CN113240184B - Building space unit cold load prediction method and system based on federal learning - Google Patents

Building space unit cold load prediction method and system based on federal learning Download PDF

Info

Publication number
CN113240184B
CN113240184B CN202110562321.0A CN202110562321A CN113240184B CN 113240184 B CN113240184 B CN 113240184B CN 202110562321 A CN202110562321 A CN 202110562321A CN 113240184 B CN113240184 B CN 113240184B
Authority
CN
China
Prior art keywords
load
data
value
weight
model
Prior art date
Legal status (The legal status 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 status listed.)
Active
Application number
CN202110562321.0A
Other languages
Chinese (zh)
Other versions
CN113240184A (en
Inventor
刘振杰
黄文君
胡斌
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University ZJU
Original Assignee
Zhejiang University ZJU
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 Zhejiang University ZJU filed Critical Zhejiang University ZJU
Priority to CN202110562321.0A priority Critical patent/CN113240184B/en
Publication of CN113240184A publication Critical patent/CN113240184A/en
Application granted granted Critical
Publication of CN113240184B publication Critical patent/CN113240184B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Business, Economics & Management (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Economics (AREA)
  • Molecular Biology (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Game Theory and Decision Science (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Quality & Reliability (AREA)
  • Operations Research (AREA)
  • Marketing (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Development Economics (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Air Conditioning Control Device (AREA)

Abstract

The invention discloses a building space unit cold load prediction method and a system based on federal learning, wherein the method respectively trains building space unit cold load prediction model weights of corresponding regions through neural network models distributed in different regions and calculates air conditioner load prediction values of the corresponding regions and unit air conditioner load prediction values; obtaining model weight, air conditioner load predicted value and unit air conditioner load predicted value of the zone bit with aggregation value, judging and classifying different zone bit models, and aggregating the weight values of the models classified into one class by using a transverse federal learning method; aggregating the network load predicted values with different weights by utilizing a longitudinal federal learning method to establish a regression prediction model, and predicting the cold load of the building space unit according to the regression prediction model; and updating the neural network model of the corresponding zone according to the model weight of the aggregated different zones. The method has stronger generalization, accelerates the network training efficiency and reduces the communication cost required by realizing the air conditioner load prediction function.

Description

Building space unit cold load prediction method and system based on federal learning
Technical Field
The invention relates to a distributed air conditioner load prediction technology, in particular to a building space unit cold load prediction method and system based on federal learning.
Background
In recent years, with the gradual enhancement of domestic environmental protection consciousness, environmental protection and energy conservation are always the goals pursued by enterprises and even social environments. The total energy consumption of China in the aspect of buildings accounts for 30% of the total social energy consumption, and the energy consumption generated by heating and ventilation equipment accounts for more than 60% of the energy consumption of the buildings, so that unnecessary energy consumption can be reduced by optimizing the operation load curve of the heating and ventilation equipment, and resources can be saved to a considerable extent. For enterprises applying the ice storage technology air conditioning equipment, considerable cost can be saved for the enterprises by replacing a method of acquiring electric energy from a power grid when the electricity price is peak by cold storage when the electricity price is low.
However, in order to realize the optimized operation of the air conditioning unit, the control mechanism timely responds to the optimization strategy, and needs to accurately predict the load demand in advance, so as to ensure the safe and stable operation of the air conditioning equipment. In recent years, there are several commonly used air conditioning load prediction methods: time series analysis, multiple regression analysis, and artificial neural networks. The time series analysis method has good prediction effect on the cold load of the stable building space unit and simple calculation, but the prediction effect is reduced when the load is influenced by non-time factors. Although the multivariate regression analysis method can take non-time factors into consideration, the model generalization is low, and the application effect of different types of buildings is different. The artificial neural network is an emerging prediction method, has strong generalization and can achieve high prediction accuracy when sufficient samples are given, but the method has lower operation speed than the former two methods and easily causes the problem that real-time control cannot be met due to insufficient prediction speed, so the conventional artificial neural network prediction method needs to be improved according to a hardware computing unit.
