation data of the analysis object in the seventh time interval collected in the first application scenario including the analysis object; the second data of the analysis object
  • the data 504 is: simulation data of the analysis object in the eighth time interval collected in the second application scenario including the analysis object.
  • the seventh time interval and the eighth time interval may be the same or different.
  • the second data 504 is more relevant to the prediction task in the second application scenario.
  • the prediction task is power consumption prediction in a second application scenario (such as Factory B) including an analysis object (such as an HVAC system);
  • the first data 501 is the power consumption simulation data in Factory A in February 2011;
  • the second data 504 is the power consumption simulation data in factory B in January 2011.
  • the first data 501 is input into the neural network model 502 including N hidden layers as training data.
  • the neural network model 502 can be trained as the prediction model 503 of the analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 504 is input into the prediction model 503 as training data to perform training again.
  • the predetermined M hidden layers in the N hidden layers of the prediction model 503 that are not related to the application environment of the second data 504 are kept fixed (that is, the M hidden layers of the M hidden layers are not updated). parameters), and the parameters of the remaining (N-M) hidden layers related to the application environment of the second data 504 are updated using the retraining process.
  • the updated prediction model 505 can be obtained.
  • prediction of the power consumption of the analysis object in Plant B can be performed using the updated prediction model 505 .
  • FIG. 6 is a fifth exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • the first data 601 of the analysis object is: the combined data collected in the first application scenario including the analysis object and including the simulation data in the ninth time interval and the actual data in the tenth time interval;
  • the second data 604 of the analysis object is: actual data of the analysis object in the eleventh time interval collected in the first application scenario including the analysis object.
  • the ninth time interval and the tenth time interval may be the same or different.
  • the second data 604 is more related to the prediction task in the first application scenario.
  • the prediction task is the electricity consumption prediction in the first application scenario (such as Factory A) including the analysis object (such as HVAC system);
  • the combined data of the actual temperature value in March 2011 in A; the second data 604 is the actual use time of the fan in January 2011 in Factory A and the actual data on electricity consumption in Factory A in January 2011.
  • the first data 601 is input into the neural network model 602 including N hidden layers as training data.
  • the neural network model 602 can be trained as the prediction model 603 of the analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 604 is input into the prediction model 603 as training data to perform training again.
  • the predetermined M hidden layers among the N hidden layers of the prediction model 603 are kept fixed (that is, the parameters of the M hidden layers are not updated), and the remaining M hidden layers are not updated by the retraining process. , the parameters of (N-M) hidden layers.
  • the updated prediction model 605 can be obtained. At this time, prediction of the power consumption of the object of analysis in the plant A can be performed using the updated prediction model 605 .
  • the first data 601 of the analysis object is: collected in the first application scenario including the analysis object, including simulation data in the ninth time interval and actual data in the tenth time interval
  • the combined data of the data; the second data 604 of the analysis object is: the actual data of the analysis object in the eleventh time interval collected in the second application scenario including the analysis object.
  • the ninth time interval and the tenth time interval may be the same or different.
  • the second data 604 is more related to the prediction task in the second application scenario.
  • the prediction task is the electricity consumption prediction in the second application scenario (such as factory B) including the analysis object (such as the HVAC system);
  • the first data 601 is the simulated fan usage time and factory data in the factory A in February 2011.
  • the second data 604 is the actual usage time of the fans in factory B in January 2011 and the actual data of electricity consumption in factory A in January 2011.
  • the first data 601 is input into the neural network model 602 including N hidden layers as training data.
  • the neural network model 602 can be trained as the prediction model 603 of the analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 604 is input into the prediction model 603 as training data to perform training again.
  • the predetermined M hidden layers in the N hidden layers of the prediction model 603 that are not related to the application environment of the second data 604 are kept fixed (that is, the M hidden layers are not updated parameters), and use the retraining process to update the parameters of the remaining (N-M) hidden layers related to the application environment of the second data 604.
  • the updated prediction model 605 can be obtained.
  • prediction of the power consumption of the object of analysis in Plant B can be performed using the updated prediction model 605 .
  • FIG. 7 is a sixth exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • the first data 701 of the analysis object is: a combination of simulation data in the twelfth time interval and actual data in the thirteenth time interval, collected in the first application scenario including the analysis object Data;
  • the second data 704 of the analysis object is: simulation data in the fourteenth time interval collected in the first application scenario including the analysis object.
  • the twelfth time interval and the thirteenth time interval may be the same or different, preferably different;
  • the fourteenth time interval and the twelfth time interval may be the same or different, preferably different.
  • the second data 704 is more relevant to the prediction task in the first application scenario.
  • the prediction task is the electricity consumption prediction in the first application scenario (such as factory A) including the analysis object (such as the HVAC system);
  • the first data 701 is the simulation fan usage time and the factory in the factory A in February 2011.
  • the second data 704 is the simulated fan usage time in January 2011 and the simulation data of electricity consumption in January 2011 in Factory A.
  • the first data 701 is input into a neural network model 702 including N hidden layers as training data.
  • the neural network model 702 can be trained as a prediction model 703 of an analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 704 is input into the prediction model 703 as training data to perform training again.
  • the predetermined M hidden layers among the N hidden layers of the prediction model 703 are kept fixed (that is, the parameters of the M hidden layers are not updated), and the remaining M hidden layers are updated using the retraining process , the parameters of (N-M) hidden layers.
  • the updated prediction model 705 can be obtained.
  • prediction of the power consumption of the analysis object in Plant A can be performed using the updated prediction model 705 .
  • the first data 701 of the analysis object is: the simulation data in the twelfth time interval and the thirteenth time interval collected in the first application scenario including the analysis object
  • the second data 704 of the analysis object is: the actual data of the analysis object in the fourteenth time interval collected in the second application scenario including the analysis object.
  • the twelfth time interval and the thirteenth time interval may be the same or different, preferably different
  • the fourteenth time interval and the twelfth time interval may be the same or different, preferably different.
  • the second data 704 is more relevant to the prediction task in the second application scenario.
  • the prediction task is the electricity consumption prediction in the second application scenario (such as factory B) including the analysis object (such as the HVAC system);
  • the first data 701 is the simulation fan usage time and the factory in the factory A in February 2011.
  • the second data 704 is the simulated fan usage time in January 2011 and the simulation data of electricity consumption in January 2011 in Factory B.
  • the first data 701 is input into a neural network model 702 including N hidden layers as training data.
  • the neural network model 702 can be trained as a prediction model 703 of an analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 704 is input into the prediction model 703 as training data to perform training again.
  • the predetermined M hidden layers in the N hidden layers of the prediction model 703 that are not related to the application environment of the second data 704 are kept fixed (that is, the M hidden layers of the M hidden layers are not updated. parameters), and the parameters of the remaining (N-M) hidden layers related to the application environment of the second data 704 are updated by using the retraining process.
  • the updated prediction model 705 can be obtained.
  • prediction of the power consumption of the object of analysis in Plant B can be performed using the updated prediction model 705 .
  • FIG. 8 is a seventh exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • the first data 801 of the analysis object is: actual data collected in the first application scenario including the analysis object and within the fifteenth time interval;
  • the second data 804 of the analysis object is: The combined data of the simulation data in the sixteenth time interval and the actual data in the seventeenth time interval collected in the first application scenario of the 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 prediction task is the electricity consumption forecast in the first application scenario (eg Factory A) including the analysis object (eg HVAC system);
  • the first data 801 includes the actual electricity consumption data in February 2011 and 2011 The actual temperature value in February of the year;
  • the second data 804 is the combined data including the power consumption simulation data in Factory A in January 2011 and the actual temperature value in Factory A in March 2011.
  • the first data 801 is input into a neural network model 802 including N hidden layers as training data.
  • the neural network model 802 can be trained as a prediction model 803 of an analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 804 is input into the prediction model 803 as training data to perform training again.
  • the predetermined M hidden layers among the N hidden layers of the prediction model 803 are kept fixed (that is, the parameters of the M hidden layers are not updated), and the remaining M hidden layers are not updated by the retraining process. , the parameters of (N-M) hidden layers.
  • the updated prediction model 805 can be obtained. At this time, prediction of the power consumption of the object of analysis in the plant A can be performed using the updated prediction model 805 .
  • the first data 801 of the analysis object is: actual data collected in the first application scenario including the analysis object and within the fifteenth time interval;
  • the second data 704 of the analysis object is: the combined data of the simulation data in the sixteenth time interval and the actual data in the seventeenth time interval, collected in the second application scenario including 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.
  • the second data 804 is more relevant to the prediction task in the second application scenario.
  • the prediction task is the electricity consumption forecast in the second application scenario (eg factory B) including the analysis object (eg HVAC system);
  • the first data 801 includes the actual electricity consumption data in February 2011 in factory A and the 2011 The actual temperature value in February of the year;
  • the second data 804 is the combined data including the power consumption simulation data in factory B in January 2011 and the actual temperature value in factory B in March 2011.
  • the first data 801 is input into a neural network model 802 including N hidden layers as training data.
  • the neural network model 802 can be trained as a prediction model 803 of an analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 804 is input into the prediction model 803 as training data to perform training again.
  • the predetermined M hidden layers in the N hidden layers of the prediction model 803 that are not related to the application environment of the second data 804 are kept fixed (that is, the M hidden layers of the M hidden layers are not updated). parameters), and use the retraining process to update the parameters of the remaining (N-M) hidden layers related to the application environment of the second data 804.
  • the updated prediction model 805 can be obtained.
  • prediction of the power consumption of the analysis object in Plant B can be performed using the updated prediction model 805 .
  • FIG. 9 is an eighth exemplary flowchart of a method for generating a prediction model of an analysis object according to an embodiment of the present invention.
  • the first data 901 of the analysis object is: the simulation data in the eighteenth time interval collected in the first application scene including the analysis object;
  • the second data 904 of the analysis object is: The combined data that is collected in the first application scenario of the object and includes the simulation data in the nineteenth time interval and the actual data in the twentieth time interval.
  • 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 prediction task is the electricity consumption prediction in the first application scenario (such as factory A) including the analysis object (such as the HVAC system);
  • the first data 801 includes the simulated fan usage time in February 2011 in factory A and the year The simulation data of electricity consumption in February;
  • the second data 804 is the combined data including the simulated fan usage time in January 2011 in Factory A and the actual data on electricity consumption in March 2011 in Factory A.
  • the first data 901 is input into the neural network model 902 including N hidden layers as training data.
  • the neural network model 902 can be trained as the prediction model 903 of the analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 904 is input into the prediction model 903 as training data to perform training again.
  • the predetermined M hidden layers among the N hidden layers of the prediction model 903 are kept fixed (that is, the parameters of these M hidden layers are not updated), and the remaining M hidden layers are not updated by the retraining process. , the parameters of (N-M) hidden layers.
  • the updated prediction model 905 can be obtained. At this time, prediction of the power consumption of the object of analysis in the plant A can be performed using the updated prediction model 805 .
  • the first data 901 of the analysis object is: simulation data in the eighteenth time interval collected in the first application scenario including the analysis object;
  • the second data 904 of the analysis object is: the combined data collected in the second application scenario including the analysis object and including the simulation data in the nineteenth time interval and the actual data in the twentieth time interval.
  • 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.
  • the second data 904 is more relevant to the prediction task in the second application scenario.
  • the prediction task is the electricity consumption prediction in the second application scenario (such as factory B) including the analysis object (such as the HVAC system);
  • the first data 801 includes the simulated fan usage time in factory A in February 2011 and the year of 2011 The simulation data of electricity consumption in February;