Nowadays, most large buildings are temperature-regulated by central heating and ventilation equipment, so that sensors need to be arranged as dispersedly and uniformly as possible when acquiring source data of building load prediction, certain communication pressure is generated if data collected by the sensors are transmitted to a prediction unit by means of communication, and the cost for arranging so many communication facilities is high. In order to solve the overhead problem of communication, a cloud-edge cooperation technology is adopted, an edge computing device collects data of surrounding sensors and carries out operation, then an operation result is transmitted to a cloud computing device, and a prediction method of an artificial neural network is optimized through a federal learning technology.
In some existing air conditioner load prediction embodiments, most of the air conditioner load prediction methods tend to utilize a neural network or a support vector machine for prediction, but the methods are single-machine computing, and under the trend of building development with larger and larger scale, the communication cost is higher and higher, so that a joint learning and cloud-side cooperation technology is needed to complete the prediction of the air conditioner load.
Disclosure of Invention
The invention aims to provide a building space unit cold load prediction method and system based on federal learning, aiming at the defects of the prior art.
The purpose of the invention is realized by the following technical scheme:
in a first aspect, the invention provides a building space unit cold load prediction method based on federal learning, which comprises the following steps:
step S1: acquiring data related to the prediction of the cold load of building space units in different locations of a building;
step S2: respectively training the weight of a building space unit cold load prediction model of the corresponding zone through neural network models distributed in different zones according to the data obtained in the step S1, and calculating the predicted value of the air conditioner load of the corresponding zone and the predicted value of the air conditioner load of the unit of the corresponding zone;
step S3: obtaining model weight, air conditioner load predicted value and unit air conditioner load predicted value of the zone bit with aggregation value, judging and classifying models of different zone bits, and aggregating the weight values of the models classified into one type by using a transverse federal learning method;
step S4: aggregating the network load predicted values with different weights to establish a regression prediction model by utilizing a longitudinal federal learning method, and predicting the cold load of the building space unit according to the regression prediction model;
step S5: and updating the neural network model of the corresponding zone according to the model weight of the aggregated different zones.
Furthermore, the data related to the predicted building space unit cold loads in different areas of the building are obtained as the outdoor air temperature at the current predicted time T, the outdoor air temperature at the last time T-1, the indoor pedestrian volume at the current time T, the indoor pedestrian volume at the last time T-1, the solar radiation intensity at the current time T, the solar radiation intensity at the last time T-1, the relative humidity at the current time T, the unit air conditioner cold load at the last time T-1, the air conditioner cold loads T-2 before two times and the air conditioner cold load at the same time T-24 before.
Furthermore, the time T is a prediction time, the prediction step period is 1 hour, 60 sampling values are obtained within one prediction period, and the air conditioning cooling load at the time T is predicted by the data source, so as to obtain the space unit cooling load.
Further, in step S2, the data obtained from different regions are respectively preprocessed and processed into input data that can be directly input into the neural network model; the data preprocessing method specifically comprises the following steps:
step S21: and eliminating sampling error data.
First, a data point is collected every minute, and the statistic F of the sampled data in a 1-hour sampling period is obtained:
Figure BDA0003078007850000021
wherein s isiIs the ith sample point and is the last sample point,
Figure BDA0003078007850000022
is the average of the sampling points and is,
Figure BDA0003078007850000023
and calculating the regression equation parameters by a least square method for the regression fitting value of the ith sampling point, wherein the regression fitting value is obtained by the regression equation of the linear function.
Then obtaining statistic and judging FαSize of (1,58), Fα(1,58) represents the critical value of the rejection region with a degree of freedom of 1 and a data point of 60, when F > Fα(1,58), eliminating data, and when F is less than or equal to FαAnd (1,58), removing the point with the maximum fitting value difference value with the sampling point from the sampling points, and judging as above.
Step S22: and taking the average value in the sampling value prediction period as the representative value of the current prediction period.
Step S23: by the formula
Figure BDA0003078007850000031
The normalization method is to sample the dataAnd (4) dimensionless.
Wherein n is1The number of sampling points after sampling error data is eliminated.
Further, the neural network model in step S2 has a BP neural network structure, the weight of the BP neural network is updated in a particle swarm optimization manner, and the particle fitness V is calculated by using the following formula:
Figure BDA0003078007850000032
wherein n is2Is the number of sample point classes, yiIs the true value of the cooling load of the ith time sequence unit, oiIs the predicted value of the cooling load of the ith time sequence unit.