  • the second data 804 is the combined data including the simulated fan usage time in January 2011 in Factory B and the actual data on electricity consumption in March 2011 in Factory B.
  • the first data 801 is input into a neural network model 802 including N hidden layers as training data.
  • the neural network model 802 can be trained as a prediction model 803 of an analysis object.
  • the parameters of N hidden layers are all determined.
  • the second data 804 is input into the prediction model 803 as training data to perform training again.
  • the predetermined M hidden layers in the N hidden layers of the prediction model 803 that are not related to the application environment of the second data 804 are kept fixed (that is, the M hidden layers of the M hidden layers are not updated). parameters), and use the retraining process to update the parameters of the remaining (N-M) hidden layers related to the application environment of the second data 804.
  • the updated prediction model 805 can be obtained.
  • prediction of the power consumption of the analysis object in Plant B can be performed using the updated prediction model 805 .
  • a simulation model of an analysis object may be constructed using the metadata, and simulation data (data set D S ) may be generated using the simulation model.
  • simulation data data set D S
  • a process of training the predictive model of the analysis object is performed.
  • the combined data D C can be used to pair the data that is preferably implemented as the LSTM model.
  • the neural network model is trained to obtain a predictive model of the object under analysis.
  • a neural network model which is preferably implemented as an LSTM model, can also be trained first by using the actual data DA of the analysis object to obtain a prediction model M A of the analysis object.
  • the MA can either be used to perform predictions on the analytic objects.
  • the prediction model MA can be trained by using the simulation data DS of the analysis object.
  • the parameters of the predetermined hidden layer of the prediction model MA are kept fixed, and the parameters of the other hidden layers except the predetermined hidden layer are kept fixed. parameters are updated.
  • the updated prediction model MA has better accuracy than the pre-updated prediction model MA.
  • the following is an exemplary comparison of the effect of updating the hidden layer and not updating the hidden layer in the application environment to which transfer learning is transferred.
  • the number of all trainable parameters is 508201.
  • the number of trainable parameters is 201. Therefore, after updating the hidden layer, the number of trainable parameters is significantly reduced, thereby increasing the training speed.
  • FIG. 10 is a flowchart of a method for predicting power consumption of an HVAC system according to an embodiment of the present invention. As shown in Figure 10, the method includes:
  • Step 1001 receive the predicted time
  • Step 1002 Generate a predicted value of power consumption corresponding to the predicted time based on a power consumption prediction model of the HVAC system; wherein the method for generating the power consumption prediction model includes: acquiring first power consumption data of the HVAC system; Acquiring second power consumption data of the HVAC system; using the first power consumption data to train a neural network model including N hidden layers as the power consumption prediction model, where N is at least 2 A positive integer; at least one hidden layer of the N hidden layers included in the power consumption prediction model is updated with the second power consumption data.
  • the first power consumption data is actual power consumption data in a first time interval
  • the second power consumption data is power consumption simulation data in a second time interval
  • the The first power consumption data is the actual power consumption data in the third time interval
  • the second power consumption data is the actual power consumption data in the fourth time interval
  • the first power consumption data is the power consumption simulation data in the fifth time interval
  • the second power consumption data is the actual power consumption data in the sixth time interval
  • the first power consumption data is in the seventh time interval
  • the second power consumption data is the power consumption simulation data in the eighth time interval
  • the first power consumption data is the power consumption simulation data in the ninth time interval.
  • the combined data of the data and the actual power consumption data in the tenth time interval, the second power consumption data is the actual power consumption data in the eleventh time interval; or, the first power consumption data is The combined data including the power consumption simulation data in the twelfth time interval and the actual power consumption data in the thirteenth time interval, the second power consumption data is the power consumption simulation in the fourteenth time interval or, the first power consumption data is the actual power consumption data in the fifteenth time interval, and the second power consumption data includes the power consumption simulation data in the sixteenth time interval and the first power consumption data.
  • the combined data of the actual data in the seventeenth time interval; or, the first power consumption data is the simulation data in the eighteenth time interval, and the second power consumption data is included in the nineteenth time interval.
  • the combined data of the power consumption simulation data and the actual power consumption data in the twentieth time interval is included in the nineteenth time interval.
  • using the second power consumption data to update at least one hidden layer of the N hidden layers included in the power consumption prediction model includes: using the second power consumption data Train the prediction model, wherein the predetermined M hidden layers in the prediction model are fixed, and the remaining hidden layers in the prediction model except 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.
  • FIG. 11 is a configuration diagram of a power consumption prediction device of an HVAC system according to an embodiment of the present invention.
  • the power consumption prediction device 80 of the HVAC system includes:
  • a receiving module 81 configured to receive the predicted time
  • a prediction module 82 configured to generate a predicted value of power consumption corresponding to the predicted time based on a predicted model of power consumption of the HVAC system; wherein the method for generating the predicted model of power consumption includes: acquiring a first value of the HVAC system electricity consumption data; 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; using the first electricity consumption data Train a neural network model including N hidden layers as the electricity consumption prediction model, where N is a positive integer that is at least 2; update the electricity consumption prediction model included in the electricity consumption using the second electricity consumption data At least one hidden layer of the N hidden layers in .
  • data enrichment can be achieved for buildings with little data. For example, if the energy consumption of an HVAC system has a seasonal pattern (eg, high cooling demand in summer and low cooling demand in winter), how to predict electricity consumption in summer using only available winter data.
  • transfer learning can also be implemented between similar buildings, or from one type of building (such as an office building) to another type of building (such as a shopping mall), or from Shift from one season (eg winter) to another (eg summer).
  • the following is an example to describe and evaluate the electricity consumption prediction method of HVAC system.
  • Data comes from cooling systems in LCD manufacturing plants. It has 33 data characteristics, including outdoor temperature, humidity, secondary chilled water circuit operating parameters (such as supply and return water temperature) and pump flow, among others.
  • the response variable is the total power consumption of the cooling system.
  • Four months of historical data (March, April, June and August 2017) are available.
  • the goal is to predict electricity consumption in June 2017 (sample size of 584) and August 2017 (sample size of 528).
  • the KPI used is: Root Mean Square Error (Mean Absolution Percentage Error, MAPE):
  • Table 1 is a schematic diagram of training prediction models in various ways.
  • the standard method of training does not use transfer learning, that is, directly uses the training data to train the prediction model.
  • transfer learning the prediction model is first trained using pre-training data (data for pretraining), and then the hidden layer in the prediction model is updated using the training data.
  • FIG. 12 is a schematic diagram of prediction of HVAC power consumption according to an embodiment of the present invention.
  • the abscissa is time, and the ordinate is power consumption.
  • the curve 61 is the actual power consumption in winter; the curve 62 is the simulated power consumption in summer; the curve 63 is the actual power consumption in summer; the curve 64 is the predicted power consumption using the embodiment of the invention; the curve 65 is the standard method forecasted electricity consumption.
  • FIG. 13 is a configuration diagram of an HVAC power consumption prediction device according to an embodiment of the present invention.
  • the HVAC power consumption prediction device 30 includes:
  • the memory 32 is used to store the electricity consumption prediction model of the heating HVAC system, wherein the method for generating the electricity consumption prediction model includes: acquiring first electricity consumption data of the HVAC system; acquiring the first electricity consumption data of the HVAC system. Second power consumption data, wherein the data type of the second power consumption data is different from the data type of the first power consumption data; using the first power consumption data to train a neural network model including multiple hidden layers to obtaining an electricity consumption prediction model of the HVAC system; training the electricity consumption prediction model using the second electricity consumption data to update at least one hidden layer of the HVAC system;
  • the processor 33 is coupled to the interface 31 and the storage 32 via the bus 34, respectively, and is configured to generate a predicted power consumption value corresponding to the predicted time based on the power consumption prediction model.
  • 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 device 50 includes:
  • the first data acquisition module 51 is used to acquire the first data of the analysis object
  • the second data acquisition module 52 is configured to acquire the 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;
  • a training module 53 configured to use the first data to train a neural network model comprising N hidden layers as a prediction model of the analysis object, where N is a positive integer at least 2;
  • the updating module 54 is configured to use the second data to update at least one hidden layer of the N hidden layers included in the prediction model.
  • the first data is actual data in a first time interval
  • the second data is simulation data in a second time interval
  • the first data is in a third time interval
  • Actual data the second data is the actual data in the fourth time interval
  • the first data is the simulation data in the fifth time interval
  • the second data is the actual data in the sixth time interval
  • the first data is the simulation data in the seventh time interval
  • the second data is the simulation data in the eighth time interval
  • the first data is the simulation data in the ninth time interval and
  • the second data is the actual data in the eleventh time interval
  • the first data includes the simulation data in the twelfth time interval and the thirteenth time interval
  • the combined data of the actual data in the time interval, the second data is the simulation data in the fourteenth time interval; or, the first data is the actual data in the fifteenth time interval, and the second data is The combined data including the simulation data in the sixteenth time interval and the actual data in the seventeenth time interval
  • the actual data and the simulated data in the combined data have data indicators that can be superimposed on time attributes or data indicators that are not superimposed on time attributes.
  • the apparatus 50 further includes: a simulation data acquisition module 55, configured to establish a simulation model of the analysis object based on predetermined analysis object metadata; and generate the simulation data based on the simulation model.
  • a simulation data acquisition module 55 configured to establish a simulation model of the analysis object based on predetermined analysis object metadata; and generate the simulation data based on the simulation model.
  • the analysis object is an HVAC system
  • the prediction model is an electricity consumption prediction model
  • the apparatus 50 further includes: a receiving module 56 for receiving a prediction time; a prediction module 57 for The amount prediction model generates a predicted value of electric power consumption corresponding to the predicted time.
  • the updating module 54 is configured to use the second data to train the prediction model, wherein the predetermined M hidden layers in the prediction model are fixed, and the M hidden layers in the prediction model are updated, except The remaining hidden layers other than the M hidden layers, where M is a positive integer of at least 2, and M is less than or equal to N.
  • FIG. 15 is an exemplary structural diagram of an apparatus for generating a prediction model of an analysis object with a memory-processor architecture according to an embodiment of the present invention.
  • the apparatus 70 for generating the prediction model of the analysis object includes a memory 72 and a processor 71; the memory 72 stores an application program executable by the processor 71 for causing the processor 71 to execute any one of the above The described method for generating a predictive model of an analysis object.
  • the hardware modules in various embodiments may be implemented mechanically or electronically.
  • a hardware module may include specially designed permanent circuits or logic devices (eg, special purpose processors, such as FPGAs or ASICs) for performing specific operations.
  • Hardware modules may also include programmable logic devices or circuits (eg, including general-purpose processors or other programmable processors) temporarily configured by software for performing particular operations.
  • programmable logic devices or circuits eg, including general-purpose processors or other programmable processors
  • the present invention also provides a machine-readable storage medium storing instructions for causing a machine to perform a method as described herein.
  • a system or device equipped with a storage medium on which software program codes for realizing the functions of any one of the above-described embodiments are stored, and make the computer (or CPU or MPU of the system or device) ) to read and execute the program code stored in the storage medium.
  • a part or all of the actual operation can also be completed by an operating system or the like operating on the computer based on the instructions of the program code.
  • the program code read from the storage medium can also be written into the memory provided in the expansion board inserted into the computer or into the memory provided in the expansion unit connected to the computer, and then the instructions based on the program code make the device installed in the computer.
  • the CPU on the expansion board or the expansion unit or the like performs part and all of the actual operations, thereby realizing the functions of any one of the above-mentioned embodiments.
  • Embodiments of storage media for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (eg, CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), Magnetic tapes, non-volatile memory cards and ROMs.
  • the program code may be downloaded from a server computer or cloud over a communications network.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Image Analysis (AREA)
  • Feedback Control In General (AREA)
  • Testing And Monitoring For Control Systems (AREA)

Abstract

一种生成分析对象的预测模型的方法、装置和存储介质,所述方法(100)包括:获取(101)分析对象的第一数据(201);获取(102)所述分析对象的第二数据(204),其中第二数据(204)与第一数据(201)表征分析对象的相同物理量,第二数据(204)的数据类型与第一数据(201)的数据类型不相同;利用第一数据(201)将包含N个隐含层的神经网络模型(202)训练(103)为分析对象的预测模型(203),其中N为至少为2的正整数;利用第二数据(204)更新(104)包含在预测模型(203)中的N个隐含层中的至少一个隐含层。该方法采用多种数据训练模型,可以减少模型对单一数据的依赖,提高模型性能,并实现迁移学习和加快模型训练过程。

Description

生成分析对象的预测模型的方法、装置和存储介质 技术领域
本发明涉及人工智能(Artificial Intelligence,AI)技术领域,尤其涉及生成分析对象的预测模型的方法、装置和存储介质。
背景技术
机器学习(Machine Learning,ML)是一种实现AI的方法。机器学习与模式识别、计算统计、人工智能等领域有着密切的联系。机器学习可以使用机器(计算机和软件)从已知数据中挖掘意义,从而给予机器学习环境的能力。机器学习算法可以包括监督学习(如分类问题)、无监督学习(如聚类问题)、半监督学习、集成学习、深度学习和强化学习,等等。
预测分析是机器学习在业务问题中的一个广泛应用。利用训练数据生成预测模型的过程即为模型训练过程。使用预测模型可以预测当输入改变后输出的变化过程。
预测模型的准确度通常严重依赖于数据可用性。目前,通常使用单一的训练数据(比如,历史数据)以训练预测模型。然而单一的训练数据通常数据量不足,难以训练出性能良好的预测模型。
发明内容
本发明实施方式提出生成分析对象的预测模型的方法、装置和存储介质。
第一方面,生成分析对象的预测模型的方法,包括:
获取分析对象的第一数据;
获取所述分析对象的第二数据,其中所述第二数据与所述第一数据表征所述分析对象的相同物理量,所述第二数据的数据类型与所述第一数据的数据类型不相同;
利用所述第一数据将包含N个隐含层的神经网络模型训练为所述分析对象的预测模型,其中N为至少为2的正整数;
利用所述第二数据更新包含在所述预测模型中的所述N个隐含层中的至少一个隐含层。
第二方面,提供生成分析对象的预测模型的装置,包括:
第一数据获取模块,用于获取分析对象的第一数据;
第二数据获取模块,用于获取所述分析对象的第二数据,其中所述第二数据与所述第一数据表征所述分析对象的相同物理量,所述第二数据的数据类型与所述第一数据的数据类型 不相同;
训练模块,用于利用所述第一数据将包含N个隐含层的神经网络模型训练为所述分析对象的预测模型,其中N为至少为2的正整数;
更新模块,用于利用所述第二数据更新包含在所述预测模型中的所述N个隐含层中的至少一个隐含层。
第三方面,提供生成分析对象的预测模型的装置,包括处理器和存储器;
所述存储器中存储有可被所述处理器执行的应用程序,用于使得所述处理器执行如上任一项所述的生成分析对象的预测模型的方法。
第四方面,提供计算机可读存储介质,其中存储有计算机可读指令,该计算机可读指令用于执行如上任一项所述的生成分析对象的预测模型的方法。
可见,本发明实施方式可以采用多种数据类型的多个数据生成分析对象的训练模型,可以减少模型对单一类型数据的依赖。而且,由于第二数据的数据特征与第一数据表征分析对象的相同物理量,因此可以利用第二数据更新预测模型中的部分隐含层,而不是重新训练整个预测模型,因此还加快了模型训练过程,实现了模型的迁移学习。
对于上述任一方面,优选地,所述第二数据的数据类型与所述第一数据的数据类型不相同,包括:
所述第一数据为第一时间区间内的实际数据,所述第二数据为第二时间区间内的仿真数据;或
所述第一数据为第三时间区间内的实际数据,所述第二数据为第四时间区间内的实际数据;或
所述第一数据为第五时间区间内的仿真数据,所述第二数据为第六时间区间内的实际数据;或
所述第一数据为第七时间区间内的仿真数据,所述第二数据为第八时间区间内的仿真数据;或
所述第一数据为包含第九时间区间内的仿真数据和第十时间区间内的实际数据的组合数据,所述第二数据为第十一时间区间内的实际数据;或
所述第一数据为包含第十二时间区间内的仿真数据和第十三时间区间内的实际数据的组合数据,所述第二数据为第十四时间区间内的仿真数据;或
所述第一数据为第十五时间区间内的实际数据,所述第二数据为包含第十六时间区间内的仿真数据和第十七时间区间内的实际数据的组合数据;或
所述第一数据为第十八时间区间内的仿真数据,所述第二数据为包含第十九时间区间内的仿真数据和第二十时间区间内的实际数据的组合数据。
优选的,所述组合数据中的实际数据和仿真数据具有时间属性上可叠加的数据指标或在时间属性上不可叠加的数据指标。
可见,本发明实施方式的第一数据和第二数据具有多种类型,丰富了训练数据,还提高了模型准确度。
对于上述任一方面,优选地,还包括:
基于预定的分析对象元数据建立所述分析对象的仿真模型;
基于所述仿真模型生成所述仿真数据。
因此,通过仿真模型可以快速获取仿真数据,提高了数据获取效率。
对于上述任一方面,优选地,分析对象为供热通风与空气调节(HVAC)系统,所述预测模型为用电量预测模型;还包括:接收预测时间;基于更新后的所述预测模型生成对应于所述预测时间的用电量预测值。
因此,可以将更新后的所述预测模型应用到HVAC系统中。
对于上述任一方面,优选地,所述利用第二数据更新包含在所述预测模型中的所述N个隐含层中的至少一个隐含层包括:
利用所述第二数据训练所述预测模型,其中固定所述预测模型中的、预定的M个隐含层,更新所述预测模型中的、除所述M个隐含层之外的剩余隐含层,其中M为至少为2的正整数,且M小于等于N。
因此,通过固定预测模型中的M个隐含层,可以将这M个隐含层作为成熟知识而保留,从而实现知识迁移,还降低了迁移后的训练工作量。
附图说明
图1为本发明实施方式的生成分析对象的预测模型的方法的示范性流程图。
图2为本发明实施方式的生成分析对象的预测模型的方法的第一示范性流程图。
图3为本发明实施方式的生成分析对象的预测模型的方法的第二示范性流程图。
图4为本发明实施方式的生成分析对象的预测模型的方法的第三示范性流程图。
图5为本发明实施方式的生成分析对象的预测模型的方法的第四示范性流程图。
图6为本发明实施方式的生成分析对象的预测模型的方法的第五示范性流程图。
图7为本发明实施方式的生成分析对象的预测模型的方法的第六示范性流程图。
图8为本发明实施方式的生成分析对象的预测模型的方法的第七示范性流程图。
图9为本发明实施方式的生成分析对象的预测模型的方法的第八示范性流程图。
图10为本发明实施方式的HVAC系统的用电量预测方法的流程图。
图11为本发明实施方式的HVAC系统的用电量预测装置的结构图。
图12为本发明实施方式的HVAC用电量的预测示意图。
图13为本发明实施方式的HVAC用电量的预测装置的结构图。
图14为本发明实施方式的生成分析对象的预测模型的装置的示范性结构图。
图15为本发明实施方式的具有存储器-处理器架构的、生成分析对象的预测模型的装置的示范性结构图。
其中,附图标记如下:
标号 含义
100 生成分析对象的预测模型的方法
201,301,401,501,601,701,801,901 第一数据
202,302,402,502,602,702,802,902 神经网络模型
203,303,403,503,603,703,803,903 预测模型
204,304,404,504,604,704,804,904 第二数据
205,305,405,505,605,705,805,905 更新后的预测模型
1000 HVAC系统的用电量预测方法
1001~1002 步骤
80 HVAC系统的用电量预测装置
81 接收模块
82 预测模块
61 冬天的实际用电量
62 夏天的仿真用电量
63 夏天的实际用电量
64 本发明实施方式的预测用电量
65 标准方法的预测用电量
30 HVAC用电量的预测装置
31 接口
32 存储器
33 处理器
34 总线
50 分析对象的预测模型的装置
51 第一数据获取模块
52 第二数据获取模块
53 训练模块
54 更新模块
55 仿真数据获取模块
56 接收模块
57 预测模块
70 分析对象的预测模型的装置
71 处理器
72 存储器
具体实施方式
为了使本发明的技术方案及优点更加清楚明白,以下结合附图及实施方式,对本发明进行进一步详细说明。应当理解,此处所描述的具体实施方式仅仅用以阐述性说明本发明,并不被配置为用于限定本发明的保护范围。
为了描述上的简洁和直观,下文通过描述若干代表性的实施方式来对本发明的方案进行阐述。实施方式中大量的细节仅被配置为用于帮助理解本发明的方案。但是很明显,本发明的技术方案实现时可以不局限于这些细节。为了避免不必要地模糊了本发明的方案,一些实施方式没有进行细致地描述,而是仅给出了框架。下文中,“包括”是指“包括但不限于”,“根据……”是指“至少根据……,但不限于仅根据……”。由于汉语的语言习惯,下文中没有特别指出一个成分的数量时,意味着该成分可以是一个也可以是多个,或可理解为至少一个。
考虑到采用单一的训练数据训练预测模型的缺陷,本发明实施方式利用第一数据训练包含隐含层的预测模型,再利用第二数据更新预测模型中的部分隐含层,可以丰富训练数据,并提高模型性能。而且,还可以解决预测模型的可迁移性问题,并加快模型训练过程。
图1为本发明实施方式的生成分析对象的预测模型的方法的示范性流程图。
如图1所示,该方法包括:
步骤101:获取分析对象的第一数据。
在这里,分析对象为需要被分析的对象,比如可以实施为热力系统或电力系统,等等。优选的,分析对象实施为供热通风与空气调节(HVAC)系统。
第一数据的数据类型可以由第一数据的数据获取方式或数据来源所决定。比如,第一数据的数据类型可以包括:
类型(1):分析对象的实际数据(比如,历史数据);
类型(2):分析对象的仿真数据;
类型(3):分析对象的、包含实际数据和仿真数据的组合数据。
第一数据表征分析对象的一种预定物理量。比如,第一数据可以用于表征分析对象的用电量、风扇使用时间、冷媒使用量,等等。
步骤102:获取分析对象的第二数据,其中第二数据与第一数据表征分析对象的相同物理量,第二数据的数据类型与第一数据的数据类型不相同。
类似地,第二数据的数据类型可以由第二数据的数据获取方式或数据来源所决定。比如,第二数据的数据类型可以包括:
类型(1):分析对象的实际数据(比如,历史数据);
类型(2):分析对象的仿真数据;
类型(3):分析对象的、包含实际数据和仿真数据的组合数据。
第二数据的数据类型与第一数据的数据类型不相同。因此,当第一数据的数据类型为类型(1)时,则第二数据的数据类型为类型(2)或类型(3);当第一数据的数据类型为类型(2)时,则第二数据的数据类型为类型(1)或类型(3);当第一数据的数据类型为类型(3)时,则第二数据的数据类型为类型(1)或类型(2)。
第二数据与第一数据表征分析对象的相同物理量。比如,当第一数据表征分析对象的用电量时,第二数据同样表征分析对象的用电量;当第一数据表征分析对象的风扇使用时间时,第二数据同样表征分析对象的风扇使用时间,等等。
优选地,相比较第一数据而言,第二数据与最终得到的预测模型的预测任务更加相关。
在一个实施方式中,第一数据为第一时间区间内的实际数据,第二数据为第二时间区间内的仿真数据。第一时间区间和第二时间区间都可以通过年、季度、月、星期或日等时间计量单位进行描述。比如,第一数据为包含分析对象(比如HVAC系统)的工厂A在2012年3月份的真实用电量数据;第二数据为包含该分析对象(比如HVAC系统)的工厂A在2012年4月份的仿真用电量数据,或包含该分析对象(比如HVAC系统)的工厂B在2012年4 月份的仿真用电量数据。
在一个实施方式中,第一数据为第三时间区间内的实际数据,第二数据为第四时间区间内的实际数据。第三时间区间和第四时间区间都可以通过年、季度、月、星期或日等时间计量单位进行描述。比如,第一数据为包含分析对象(比如HVAC系统)的工厂A在2012年3月份的真实用电量数据;第二数据为包含分析对象(比如HVAC系统)的工厂A在2012年4月份的真实用电量数据,或包含分析对象(比如HVAC系统)的工厂B在2012年4月份的真实用电量数据。
在一个实施方式中,第一数据为第五时间区间内的仿真数据,第二数据为第六时间区间内的实际数据。第五时间区间和第六时间区间都可以通过年、季度、月、星期或日等时间计量单位进行描述。比如,第一数据为包含分析对象(比如HVAC系统)的工厂A在2012年3月份的仿真用电量数据;第二数据为包含分析对象(比如HVAC系统)的工厂A在2012年4月份的真实用电量数据,或包含分析对象(比如HVAC系统)的工厂B在2012年4月份的真实用电量数据。
在一个实施方式中,第一数据为第七时间区间内的仿真数据,第二数据为第八时间区间内的仿真数据。第七时间区间和第八时间区间都可以通过年、季度、月、星期或日等时间计量单位进行描述。比如,第一数据为包含分析对象(比如HVAC系统)的工厂A在2012年3月份的仿真用电量数据;第二数据为包含分析对象(比如HVAC系统)的工厂A在2012年4月份的仿真用电量数据,或包含分析对象(比如HVAC系统)的工厂B在2012年4月份的仿真用电量数据。
在一个实施方式中,第一数据为包含第九时间区间内的仿真数据和第十时间区间内的实际数据的组合数据,第二数据为第十一时间区间内的实际数据。第九时间区间、第十时间区间和第十一时间区间都可以通过年、季度、月、星期或日等时间计量单位进行描述。比如,第一数据为包含分析对象(比如HVAC系统)的工厂A在2012年4月份的仿真用电量数据和2012年3月份的真实用电量数据的组合数据;第二数据为包含分析对象(比如HVAC系统)的工厂A在2012年2月份的真实用电量数据,或包含分析对象(比如HVAC系统)的工厂B在2012年2月份的真实用电量数据。