The velocity and position update formula of the particle is:
Figure BDA0003078007850000033
Figure BDA0003078007850000034
wherein,
Figure BDA0003078007850000035
respectively the speed and position of the ith particle in the t iteration, u is the inertia weight, c1And c2As an acceleration factor, c3As disturbance factor, c3Usually a constant between (0,1),
Figure BDA0003078007850000036
for the extreme value of the ith particle for the t iteration,
Figure BDA0003078007850000037
the extreme value of the total particles in the t-th iteration,
Figure BDA0003078007850000038
for the position of the ith particle in the t iteration, r1And r2Is a random number between (0, 1).
Further, the method for determining the aggregated valuable area in step S3 is as follows:
and when the calculated weight update values of the neural networks deployed in different locations meet the following formula, the values are considered to be aggregated:
Figure BDA0003078007850000039
Figure BDA00030780078500000310
Figure BDA00030780078500000311
wherein, wtThe weight value of the t-th iteration update of the position,
Figure BDA0003078007850000041
weight gradient for the tth iteration update of the location, mhAs the number of history weights used for fitting, mhThe value is greater than 5, m is the total number of zone bits, alpha is a learning rate, and alpha is usually (0,1)]Is constant.
Further, the method for judging and classifying models of different regions in step S3 includes the following steps:
step S31: arranging the historical values of the weights from different regions according to a time sequence:
Figure BDA0003078007850000042
wherein,
Figure BDA0003078007850000043
weight vector for the t iteration of the jth zone bitH is the number of historical weights used for judging relevance, and the value of h is larger than 10.
Step S32: normalizing the weight value:
Figure BDA0003078007850000044
step S33: generating a differencing sequence for the compared locational network weights:
Figure BDA0003078007850000045
wherein
Figure BDA0003078007850000046
Is the normalized weight vector for the compared location o.
Step S34: determining the correlation with the compared locational network weight:
Figure BDA0003078007850000047
step S35: go through m above steps S31-S34v1 pass formation of a matrix of correlations between weights of different locations, mvFor the number of locations that are of aggregation value, the location weight above a relevance threshold in the square matrix is classified as a class model weight.
Further, the method of horizontal federal learning in step S3 is as follows:
Figure BDA0003078007850000048
wherein,
Figure BDA0003078007850000049
is the weight of the aggregation after the t round of iteration, msThe number of zone bits belonging to the same class model.
Further, after performing multiple iterations,if it is not
Figure BDA00030780078500000410
When the convergence is not achieved, the polymerization method is switched to:
Figure BDA00030780078500000411
where λ is a velocity constant, typically a constant between (0, 1).
Further, the regression prediction model building method in step S4 includes the following steps:
step S41: regarding the air-conditioning cold load prediction data from different models as a multiple regression independent variable of the total air-conditioning cold load of the building, establishing a multiple linear regression model of the air-conditioning cold load prediction data and the building total air-conditioning cold load:
Figure BDA0003078007850000051
wherein, betalRegression coefficient, x, for predicted values of class I zone bitslIs the sum of predicted values of air conditioner load in class I zone0Is the constant term coefficient of the regression model, ε is the noise term, N is a Gaussian distribution, σ is the standard deviation of white noise, mkFor different types of zone numbers, y is the actual total air conditioning load of the building.
Step S42: the regression coefficient is obtained by the following formula:
Figure BDA0003078007850000052
wherein m is1For the number of data used to determine the sum of the historical air conditioning load predictions for each location of the regression coefficients,
Figure BDA0003078007850000053
for the l-type region1Sum of predicted values of load of the wheel air conditioners.
Further, the prediction method in step S4 is: and after the air conditioner load predicted values of different regions are obtained, calculating the air conditioner predicted load of the whole building through a regression prediction model, and obtaining the indoor building area of the whole building, wherein the ratio of the air conditioner predicted load of the whole building to the indoor building area is the building space unit cold load predicted value.
Further, the method for updating the neural network models corresponding to different areas in step S5 includes: and replacing the neural network weight of the corresponding zone with the aggregated model weight of the different zone.