在一个实施方式中,第一数据为包含第十二时间区间内的仿真数据和第十三时间区间内的实际数据的组合数据,第二数据为第十四时间区间内的仿真数据。第十二时间区间、第十三时间区间和第十四时间区间都可以通过年、季度、月、星期或日等时间计量单位进行描述。比如,第一数据为包含分析对象(比如HVAC系统)的工厂A在2012年4月份的仿真用电 量数据和2012年3月份的真实用电量数据的组合数据;第二数据为包含分析对象(比如HVAC系统)的工厂A在2012年5月份的仿真用电量数据,或包含分析对象(比如HVAC系统)的工厂B在2012年5月份的仿真用电量数据。
在一个实施方式中,第一数据为第十五时间区间内的实际数据,第二数据为包含第十六时间区间内的仿真数据和第十七时间区间内的实际数据的组合数据。第十五时间区间、第十六时间区间和第十七时间区间都可以通过年、季度、月、星期或日等时间计量单位进行描述。比如,第一数据为包含分析对象(比如HVAC系统)的工厂A在2012年5月份的真实用电量数据,第二数据为包含分析对象(比如HVAC系统)的工厂A在2012年4月份的仿真用电量数据和2012年3月份的真实用电量数据的组合数据,或包含分析对象(比如HVAC系统)的工厂B在2012年4月份的仿真用电量数据和2012年3月份的真实用电量数据的组合数据。
在一个实施方式中,第一数据为第十八时间区间内的仿真数据,所述第二数据为包含第十九时间区间内的仿真数据和第二十时间区间内的实际数据的组合数据。第十八时间区间、第十九时间区间和第二十时间区间都可以通过年、季度、月、星期或日等时间计量单位进行描述。比如,第一数据为包含分析对象(比如HVAC系统)的工厂A在2012年5月份的仿真用电量数据;第二数据为包含分析对象(比如HVAC系统)的工厂A在2012年4月份的仿真用电量数据和2012年3月份的真实用电量数据的组合数据,或包含分析对象(比如HVAC系统)的工厂B在2012年4月份的仿真用电量数据和2012年3月份的真实用电量数据的组合数据。
在一个实施方式中,组合数据中的实际数据和仿真数据具有时间属性上可叠加的数据指标。
比如,组合数据中的实际数据和仿真数据分别实施为在可叠加月份上的用电量数据。举例,实际数据包括2011年1月份的用电量数据和2011年2月份的用电量数据,仿真数据为2011年3月份的用电量数据,则组合数据为2011年第一季度(包括1月份~3月份)的用电量数据。
在一个实施方式中,组合数据中的实际数据和仿真数据具有在时间属性上不可叠加的数据指标。
比如,实际数据为2011年1月份的用电量数据,仿真数据为2011年1月份的房间温度值。再比如,实际数据为2011年1月份的用电量数据,仿真数据为2011年2月份的风扇使用时间。
以上示范性描述了第一数据和第二数据的典型实例,本领域技术人员可以意识到,这种描述仅是示范性的,并不用于限定本发明实施方式的保护范围。
可以利用分析对象的仿真模型获取仿真数据。具体包括:基于预定的分析对象元数据(比如,建筑信息模型、设计图纸等)建立分析对象的仿真模型;基于仿真模型生成仿真数据。
步骤103:利用第一数据将包含N个隐含层的神经网络模型训练为分析对象的预测模型,其中N为至少为2的正整数。
具体地,神经网络模型可以实施为:前馈神经网络模型、径向基神经网络模型、长短期记忆(LSTM)网络模型、回声状态网络(ESN)、门循环单元(GRU)网络模型或深度残差网络模型,等等。优选地,神经网络模型实施为LSTM网络模型。
在这里,利用第一数据将包含N个隐含层的神经网络模型训练为分析对象的预测模型。基于步骤103的训练步骤,可以确定出各个隐含层的参数(比如,权重)。
步骤104:利用第二数据更新包含在预测模型中的N个隐含层中的至少一个隐含层。
步骤103中利用第一数据训练出的分析对象的预测模型,即可用于对应用到各个场景中的分析对象执行预测分析。
在一个实施方式中,当第二数据的应用场景与第一数据的应用场景相同时,优选利用第二数据对第一数据训练出的分析对象的预测模型中的、预定的部分隐含层进行更新,从而提高预测模型的准确度。在这里,可以基于更新和不更新某特定隐含层的预测效果对比,确定出是否将该特定隐含层作为需要被更新的隐含层。比如,当更新和不更新某特定隐含层的预测效果有显著提升(比如,预测准确率的提升幅度大于预定门限值)时,则该特定隐含层作为需要被更新的隐含层;当更新和不更新某特定隐含层的预测效果没有显著提升(比如,预测准确率的提升幅度小于等于预定门限值)时,则该特定隐含层不作为需要被更新的隐含层。
实际上,由于第一数据通常源自特定场景,而分析对象可能会应用到与该特定场景具有相关性的另外应用场景中,因此优选利用源自该另外应用场景的第二数据对第一数据训练出的分析对象的预测模型中的、与该第二数据的应用场景具有关联性的隐含层进行更新,从而使得该更新后的预测模型相比较更新前的预测模型更加适用于该另外应用场景,从而实现知识迁移。在这里,与第二数据的应用场景具有关联性的隐含层的确定方式包括:
(1)、基于对该第二数据的应用场景的模型分析,从理论上推导出与该第二数据的应用场景具有关联性的隐含层。
(2)、基于更新和不更新某特定隐含层的预测效果对比,确定该特定隐含层是否与第二数据的应用场景具有关联性。比如,当更新和不更新某特定隐含层的预测效果有显著提升(比 如,预测准确率的提升幅度大于预定门限值)时,则该特定隐含层与第二数据的应用场景具有关联性,因此作为需要被更新的隐含层;当更新和不更新某特定隐含层的预测效果没有显著提升(比如,预测准确率的提升幅度小于等于预定门限值)时,则该特定隐含层与第二数据的应用场景不具有关联性,因此不作为需要被更新的隐含层。
在一个实施方式中,利用第二数据更新包含在预测模型中的N个隐含层中的至少一个隐含层包括:利用第二数据训练预测模型,其中固定预测模型中的、预定的M个隐含层,更新预测模型中的、除M个隐含层之外的剩余隐含层,其中M为至少为2的正整数,且M小于等于N。
下面描述生成分析对象的预测模型的典型过程。
图2为本发明实施方式的生成分析对象的预测模型的方法的第一示范性流程图。
在图2的一个实施方式中,分析对象的第一数据201为:在包含分析对象的第一应用场景中所采集的、分析对象在第一时间区间内的实际数据;分析对象的第二数据204为:在包含分析对象的第一应用场景中所采集的、分析对象在第二时间区间内的仿真数据。其中,第一时间区间和第二时间区间可以相同,也可以不同,优选为不同。
比如,预测任务为包含分析对象(比如HVAC系统)的第一应用场景(比如工厂A)中的用电量预测;第一数据201为工厂A中2011年1月份的风扇实际使用时间;第二数据204为工厂A中2011年2月份的风扇仿真使用时间。
首先,将第一数据201作为训练数据输入到包含N个隐含层的神经网络模型202中,经过训练过程,可以将该神经网络模型202训练为分析对象的预测模型203。在该预测模型203中,N个隐含层的参数都得到确定。
然后,将第二数据204作为训练数据输入到预测模型203中以再次执行训练。在该再次训练过程中:将预测模型203的N个隐含层中的、预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型205。此时,可以利用该更新后的预测模型205对工厂A中的分析对象的用电量执行预测。
在图2的另一个实施方式中,分析对象的第一数据201为:在包含分析对象的第一应用场景中所采集的、分析对象在第一时间区间内的实际数据;分析对象的第二数据204为:在包含分析对象的第二应用场景中所采集的、分析对象在第二时间区间内的仿真数据。其中,第一时间区间和第二时间区间可以相同,也可以不同,优选为不同。相比较第一数据201而言,第二数据204与第二应用场景中的预测任务更加相关。
比如,预测任务为包含分析对象(比如HVAC系统)的第二应用场景(比如工厂B)中的用电量预测;第一数据201为工厂A中2011年2月份的风扇使用时间;第二数据204为工厂B中2011年1月份的风扇仿真使用时间。
首先,将第一数据201作为训练数据输入到包含N个隐含层的神经网络模型202中,经过训练过程,可以将该神经网络模型202训练为分析对象的预测模型203。在该预测模型203中,N个隐含层的参数都得到确定。
然后,将第二数据204作为训练数据输入到预测模型203中以再次执行训练。在该再次训练过程中:将预测模型203的N个隐含层中的、与第二数据204的应用场景不相关的预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、与第二数据204的应用场景相关的(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型205。此时,可以利用该更新后的预测模型205对工厂B中的分析对象的用电量执行预测。
图3为本发明实施方式的生成分析对象的预测模型的方法的第二示范性流程图。
在图3的一个实施方式中,分析对象的第一数据301为:在包含分析对象的第一应用场景中所采集的、分析对象在第三时间区间内的实际数据;分析对象的第二数据304为:在包含分析对象的第一应用场景中所采集的、分析对象在第四时间区间内的实际数据。其中,第三时间区间和第四时间区间可以相同或不同。
比如,预测任务为包含分析对象(比如HVAC系统)的第一应用场景(比如工厂A)中的用电量预测;第一数据301为工厂A中2011年2月份的用电量实际数据;第二数据304为工厂A中2011年1月份的用电量实际数据。
首先,将第一数据301作为训练数据输入到包含N个隐含层的神经网络模型302中,经过训练过程,可以将该神经网络模型302训练为分析对象的预测模型303。在该预测模型303中,N个隐含层的参数都得到确定。
然后,将第二数据304作为训练数据输入到预测模型303中以再次执行训练。在该再次训练过程中:将预测模型303的N个隐含层中的、预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型305。此时,可以利用该更新后的预测模型305对工厂A中的分析对象的用电量执行预测。
在图3的另一个实施方式中,分析对象的第一数据301为:在包含分析对象的第一应用场景中所采集的、分析对象在第三时间区间内的实际数据;分析对象的第二数据304为:在 包含分析对象的第二应用场景中所采集的、分析对象在第四时间区间内的实际数据。其中,第三时间区间和第四时间区间可以相同或不同。相比较第一数据301而言,第二数据304与第二应用场景中的预测任务更加相关。
比如,预测任务为包含分析对象(比如HVAC系统)的第二应用场景(比如工厂B)中的用电量预测;第一数据301为工厂A中2011年2月份的用电量实际数据;第二数据304为工厂B中2011年1月份的用电量实际数据。
首先,将第一数据301作为训练数据输入到包含N个隐含层的神经网络模型302中,经过训练过程,可以将该神经网络模型302训练为分析对象的预测模型303。在该预测模型303中,N个隐含层的参数都得到确定。
然后,将第二数据304作为训练数据输入到预测模型303中以再次执行训练。在该再次训练过程中:将预测模型303的N个隐含层中的、与第二数据304的应用场景不相关的预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、与第二数据304的应用场景相关的(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型305。此时,可以利用该更新后的预测模型305对工厂B中的分析对象的用电量执行预测。
图4为本发明实施方式的生成分析对象的预测模型的方法的第三示范性流程图。
在图4中,分析对象的第一数据401为:在包含分析对象的第一应用场景中所采集的、分析对象在第五时间区间内的仿真数据;分析对象的第二数据404为:在包含分析对象的第一应用场景中所采集的、分析对象在第六时间区间内的实际数据。其中,第五时间区间和第六时间区间可以相同,也可以不同,优选为不同。
比如,预测任务为包含分析对象(比如HVAC系统)的第一应用场景(比如工厂A)中的用电量预测;第一数据401为工厂A中2011年2月份的风扇仿真使用时间;第二数据404为工厂A中2011年1月份的风扇实际使用时间。
首先,将第一数据401作为训练数据输入到包含N个隐含层的神经网络模型402中,经过训练过程,可以将该神经网络模型402训练为分析对象的预测模型403。在该预测模型403中,N个隐含层的参数都得到确定。
然后,将第二数据404作为训练数据输入到预测模型403中以再次执行训练。在该再次训练过程中:将预测模型403的N个隐含层中的、预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型405。此时,可以利用该更新后的预测模型405对工 厂A中的分析对象的用电量执行预测。
在图4的另一个实施方式中,分析对象的第一数据401为:在包含分析对象的第一应用场景中所采集的、分析对象在第五时间区间内的仿真数据;分析对象的第二数据404为:在包含分析对象的第二应用场景中所采集的、分析对象在第六时间区间内的实际数据。其中,第五时间区间和第六时间区间可以相同或不同。相比较第一数据401而言,第二数据404与第二应用场景中的预测任务更加相关。
比如,预测任务为包含分析对象(比如HVAC系统)的第二应用场景(比如工厂B)中的用电量预测;第一数据401为工厂A中2011年2月份的风扇仿真使用时间;第二数据404为工厂B中2011年1月份的风扇实际使用时间。
首先,将第一数据401作为训练数据输入到包含N个隐含层的神经网络模型402中,经过训练过程,可以将该神经网络模型402训练为分析对象的预测模型403。在该预测模型403中,N个隐含层的参数都得到确定。
然后,将第二数据404作为训练数据输入到预测模型403中以再次执行训练。在该再次训练过程中:将预测模型403的N个隐含层中的、与第二数据404的应用场景不相关的M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、与第二数据404的应用场景相关的(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型405。此时,可以利用该更新后的预测模型405对工厂B中的分析对象的用电量执行预测。
图5为本发明实施方式的生成分析对象的预测模型的方法的第四示范性流程图。
在图5中,分析对象的第一数据501为:在包含分析对象的第一应用场景中所采集的、分析对象在第七时间区间内的仿真数据;分析对象的第二数据504为:在包含分析对象的第一应用场景中所采集的、分析对象在第八时间区间内的仿真数据。其中,第七时间区间和第八时间区间可以相同或不同。
比如,预测任务为包含分析对象(比如HVAC系统)的第一应用场景(比如工厂A)中的用电量预测;第一数据501为工厂A中2011年2月份的用电量仿真数据;第二数据404为工厂A中2011年1月份的用电量仿真数据。
首先,将第一数据501作为训练数据输入到包含N个隐含层的神经网络模型502中,经过训练过程,可以将该神经网络模型502训练为分析对象的预测模型503。在该预测模型503中,N个隐含层的参数都得到确定。
然后,将第二数据504作为训练数据输入到预测模型503中以再次执行训练。在该再次 训练过程中:将预测模型503的N个隐含层中的、预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型505。此时,可以利用该更新后的预测模型505对工厂A中的分析对象的用电量执行预测。
在图5的另一个实施方式中,分析对象的第一数据501为:在包含分析对象的第一应用场景中所采集的、分析对象在第七时间区间内的仿真数据;分析对象的第二数据504为:在包含分析对象的第二应用场景中所采集的、分析对象在第八时间区间内的仿真数据。其中,第七时间区间和第八时间区间可以相同或不同。相比较第一数据501而言,第二数据504与第二应用场景中的预测任务更加相关。
比如,预测任务为包含分析对象(比如HVAC系统)的第二应用场景(比如工厂B)中的用电量预测;第一数据501为工厂A中2011年2月份的用电量仿真数据;第二数据504为工厂B中2011年1月份的用电量仿真数据。
首先,将第一数据501作为训练数据输入到包含N个隐含层的神经网络模型502中,经过训练过程,可以将该神经网络模型502训练为分析对象的预测模型503。在该预测模型503中,N个隐含层的参数都得到确定。
然后,将第二数据504作为训练数据输入到预测模型503中以再次执行训练。在该再次训练过程中:将预测模型503的N个隐含层中的、与第二数据504的应用环境不相关的预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、与第二数据504的应用环境相关的(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型505。此时,可以利用该更新后的预测模型505对工厂B中的分析对象的用电量执行预测。
图6为本发明实施方式的生成分析对象的预测模型的方法的第五示范性流程图。
在图6中,分析对象的第一数据601为:在包含分析对象的第一应用场景中所采集的、包含第九时间区间内的仿真数据和第十时间区间内的实际数据的组合数据;分析对象的第二数据604为:在包含分析对象的第一应用场景中所采集的、分析对象在第十一时间区间内的实际数据。其中,第九时间区间和第十时间区间可以相同或不同。相比较第一数据601而言,第二数据604与第一应用场景中的预测任务更加相关。
比如,预测任务为包含分析对象(比如HVAC系统)的第一应用场景(比如工厂A)中的用电量预测;第一数据601为包含工厂A中2011年2月份的仿真风扇使用时间和工厂A中2011年3月份的实际温度值的组合数据;第二数据604为工厂A中2011年1月份的风扇 实际使用时间和工厂A中2011年1月份的用电量实际数据。
首先,将第一数据601作为训练数据输入到包含N个隐含层的神经网络模型602中,经过训练过程,可以将该神经网络模型602训练为分析对象的预测模型603。在该预测模型603中,N个隐含层的参数都得到确定。
然后,将第二数据604作为训练数据输入到预测模型603中以再次执行训练。在该再次训练过程中:将预测模型603的N个隐含层中的、预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型605。此时,可以利用该更新后的预测模型605对工厂A中的分析对象的用电量执行预测。
在图6的另一个实施方式中,分析对象的第一数据601为:在包含分析对象的第一应用场景中所采集的、包含第九时间区间内的仿真数据和第十时间区间内的实际数据的组合数据;分析对象的第二数据604为:在包含分析对象的第二应用场景中所采集的、分析对象在第十一时间区间内的实际数据。其中,第九时间区间和第十时间区间可以相同或不同。相比较第一数据601而言,第二数据604与第二应用场景中的预测任务更加相关。
比如,预测任务为包含分析对象(比如HVAC系统)的第二应用场景(比如工厂B)中的用电量预测;第一数据601为包含工厂A中2011年2月份的仿真风扇使用时间和工厂A中2011年3月份的实际温度值的组合数据;第二数据604为工厂B中2011年1月份的风扇实际使用时间和工厂A中2011年1月份的用电量实际数据。
首先,将第一数据601作为训练数据输入到包含N个隐含层的神经网络模型602中,经过训练过程,可以将该神经网络模型602训练为分析对象的预测模型603。在该预测模型603中,N个隐含层的参数都得到确定。
然后,将第二数据604作为训练数据输入到预测模型603中以再次执行训练。在该再次训练过程中:将预测模型603的N个隐含层中的、与第二数据604的应用环境不相关的预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、与第二数据604的应用环境相关的(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型605。此时,可以利用该更新后的预测模型605对工厂B中的分析对象的用电量执行预测。
图7为本发明实施方式的生成分析对象的预测模型的方法的第六示范性流程图。
在图7中,分析对象的第一数据701为:在包含分析对象的第一应用场景中所采集的、包含第十二时间区间内的仿真数据和第十三时间区间内的实际数据的组合数据;分析对象的 第二数据704为:在包含分析对象的第一应用场景中所采集的、第十四时间区间内的仿真数据。其中,第十二时间区间和第十三时间区间可以相同或不同,优选不同;第十四时间区间与第十二时间区间可以相同或不同,优选不同。相比较第一数据701而言,第二数据704与第一应用场景中的预测任务更加相关。
比如,预测任务为包含分析对象(比如HVAC系统)的第一应用场景(比如工厂A)中的用电量预测;第一数据701为包含工厂A中2011年2月份的仿真风扇使用时间和工厂A中2011年3月份的实际用电量的组合数据;第二数据704为工厂A中2011年1月份的仿真风扇使用时间和2011年1月份的用电量仿真数据。
首先,将第一数据701作为训练数据输入到包含N个隐含层的神经网络模型702中,经过训练过程,可以将该神经网络模型702训练为分析对象的预测模型703。在该预测模型703中,N个隐含层的参数都得到确定。
然后,将第二数据704作为训练数据输入到预测模型703中以再次执行训练。在该再次训练过程中:将预测模型703的N个隐含层中的、预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型705。此时,可以利用该更新后的预测模型705对工厂A中的分析对象的用电量执行预测。
在图7的另一个实施方式中,分析对象的第一数据701为:在包含分析对象的第一应用场景中所采集的、包含第十二时间区间内的仿真数据和第十三时间区间内的实际数据的组合数据;分析对象的第二数据704为:在包含分析对象的第二应用场景中所采集的、分析对象在第十四时间区间内的实际数据。其中,第十二时间区间和第十三时间区间可以相同或不同,优选不同;第十四时间区间与第十二时间区间可以相同或不同,优选不同。相比较第一数据701而言,第二数据704与第二应用场景中的预测任务更加相关。
比如,预测任务为包含分析对象(比如HVAC系统)的第二应用场景(比如工厂B)中的用电量预测;第一数据701为包含工厂A中2011年2月份的仿真风扇使用时间和工厂A中2011年3月份的实际用电量的组合数据;第二数据704为工厂B中2011年1月份的仿真风扇使用时间和2011年1月份的用电量仿真数据。
首先,将第一数据701作为训练数据输入到包含N个隐含层的神经网络模型702中,经过训练过程,可以将该神经网络模型702训练为分析对象的预测模型703。在该预测模型703中,N个隐含层的参数都得到确定。
然后,将第二数据704作为训练数据输入到预测模型703中以再次执行训练。在该再次 训练过程中:将预测模型703的N个隐含层中的、与第二数据704的应用环境不相关的预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、与第二数据704的应用环境相关的(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型705。此时,可以利用该更新后的预测模型705对工厂B中的分析对象的用电量执行预测。
图8为本发明实施方式的生成分析对象的预测模型的方法的第七示范性流程图。
在图8中,分析对象的第一数据801为:在包含分析对象的第一应用场景中所采集的、第十五时间区间内的实际数据;分析对象的第二数据804为:在包含分析对象的第一应用场景中所采集的、第十六时间区间内的仿真数据和第十七时间区间内的实际数据的组合数据。其中,第十六时间区间和第十七时间区间可以相同或不同,优选不同;第十五时间区间与第十七时间区间可以相同或不同,优选不同。优选不同。
比如,预测任务为包含分析对象(比如HVAC系统)的第一应用场景(比如工厂A)中的用电量预测;第一数据801包含工厂A中2011年2月份的用电量实际数据和2011年2月份的实际温度值;第二数据804为包含工厂A中2011年1月份的用电量仿真数据和工厂A中2011年3月份的实际温度值的组合数据。
首先,将第一数据801作为训练数据输入到包含N个隐含层的神经网络模型802中,经过训练过程,可以将该神经网络模型802训练为分析对象的预测模型803。在该预测模型803中,N个隐含层的参数都得到确定。
然后,将第二数据804作为训练数据输入到预测模型803中以再次执行训练。在该再次训练过程中:将预测模型803的N个隐含层中的、预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型805。此时,可以利用该更新后的预测模型805对工厂A中的分析对象的用电量执行预测。
在图8的另一个实施方式中,分析对象的第一数据801为:在包含分析对象的第一应用场景中所采集的、第十五时间区间内的实际数据;分析对象的第二数据704为:在包含分析对象的第二应用场景中所采集的、第十六时间区间内的仿真数据和第十七时间区间内的实际数据的组合数据。第十六时间区间和第十七时间区间可以相同或不同,优选不同;第十五时间区间与第十七时间区间可以相同或不同,优选不同。优选不同。相比较第一数据801而言,第二数据804与第二应用场景中的预测任务更加相关。
比如,预测任务为包含分析对象(比如HVAC系统)的第二应用场景(比如工厂B)中 的用电量预测;第一数据801包含工厂A中2011年2月份的用电量实际数据和2011年2月份的实际温度值;第二数据804为包含工厂B中2011年1月份的用电量仿真数据和工厂B中2011年3月份的实际温度值的组合数据。
首先,将第一数据801作为训练数据输入到包含N个隐含层的神经网络模型802中,经过训练过程,可以将该神经网络模型802训练为分析对象的预测模型803。在该预测模型803中,N个隐含层的参数都得到确定。
然后,将第二数据804作为训练数据输入到预测模型803中以再次执行训练。在该再次训练过程中:将预测模型803的N个隐含层中的、与第二数据804的应用环境不相关的预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、与第二数据804的应用环境相关的(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型805。此时,可以利用该更新后的预测模型805对工厂B中的分析对象的用电量执行预测。
图9为本发明实施方式的生成分析对象的预测模型的方法的第八示范性流程图。
在图9中,分析对象的第一数据901为:在包含分析对象的第一应用场景中所采集的、第十八时间区间内的仿真数据;分析对象的第二数据904为:在包含分析对象的第一应用场景中所采集的、包含第十九时间区间内的仿真数据和第二十时间区间内的实际数据的组合数据。其中,第十九时间区间和第二十时间区间可以相同或不同,优选不同;第十九时间区间与第十八时间区间可以相同或不同,优选不同。优选不同。
比如,预测任务为包含分析对象(比如HVAC系统)的第一应用场景(比如工厂A)中的用电量预测;第一数据801包含工厂A中2011年2月份的仿真风扇使用时间和2011年2月份的用电量仿真数据;第二数据804为包含工厂A中2011年1月份的仿真风扇使用时间和工厂A中2011年3月份的用电量实际数据的组合数据。
首先,将第一数据901作为训练数据输入到包含N个隐含层的神经网络模型902中,经过训练过程,可以将该神经网络模型902训练为分析对象的预测模型903。在该预测模型903中,N个隐含层的参数都得到确定。
然后,将第二数据904作为训练数据输入到预测模型903中以再次执行训练。在该再次训练过程中:将预测模型903的N个隐含层中的、预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型905。此时,可以利用该更新后的预测模型805对工厂A中的分析对象的用电量执行预测。
在图9的另一个实施方式中,分析对象的第一数据901为:在包含分析对象的第一应用场景中所采集的、第十八时间区间内的仿真数据;分析对象的第二数据904为:在包含分析对象的第二应用场景中所采集的、包含第十九时间区间内的仿真数据和第二十时间区间内的实际数据的组合数据。其中,第十九时间区间和第二十时间区间可以相同或不同,优选不同;第十九时间区间与第十八时间区间可以相同或不同,优选不同。优选不同。相比较第一数据901而言,第二数据904与第二应用场景中的预测任务更加相关。