Furthermore, the above model needs to be trained by historical sampling data before prediction, and then parameters of the prediction model are further corrected by real-time sampling data and predicted in real time.
In a second aspect, the invention provides a building space unit cold load prediction system based on federal learning, which comprises a data acquisition module, an edge calculation module and a cloud calculation module; each zone bit of the building corresponds to one edge calculation module, each edge calculation module is connected with a plurality of data acquisition modules, and the plurality of edge calculation modules are connected with one cloud calculation module.
The data acquisition module is used for acquiring data related to the cold load of the building space unit predicted at different locations of the building;
the edge calculation module receives and processes data collected by corresponding zone bits, trains building space unit cold load prediction model weight of the zone bits through a neural network model of the zone bits, and calculates a zone bit air conditioner load prediction value and a corresponding zone bit unit air conditioner load prediction value; judging whether the model weight of the local zone bit has the aggregation value uploaded to the cloud computing module;
the cloud computing module is used for receiving and processing the weight, the load predicted value and the unit load predicted value sent by the edge computing module of different areas of the building, judging and classifying models of the different areas, and aggregating the weight of the models classified into one type by using a transverse federal learning method; aggregating the network load predicted values with different weights to establish a regression prediction model by utilizing a longitudinal federal learning method, and predicting the cold load of the building space unit according to the regression prediction model; and transmitting the aggregated model weight values of different positions back to the edge calculation module of different positions.
The cloud computing module obtains training results of the edge computing modules in different regions and returns model training parameters, so that the cost of communication facilities and data transmission is reduced, the model computing speed is increased, and corresponding control strategies are guaranteed to be made and executed in real time.
Furthermore, after the data are acquired by the data acquisition modules in a certain zone bit and converted into digital quantity, the model input source data are sent to the edge calculation module, the edge calculation module preprocesses the received data after receiving the data sent by the data acquisition module and trains or predicts the data, and after the edge calculation module processes the data, the cloud calculation module aggregates the processing results of different zone bits and reversely sends the aggregated information to the edge calculation module.
Further, the data collection module collects data associated with predicting air conditioner cooling load. The data acquisition module acquires the outdoor air temperature, the indoor pedestrian flow, the solar radiation intensity, the relative humidity and the air conditioner cold load once per minute. The outdoor air temperature acquisition unit, the solar radiation intensity acquisition unit and the relative humidity acquisition unit of the data acquisition module of one zone are uniformly distributed in the zone to acquire data of different positions in the same area, the indoor people flow rate acquisition unit counts the people flow rate in the zone, and the air conditioner cold load acquisition unit acquires the air conditioner cold load consumed by the zone. And after the data acquisition is finished, the data acquisition module converts the data into digital quantity and transmits the digital quantity to the edge calculation module of the corresponding zone bit.
Furthermore, the edge calculation module receives and processes the data collected by the corresponding zone bit and predicts the data. The edge calculation module comprises a data preprocessing unit, a neural network model and a weight uploading judgment unit. After the edge calculation module receives the source data sent by the data acquisition module, the data is preprocessed, and the preprocessed data is used as input data of a neural network model, so that parameters of the neural network are optimized. After the judgment processing of the aggregation value of the weight uploading judgment unit, the edge calculation module sends the network weight with the aggregation value and the load predicted value to the cloud calculation module, and receives weight adjustment data sent reversely by the cloud calculation module.
Furthermore, the cloud computing module comprises a model similarity judging unit, a model aggregation unit and a multiple regression model. And when the cloud computing module receives the data sent by the edge computing module, judging the zone bits of the same type of prediction models according to the received updated weights, and aggregating the updated weights and reversely sending the aggregated weights to the edge computing unit by using a transverse federal learning method aiming at the same type of prediction models. And aiming at the load predicted values of different types of prediction models, determining the contribution degrees of the predicted values of the different zone models to the total air-conditioning load by establishing a regression prediction model, and obtaining the total building area and calculating the predicted value of the cold load of the building space unit after the total air-conditioning load is calculated by the cloud computing module.
Furthermore, the building space unit cold load prediction system based on federal learning determines parameters of all prediction models according to historical data, and then performs real-time parameter adjustment along with operation of the heating and ventilation system and predicts the building space unit cold load in real time.
Further, the edge calculation module calculates once according to the sampling data per hour, the data acquisition unit sends the acquired data to the edge calculation module after completing data acquisition for 60 times, after the edge calculation module calculates k times, the optimized weight and the predicted value are transmitted to the cloud calculation module, and the cloud calculation module aggregates the weight data and returns the aggregated weight to the edge calculation module.
Furthermore, the single edge computing module is responsible for data processing of one zone, a building requiring unit cooling load prediction is divided into a plurality of typical zones, when the difference between the selected zone and the zone is large, the selected zone is determined as the selected typical zone, and the selected zone may not cover the whole building.
Further, the temperature acquisition unit is a temperature sensor.
Further, the solar radiation collection unit is a photosensitive sensor.
Further, the relative humidity acquisition unit is a humidity sensor.
Furthermore, the indoor people flow rate acquisition unit is an infrared sensor, the sensor is arranged at a passage port where people enter and exit corresponding areas, and people flow rate is obtained by calculating the shielded times and time.
Further, the air conditioner cold load acquisition unit is a temperature sensor and an air speed detector. And acquiring the cold load of the air conditioner by acquiring the temperature difference between the indoor temperature and the air outlet of the air conditioner at the zone, the cross section area of the air inlet and the air speed.
Further, the edge computing module is an MCU, and the cloud computing module is an MCU.
In conclusion, the invention has the following beneficial effects:
according to the building space unit cold load prediction method based on federal learning, the problem of difference of prediction models of different locations of a building is considered, the models of the different locations are identified through a neural network, the contribution degree of network model outputs of different types to the total load is searched through a multiple regression model, the prediction speed is improved compared with single network prediction, and the method is high in generalization. The neural network training efficiency is improved through a federal learning mechanism, the communication cost is reduced, and when the convergence abnormal condition is found during training, the stability of convergence is enhanced by adjusting the training method.
The building space unit cold load prediction system based on federal learning provided by the invention reduces the communication pressure of the sensing equipment and the calculation unit. The real-time performance of the prediction process is enhanced through the cloud edge cooperative computing architecture. Only the typical zone bit is selected to arrange the data acquisition module and the edge calculation module, so that the hardware cost is saved.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a system framework diagram in an embodiment of the invention;
FIG. 2 is a flow chart of a method in an embodiment of the invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
As shown in fig. 1, a building space unit cold load prediction system based on federal learning according to an embodiment of the present invention includes a data acquisition module, an edge calculation module, and a cloud calculation module.
The data acquisition module is used for acquiring data related to the prediction of the cold load of the building space units in different regions of the building;
the edge calculation module is used for receiving and processing data collected by corresponding zone bits, training the weight of a building space unit cold load prediction model of the zone bit through a neural network model of the zone bit, and calculating a zone bit air conditioner load prediction value and a corresponding zone bit unit air conditioner load prediction value; judging whether the model weight of the local zone bit has the aggregation value uploaded to the cloud computing module;
the cloud computing module is used for receiving and processing the weight, the load predicted value and the unit load predicted value sent by the edge computing module of different areas of the building, judging and classifying models of the different areas, and aggregating the weight of the models classified into one type by using a transverse federal learning method; aggregating the network load predicted values with different weights to establish a regression prediction model by utilizing a longitudinal federal learning method, and predicting the cold load of the building space unit according to the regression prediction model; the model weight of the aggregated different zone bits is transmitted back to the edge calculation module of the different zone bits;
each zone bit of the building corresponds to one edge calculation module, each edge calculation module is connected with a plurality of data acquisition modules, and the plurality of edge calculation modules are connected with one cloud calculation module.
The single edge calculation module is responsible for data processing of one zone bit, a building needing unit cold load prediction is divided into a plurality of typical zone bits, when the difference between the selected zone bit and the zone bit is large, the selected zone bit is determined to be selected as the typical zone bit, and the selected zone bit can not cover the whole building.