比如,预测任务为包含分析对象(比如HVAC系统)的第二应用场景(比如工厂B)中的用电量预测;第一数据801包含工厂A中2011年2月份的仿真风扇使用时间和2011年2月份的用电量仿真数据;第二数据804为包含工厂B中2011年1月份的仿真风扇使用时间和工厂B中2011年3月份的用电量实际数据的组合数据。
首先,将第一数据801作为训练数据输入到包含N个隐含层的神经网络模型802中,经过训练过程,可以将该神经网络模型802训练为分析对象的预测模型803。在该预测模型803中,N个隐含层的参数都得到确定。
然后,将第二数据804作为训练数据输入到预测模型803中以再次执行训练。在该再次训练过程中:将预测模型803的N个隐含层中的、与第二数据804的应用环境不相关的预定M个隐含层保持固定(即不更新这M个隐含层的参数),而利用该再次训练过程更新剩余的、与第二数据804的应用环境相关的(N-M)个隐含层的参数。经过再次执行训练,可以得到更新后的预测模型805。此时,可以利用该更新后的预测模型805对工厂B中的分析对象的用电量执行预测。
在本发明实施方式中,可以使用元数据构建分析对象的仿真模型,而且使用仿真模型生成仿真数据(数据集D S)。接着,执行训练分析对象的预测模型的过程。在该过程中,可以将仿真数据D S与分析对象的实际数据D A相组合得到组合数据D C,即D c=D S+D A,并利用组合数据D C对优选实施为LSTM模型的神经网络模型进行训练以得到分析对象的预测模型。在该过程中,还可以首先利用分析对象的实际数据D A对优选实施为LSTM模型的神经网络模型进行训练以得到分析对象的预测模型M A。该M A既可以用于对分析对象执行预测。进一步地,可以利用分析对象的仿真数据D S对预测模型M A进行训练。在利用分析对象的仿真数据D S对预测模型M A进行训练的过程中,将预测模型M A的预定隐含层的参数保持固定,而针对除了该预定隐含层之外的其它隐含层的参数进行更新。经过更新后的预测模型M A相比较更新前的预测模型M A具有更好的准确度。
下面对迁移学习转移到的应用环境中,更新隐含层与不更新隐含层的效果进行示范性对 比。对于包含2个隐含层的LSTM网络模型,当当应用环境中不采用更新隐含层方式时,全部可训练的参数的数目为508201个。当应用环境采用更新隐含层方式时,可训练的参数的数目为201个。因此,更新隐含层后,可训练的参数数目显著降低,从而提高了训练速度。
在本发明实施方式中,还提出HVAC系统的用电量预测方法。
图10为本发明实施方式的HVAC系统的用电量预测方法的流程图。如图10所示,该方法包括:
步骤1001:接收预测时间;
步骤1002:基于HVAC系统的用电量预测模型生成对应于所述预测时间的用电量预测值;其中生成所述用电量预测模型的方法包括:获取HVAC系统的第一用电量数据;获取所述HVAC系统的第二用电量数据;利用所述第一用电量数据将包含N个隐含层的神经网络模型训练为所述用电量预测模型,其中N为至少为2的正整数;利用所述第二用电量数据更新包含在所述用电量预测模型中的所述N个隐含层中的至少一个隐含层。
在一个实施方式中,第一用电量数据为第一时间区间内的用电量实际数据,所述第二用电量数据为第二时间区间内的用电量仿真数据;或,所述第一用电量数据为第三时间区间内的用电量实际数据,所述第二用电量数据为第四时间区间内的用电量实际数据;或,所述第一用电量数据为第五时间区间内的用电量仿真数据,所述第二用电量数据为第六时间区间内的用电量实际数据;或,所述第一用电量数据为第七时间区间内的用电量仿真数据,所述第二用电量数据为第八时间区间内的用电量仿真数据;或,所述第一用电量数据为包含第九时间区间内的用电量仿真数据和第十时间区间内的用电量实际数据的组合数据,所述第二用电量数据为第十一时间区间内的用电量实际数据;或,所述第一用电量数据为包含第十二时间区间内的用电量仿真数据和第十三时间区间内的用电量实际数据的组合数据,所述第二用电量数据为第十四时间区间内的用电量仿真数据;或,所述第一用电量数据为第十五时间区间内的用电量实际数据,所述第二用电量数据为包含第十六时间区间内的用电量仿真数据和第十七时间区间内的实际数据的组合数据;或,所述第一用电量数据为第十八时间区间内的仿真数据,所述第二用电量数据为包含第十九时间区间内的用电量仿真数据和第二十时间区间内的用电量实际数据的组合数据。
在一个实施方式中,利用第二用电量数据更新包含在所述用电量预测模型中的所述N个隐含层中的至少一个隐含层包括:利用所述第二用电量数据训练所述预测模型,其中固定所述预测模型中的、预定的M个隐含层,更新所述预测模型中的、除所述M个隐含层之外的剩余隐含层,其中M为至少为2的正整数,且M小于等于N。
图11为本发明实施方式的HVAC系统的用电量预测装置的结构图。
如图11所示,HVAC系统的用电量预测装置80包括:
接收模块81,用于接收预测时间;
预测模块82,用于基于HVAC系统的用电量预测模型生成对应于所述预测时间的用电量预测值;其中生成所述用电量预测模型的方法包括:获取所述HVAC系统的第一用电量数据;获取所述HVAC系统的第二用电量数据,其中第二用电量数据的数据类型与第一用电量数据的数据类型不相同;利用所述第一用电量数据将包含N个隐含层的神经网络模型训练为所述用电量预测模型,其中N为至少为2的正整数;利用所述第二用电量数据更新包含在所述用电量预测模型中的所述N个隐含层中的至少一个隐含层。
下面以将建筑物中的HVAC为例,描述本发明实施方式的实现过程。
建筑物为用电大户,其HVAC系统同占总用电量的30%至40%。建筑物业主或运营商期望可以降低能源成本,同时满足舒适的室温、良好的空气质量等运行要求。使用AI技术可以平衡需求方和供应方之间的能源浪费,提高能源利用效率并降低能耗。但是,目前关于HVAC的AI技术在应用中存在挑战。这些挑战包括:
(1)、存在数据可用性问题。高质量的足够数据对于数据驱动的模型非常重要。缺乏数据,就无法训练和测试模型,更不用说提供见解或建议。在数字化和物联网技术尚未得到广泛应用时,现有建筑物中的很大一部分已经建成。一些建筑物已经安装了一些用于系统监视的传感器,但是从这些传感器收集的数据不足以进行模型训练和测试。另外,在当前系统中安装新传感器既困难又昂贵。
(2)、存在解决方案的可迁移性问题。如果解决方案可以轻松地从一个客户转移到另一位客户,甚至从一种类型的客户(例如,办公楼)转移到另一种类型的客户(例如,商业楼),那么研发成本将显著降低。
在本发明实施方式中,可以为数据很少的建筑物的实现数据丰富。例如,暖通空调系统的能耗具有季节性规律(例如,夏季制冷需求高而冬季制冷需求低),如何仅利用可用的冬季数据来预测夏季的电耗。在本发明实施方式中,还可以在相似的建筑物之间实现迁移学习,或从一种类型的建筑物(例如办公楼)转移到另一种类型的建筑物(例如购物商场),或从一个季节(例如冬天)转移到另一个季节(例如夏天)。
下面以举例描述和评估HVAC系统的用电量预测方法。数据来自LCD制造工厂的冷却系统。它具有33个数据特征,包括室外温度、湿度、二次冷冻水回路运行参数(例如供水和回水温度)和泵的流量,等等。响应变量是冷却系统的总耗电量。可以提供四个月(2017年 3月、4月、6月和8月)的历史数据。
假设:提供了2017年3月的实际数据(样本量为568)以及提供2017年4月的仿真数据(样本量为617)。
目标为:预测2017年6月(样本量为584)和2017年8月(样本量为528)的用电量。其中,采用的KPI为:均方根误差(Mean Absolution Percentage Error,MAPE):
Figure PCTCN2020132933-appb-000001
表1为采用各种方式训练预测模型的示意表。
在表中,训练的标准方法中不采用迁移学习,即直接利用训练数据训练出预测模型。在迁移学习中,首先利用预训练数据(data for pretraining)训练出预测模型,再利用训练数据对预测模型中的隐含层进行更新。
由表1可见,采用了迁移学习后,训练的运行时间显著降低,而且MAPE也显著减少,从而还提高了预测模型的准确度。
Figure PCTCN2020132933-appb-000002
表1
图12为本发明实施方式的HVAC用电量的预测示意图。
在图12中,横坐标为时间,纵坐标为用电量。曲线61为冬天的实际用电量;曲线62为夏天的仿真用电量;曲线63为夏天的实际用电量;曲线64为采用本发明实施方式的预测用电量;曲线65为采用标准方法的预测用电量。
图13为本发明实施方式的HVAC用电量的预测装置的结构图。
在图13中,HVAC用电量的预测装置30包括:
接口31,用于接收预测时间;
存储器32,用于保存供热HVAC系统的用电量预测模型,其中生成所述用电量预测模型的方法包括:获取所述HVAC系统的第一用电量数据;获取所述HVAC系统的第二用电量数据,其中第二用电量数据的数据类型与第一用电量数据的数据类型不相同;利用所述第一用电量数据训练包含多个隐含层的神经网络模型以获取所述HVAC系统的用电量预测模型;利用所述第二用电量数据训练所述用电量预测模型以更新所述HVAC系统的至少一个隐含层;
处理器33,经由总线34与所述接口31和所述存储32器分别耦合,用于基于所述用电量预测模型生成对应于所述预测时间的用电量预测值。
图14为本发明实施方式的生成分析对象的预测模型的装置的示范性结构图。该装置50包括:
第一数据获取模块51,用于获取分析对象的第一数据;
第二数据获取模块52,用于获取所述分析对象的第二数据,其中第二数据与第一数据表征分析对象的相同物理量,第二数据的数据类型与第一数据的数据类型不相同;
训练模块53,用于利用所述第一数据将包含N个隐含层的神经网络模型训练为所述分析对象的预测模型,其中N为至少为2的正整数;
更新模块54,用于利用所述第二数据更新包含在所述预测模型中的所述N个隐含层中的至少一个隐含层。
在一个实施方式中,所述第一数据为第一时间区间内的实际数据,所述第二数据为第二时间区间内的仿真数据;或,所述第一数据为第三时间区间内的实际数据,所述第二数据为第四时间区间内的实际数据;或,所述第一数据为第五时间区间内的仿真数据,所述第二数据为第六时间区间内的实际数据;或,所述第一数据为第七时间区间内的仿真数据,所述第二数据为第八时间区间内的仿真数据;或,所述第一数据为包含第九时间区间内的仿真数据和第十时间区间内的实际数据的组合数据,所述第二数据为第十一时间区间内的实际数据;或,所述第一数据为包含第十二时间区间内的仿真数据和第十三时间区间内的实际数据的组合数据,所述第二数据为第十四时间区间内的仿真数据;或,所述第一数据为第十五时间区间内的实际数据,所述第二数据为包含第十六时间区间内的仿真数据和第十七时间区间内的实际数据的组合数据;或,所述第一数据为第十八时间区间内的仿真数据,所述第二数据为包含第十九时间区间内的仿真数据和第二十时间区间内的实际数据的组合数据。
在一个实施方式中,所述组合数据中的实际数据和仿真数据具有时间属性上可叠加的数据指标或在时间属性上不可叠加的数据指标。
在一个实施方式中,该装置50还包括:仿真数据获取模块55,用于基于预定的分析对 象元数据建立所述分析对象的仿真模型;基于所述仿真模型生成所述仿真数据。
在一个实施方式中,分析对象为HVAC系统,所述预测模型为用电量预测模型;该装置50还包括:接收模块56,用于接收预测时间;预测模块57,用于基于所述用电量预测模型生成对应于所述预测时间的用电量预测值。
在一个实施方式中,更新模块54,用于利用所述第二数据训练所述预测模型,其中固定所述预测模型中的、预定的M个隐含层,更新所述预测模型中的、除所述M个隐含层之外的剩余隐含层,其中M为至少为2的正整数,且M小于等于N。
图15为本发明实施方式的具有存储器-处理器架构的、生成分析对象的预测模型的装置的示范性结构图。
在图15中,生成分析对象的预测模型的装置70包括一个存储器72和一个处理器71;存储器72中存储有可被处理器71执行的应用程序,用于使得处理器71执行如上任一项所述的生成分析对象的预测模型的方法。
需要说明的是,上述各流程和各结构图中不是所有的步骤和模块都是必须的,可以根据实际的需要忽略某些步骤或模块。各步骤的执行顺序不是固定的,可以根据需要进行调整。各模块的划分仅仅是为了便于描述采用的功能上的划分,实际实现时,一个模块可以分由多个模块实现,多个模块的功能也可以由同一个模块实现,这些模块可以位于同一个设备中,也可以位于不同的设备中。
各实施方式中的硬件模块可以以机械方式或电子方式实现。例如,一个硬件模块可以包括专门设计的永久性电路或逻辑器件(如专用处理器,如FPGA或ASIC)用于完成特定的操作。硬件模块也可以包括由软件临时配置的可编程逻辑器件或电路(如包括通用处理器或其它可编程处理器)用于执行特定操作。至于具体采用机械方式,或是采用专用的永久性电路,或是采用临时配置的电路(如由软件进行配置)来实现硬件模块,可以根据成本和时间上的考虑来决定。
本发明还提供了一种机器可读的存储介质,存储用于使一机器执行如本文所述方法的指令。具体地,可以提供配有存储介质的系统或者装置,在该存储介质上存储着实现上述实施例中任一实施方式的功能的软件程序代码,且使该系统或者装置的计算机(或CPU或MPU)读出并执行存储在存储介质中的程序代码。此外,还可以通过基于程序代码的指令使计算机上操作的操作系统等来完成部分或者全部的实际操作。还可以将从存储介质读出的程序代码写到插入计算机内的扩展板中所设置的存储器中或者写到与计算机相连接的扩展单元中设置的存储器中,随后基于程序代码的指令使安装在扩展板或者扩展单元上的CPU等来执行部分 和全部实际操作,从而实现上述实施方式中任一实施方式的功能。用于提供程序代码的存储介质实施方式包括软盘、硬盘、磁光盘、光盘(如CD-ROM、CD-R、CD-RW、DVD-ROM、DVD-RAM、DVD-RW、DVD+RW)、磁带、非易失性存储卡和ROM。可选择地,可以由通信网络从服务器计算机或云上下载程序代码。
上文通过附图和优选实施例对本发明进行了详细展示和说明,然而本发明不限于这些已揭示的实施例。基与上述多个实施例,本领域技术人员可以知晓,可以组合上述不同实施例中的代码审核手段得到本发明更多的实施例,这些实施例也在本发明的保护范围之内。

Claims (14)

  1. 生成分析对象的预测模型的方法(100),其特征在于,包括:
    获取(101)分析对象的第一数据;
    获取(102)所述分析对象的第二数据,其中所述第二数据与所述第一数据表征所述分析对象的相同物理量,所述第二数据的数据类型与所述第一数据的数据类型不相同;
    利用所述第一数据将包含N个隐含层的神经网络模型训练(103)为所述分析对象的预测模型,其中N为至少为2的正整数;
    利用所述第二数据更新(104)包含在所述预测模型中的所述N个隐含层中的至少一个隐含层。
  2. 根据权利要求1所述的方法(100),其特征在于,所述第二数据的数据类型与所述第一数据的数据类型不同,包括:
    所述第一数据为第一时间区间内的实际数据,所述第二数据为第二时间区间内的仿真数据;或
    所述第一数据为第三时间区间内的实际数据,所述第二数据为第四时间区间内的实际数据;或
    所述第一数据为第五时间区间内的仿真数据,所述第二数据为第六时间区间内的实际数据;或
    所述第一数据为第七时间区间内的仿真数据,所述第二数据为第八时间区间内的仿真数据;或
    所述第一数据为包含第九时间区间内的仿真数据和第十时间区间内的实际数据的组合数据,所述第二数据为第十一时间区间内的实际数据;或
    所述第一数据为包含第十二时间区间内的仿真数据和第十三时间区间内的实际数据的组合数据,所述第二数据为第十四时间区间内的仿真数据;或
    所述第一数据为第十五时间区间内的实际数据,所述第二数据为包含第十六时间区间内的仿真数据和第十七时间区间内的实际数据的组合数据;或
    所述第一数据为第十八时间区间内的仿真数据,所述第二数据为包含第十九时间区间内的仿真数据和第二十时间区间内的实际数据的组合数据。
  3. 根据权利要求2所述的方法(100),其特征在于,所述组合数据中的实际数据和仿真数据具有时间属性上可叠加的数据指标或在时间属性上不可叠加的数据指标。
  4. 根据权利要求2所述的方法(100),其特征在于,还包括:
    基于预定的分析对象元数据建立所述分析对象的仿真模型;
    基于所述仿真模型生成所述仿真数据。
  5. 根据权利要求1所述的方法(100),其特征在于,所述分析对象为供热通风与空气调节HVAC系统,所述预测模型为用电量预测模型;该方法(100)还包括:
    接收预测时间;
    基于更新后的所述预测模型生成对应于所述预测时间的用电量预测值。
  6. 根据权利要求1-5中任一项所述的方法(100),其特征在于,
    所述利用第二数据更新(104)包含在所述预测模型中的所述N个隐含层中的至少一个隐含层包括:
    利用所述第二数据训练所述预测模型,其中固定所述预测模型中的、预定的M个隐含层,更新所述预测模型中的、除所述M个隐含层之外的剩余隐含层,其中M为至少为2的正整数,且M小于等于N。
  7. 生成分析对象的预测模型的装置(50),其特征在于,包括:
    第一数据获取模块(51),用于获取分析对象的第一数据;
    第二数据获取模块(52),用于获取所述分析对象的第二数据,其中所述第二数据与所述第一数据表征所述分析对象的相同物理量,所述第二数据的数据类型与所述第一数据的数据类型不相同;
    训练模块(53),用于利用所述第一数据将包含N个隐含层的神经网络模型训练为所述分析对象的预测模型,其中N为至少为2的正整数;
    更新模块(54),用于利用所述第二数据更新包含在所述预测模型中的所述N个隐含层中的至少一个隐含层。
  8. 根据权利要求7所述的装置(50),其特征在于,所述第二数据的数据类型与所述第一数据的数据类型不相同,包括:
    所述第一数据为第一时间区间内的实际数据,所述第二数据为第二时间区间内的仿真数据;或
    所述第一数据为第三时间区间内的实际数据,所述第二数据为第四时间区间内的实际数据;或
    所述第一数据为第五时间区间内的仿真数据,所述第二数据为第六时间区间内的实际数据;或
    所述第一数据为第七时间区间内的仿真数据,所述第二数据为第八时间区间内的仿真数据;或
    所述第一数据为包含第九时间区间内的仿真数据和第十时间区间内的实际数据的组合数 据,所述第二数据为第十一时间区间内的实际数据;或
    所述第一数据为包含第十二时间区间内的仿真数据和第十三时间区间内的实际数据的组合数据,所述第二数据为第十四时间区间内的仿真数据;或
    所述第一数据为第十五时间区间内的实际数据,所述第二数据为包含第十六时间区间内的仿真数据和第十七时间区间内的实际数据的组合数据;或
    所述第一数据为第十八时间区间内的仿真数据,所述第二数据为包含第十九时间区间内的仿真数据和第二十时间区间内的实际数据的组合数据。
  9. 根据权利要求8所述的装置(50),其特征在于,
    所述组合数据中的实际数据和仿真数据具有时间属性上可叠加的数据指标或在时间属性上不可叠加的数据指标。
  10. 根据权利要求7所述的装置(50),其特征在于,还包括:
    仿真数据获取模块(55),用于基于预定的分析对象元数据建立所述分析对象的仿真模型;基于所述仿真模型生成所述仿真数据。
  11. 根据权利要求7所述的装置(50),其特征在于,
    所述分析对象为供热通风与空气调节HVAC系统,所述预测模型为用电量预测模型;该装置(50)还包括:
    接收模块(56),用于接收预测时间;
    预测模块(57),用于基于所述用电量预测模型生成对应于所述预测时间的用电量预测值。
  12. 根据权利要求7-11中任一项所述的装置(50),其特征在于,
    所述更新模块(54),用于利用所述第二数据训练所述预测模型,其中固定所述预测模型中的、预定的M个隐含层,更新所述预测模型中的、除所述M个隐含层之外的剩余隐含层,其中M为至少为2的正整数,且M小于等于N。
  13. 生成分析对象的预测模型的装置(70),其特征在于,包括处理器(71)和存储器(72);
    所述存储器(72)中存储有可被所述处理器(71)执行的应用程序,用于使得所述处理器(71)执行如权利要求1至6中任一项所述的生成分析对象的预测模型的方法(100)。
  14. 计算机可读存储介质,其特征在于,其中存储有计算机可读指令,该计算机可读指令用于执行如权利要求1至6中任一项所述的生成分析对象的预测模型的方法(100)。
PCT/CN2020/132933 2020-11-30 2020-11-30 生成分析对象的预测模型的方法、装置和存储介质 WO2022110213A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202080105837.6A CN116235188A (zh) 2020-11-30 2020-11-30 生成分析对象的预测模型的方法、装置和存储介质
PCT/CN2020/132933 WO2022110213A1 (zh) 2020-11-30 2020-11-30 生成分析对象的预测模型的方法、装置和存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/132933 WO2022110213A1 (zh) 2020-11-30 2020-11-30 生成分析对象的预测模型的方法、装置和存储介质