The data acquisition module located in the same zone acquires the outdoor air temperature, the indoor pedestrian volume, the solar radiation intensity, the relative humidity and the air conditioner cold load once per minute. The outdoor air temperature acquisition unit, the solar radiation intensity acquisition unit and the relative humidity acquisition unit of the data acquisition module in one area are uniformly distributed in the area to acquire data at different positions in the same area, the indoor people flow rate acquisition unit counts the people flow rate in the area, and the air conditioner cold load acquisition unit acquires the air conditioner cold load consumed by the area. Wherein, the temperature acquisition unit is a temperature sensor. The solar radiation collecting unit is a photosensitive sensor. The relative humidity acquisition unit is a humidity sensor. The indoor flow rate acquisition unit is an infrared sensor, the sensor is arranged at a passage port where people enter and exit corresponding zone bits, and the flow rate is acquired by calculating the shielded times and time. The air conditioner cold load acquisition unit is a temperature sensor and a wind speed detector. And acquiring the air conditioner cooling load by acquiring the temperature difference between the indoor temperature and the air outlet of the air conditioner at the zone, the cross section area of the air outlet and the air speed.
And aiming at the air conditioner cold load at the predicted time T, acquiring the following data related to the predicted building space unit cold load of different regions of the large building every hour: the outdoor air temperature at the current time T, the outdoor air temperature at the last time T-1, the indoor pedestrian volume at the current time T, the indoor pedestrian volume at the last time T-1, the solar radiation intensity at the current time T, the solar radiation intensity at the last time T-1, the relative humidity at the current time T, the unit air conditioner cooling load at the last time T-1, the air conditioner cooling load T-2 two times before and the air conditioner cooling load at the same time T-24 the previous time. There are 60 sets of data for each type acquired during a prediction period, with one minute sample between each set of data.
The data acquisition module acquires enough data and then sends the data to the edge calculation module of the corresponding zone bit, and the edge calculation module is an MCU. And the edge calculation module is used for preprocessing the data after receiving the data, wherein the preprocessing comprises the elimination and dimensionless of the error sampling value.
The method for eliminating the error sampling data comprises the following steps:
first, a data point is collected every minute, and the statistic F of the sampled data in a 1-hour sampling period is obtained:
Figure BDA0003078007850000091
wherein s isiIs the ith sample point and is the last sample point,
Figure BDA0003078007850000092
is the average of the sampling points and is,
Figure BDA0003078007850000093
the regression fit for the ith sample point. The regression fit value is obtained from a regression equation of a linear function, and the parameters of the regression equation are calculated by a least square method.
Obtaining statistic and judging FαSize of (1,58), Fα(1,58) represents the threshold value of the rejection region with a degree of freedom of 1 and a data point of 60, when F > Fα(1,58), there is no need to eliminate data, when F is less than or equal to FαAnd (1,58), removing the point with the maximum fitting value difference value with the sampling point from the sampling points, and judging as above.
The method for nondimensionalizing data comprises the following steps:
using formulas
Figure BDA0003078007850000094
And realizing the dimensionless of the sampling points.
Wherein n is1The number of sampling points after sampling error data is eliminated.
The data are preprocessed through the steps and then serve as input data of the BP neural network. And adjusting the weight of the BP neural network by using a particle swarm optimization method.
Firstly, the individual length of the particles is constructed according to the structural characteristics of a BP neural network:
L＝S1*S2+S2*S3+S2+S3 (2)
wherein S is1，S2And S3Node parameters of an input layer, a hidden layer and an output layer of the BP neural network are respectively.
Secondly, calculating the particle fitness V, wherein the fitness is calculated by adopting the following formula:
Figure BDA0003078007850000095
wherein n is2Is the number of sample point classes, yiIs the true value of the cooling load of the ith time sequence unit, oiIs the predicted value of the cooling load of the ith time sequence unit.
The fitness of the different particles is then compared and the particle with the best fitness is recorded as the best particle individual for this iteration of the population.
And finally, updating the speed and the position of each particle, wherein the updating formula is as follows:
Figure BDA0003078007850000101
Figure BDA0003078007850000102
wherein,
Figure BDA0003078007850000103
respectively the speed and position of the ith particle in the t iteration, u is the inertia weight, c1And c2As an acceleration factor, c3As disturbance factor, c3Usually a constant between (0,1),
Figure BDA0003078007850000104
for the extreme value of the ith particle for the t iteration,
Figure BDA0003078007850000105
the extreme value of the total particles in the t-th iteration,
Figure BDA0003078007850000106
for the position of the ith particle in the t iteration, r1And r2Is a random number between (0, 1).