Publications (1)

Publication Number Publication Date
WO2022110213A1 true WO2022110213A1 (zh) 2022-06-02

Family

ID=81753955

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/132933 WO2022110213A1 (zh) 2020-11-30 2020-11-30 生成分析对象的预测模型的方法、装置和存储介质

Country Status (2)

Country Link
CN (1) CN116235188A (zh)
WO (1) WO2022110213A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115791957A (zh) * 2022-11-25 2023-03-14 国网四川省电力公司映秀湾水力发电总厂 高压电缆铅封缺陷脉冲涡流检测方法、系统、终端及介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109685277A (zh) * 2018-12-28 2019-04-26 国网冀北电力有限公司经济技术研究院 用电量预测方法及装置
US20190197550A1 (en) * 2017-12-21 2019-06-27 Paypal, Inc. Generic learning architecture for robust temporal and domain-based transfer learning
CN110703899A (zh) * 2019-09-09 2020-01-17 创新奇智(南京)科技有限公司 一种基于迁移学习的数据中心能效优化方法
CN110995475A (zh) * 2019-11-20 2020-04-10 国网湖北省电力有限公司信息通信公司 一种基于迁移学习的电力通信网故障检测方法
CN111664823A (zh) * 2020-05-25 2020-09-15 重庆大学 基于介质热传导系数差异的均压电极垢层厚度检测方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190197550A1 (en) * 2017-12-21 2019-06-27 Paypal, Inc. Generic learning architecture for robust temporal and domain-based transfer learning
CN109685277A (zh) * 2018-12-28 2019-04-26 国网冀北电力有限公司经济技术研究院 用电量预测方法及装置
CN110703899A (zh) * 2019-09-09 2020-01-17 创新奇智(南京)科技有限公司 一种基于迁移学习的数据中心能效优化方法
CN110995475A (zh) * 2019-11-20 2020-04-10 国网湖北省电力有限公司信息通信公司 一种基于迁移学习的电力通信网故障检测方法
CN111664823A (zh) * 2020-05-25 2020-09-15 重庆大学 基于介质热传导系数差异的均压电极垢层厚度检测方法

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115791957A (zh) * 2022-11-25 2023-03-14 国网四川省电力公司映秀湾水力发电总厂 高压电缆铅封缺陷脉冲涡流检测方法、系统、终端及介质

Also Published As

Publication number Publication date
CN116235188A (zh) 2023-06-06

Similar Documents

Publication Publication Date Title
US11238055B2 (en) Building management system with eventseries processing
US10552761B2 (en) Non-intrusive fine-grained power monitoring of datacenters
WO2021174811A1 (zh) 车流量时间序列的预测方法及预测装置
CN103766015B (zh) 用于数据中心温度和气流值辅助预测的测量系统与方法
US9429921B2 (en) Method and system for energy control management
US9753477B2 (en) Load forecasting for residential sector demand response
US20170169344A1 (en) Methods, systems, and computer readable media for a data-driven demand response (dr) recommender
CN112766596A (zh) 建筑能耗预测模型的构建方法、能耗预测方法及装置
CN112183869A (zh) 一种基于大数据的建筑能耗预测方法及系统
CN117121025A (zh) 暖通空调系统的管理方法、系统及存储介质
CN116796141A (zh) 一种基于gbdt回归模型的办公建筑能耗预测方法
WO2022110213A1 (zh) 生成分析对象的预测模型的方法、装置和存储介质
Long et al. Development of a metamodelling framework for building energy models with application to fifth-generation district heating and cooling networks
Ghassemi et al. Optimal surrogate and neural network modeling for day-ahead forecasting of the hourly energy consumption of university buildings
CN116432524A (zh) 一种变压器油温预测方法、装置、设备及存储介质
Zhang et al. Investigations on machine learning-based control-oriented modeling using historical thermal data of buildings
US20200278130A1 (en) Operation control method, storage medium, and operation control device
Zheng et al. Melody: Adaptive task definition of cop prediction with metadata for hvac control and electricity saving
CN113170592B (zh) 基于监测/控制机制的热控制优化
Shchetinin Development of Energy Saving Technologies for Smart Buildings by Using Computer Algebra
CN116528270B (zh) 一种基站节能潜力评估方法、装置、设备及存储介质
US20160132619A1 (en) Energy management apparatus, energy management method, and non-transitory tangible computer readable medium thereof
Hu et al. An Integrated Prediction Model for Building Energy Consumption: A Case Study
Kammeyer et al. Towards an HPC cluster digital twin and scheduling framework for improved energy efficiency
CN113886927A (zh) 空调系统末端设备逐时能耗拆分方法、装置及设备及介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20963101

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20963101

Country of ref document: EP

Kind code of ref document: A1