And repeating the steps through iteration until the network weight is converged, and predicting the predicted value of the air conditioning load and the predicted value of the unit air conditioning load once the weight is updated.
After k rounds of calculation are carried out on the edge calculation modules of different zone bits, whether the updated weight has a polymerization value is judged, and when the weight update value meets the following formula, the polymerization value is considered to be present:
Figure BDA0003078007850000107
Figure BDA0003078007850000108
Figure BDA0003078007850000109
wherein, wtThe updated weight value of the Tth iteration of the position,
Figure BDA00030780078500001010
weight gradient for the tth iteration update of the location, mhAs the number of history weights used for fitting, mhThe value is more than 5, m is the total number of zone bits, alpha is a learning rate, and alpha is usually (0,1)]Is constant.
And after the judgment criterion is met, the edge calculation unit sends the model weight, the air conditioner load predicted value and the unit air conditioner load predicted value to the cloud calculation module.
The cloud computing module is an MCU (microprogrammed control unit), and aggregates the model weights after receiving the model weights, the air conditioner load predicted values and the unit air conditioner load predicted values sent by different zones.
Firstly, judging the model coincidence conditions from different regions:
the method for judging whether the models are overlapped comprises the following steps: arranging the weight historical values from different zone bits according to a time sequence:
Figure BDA00030780078500001011
wherein,
Figure BDA00030780078500001012
the weight vector of the t iteration of the jth zone bit, h is the number of historical weights used for judging the relevance, and the value of h is larger than 10.
Then, the weight value is normalized:
Figure BDA0003078007850000111
further, a differencing sequence is generated for the compared locational network weights:
Figure BDA0003078007850000112
finally, judging the relevance degree of the compared locational network weight:
Figure BDA0003078007850000113
go through the above steps to mv-1-pass formation of a square matrix of correlation degrees between weights of different locations，mvFor the number of the zone bits with aggregation value, the zone bit weight values in the square matrix which are higher than a relevance threshold value are classified as a class model weight value. The value rule of the relevance threshold is shown in table 1.
TABLE 1
Correlation coefficient Degree of correlation
0.8～1.0 Very strong correlation
0.6～0.8 Strong correlation
0.4～0.6 Medium intensity correlation
0.2～0.4 Weak correlation
0.0～0.2 Very weak or no correlation
When the weight models of the same class are judged, the weights from different regions are aggregated by utilizing a transverse federal learning technology, and the weights of the model of the same class are aggregated by utilizing a parameter averaging method:
Figure BDA0003078007850000114
wherein,
Figure BDA0003078007850000115
is the weight of the aggregation after the t-th iteration, msThe number of zone bits belonging to the same class model.
And if the models of all the zones are the same, averaging the received unit air conditioner load predicted values to obtain the unit air conditioner load predicted values as the building space unit cold load predicted values of the large-scale building.
If after multiple iterations, it appears
Figure BDA0003078007850000116
And when the convergence is not achieved, switching the aggregation algorithm. The method for judging whether convergence occurs is that after a plurality of iterations,
Figure BDA0003078007850000117
whether it is a constant value with noise. When the convergence is judged not to be converged, the aggregation method is switched to:
Figure BDA0003078007850000121
where λ is a velocity constant, typically a constant between (0, 1).
When the cloud computing unit detects that two or more than two model weights exist, the weights belonging to the same model are searched and aggregated through a formula (13), and whether the weights belong to the same model or not is judged through comparing the weights with a relevance threshold through a computing formula (12). Regarding the air-conditioning cold load prediction data from different models as a multiple regression independent variable of the total air-conditioning cold load of the building, establishing a multiple linear regression model of the air-conditioning cold load prediction data and the building total air-conditioning cold load:
Figure BDA0003078007850000122
wherein, betalRegression coefficient, x, for predicted values of class I zone bitslIs the sum of predicted values of air conditioner load in class I zone0Constant term coefficients for regression modelsε is the noise term, N is a Gaussian distribution, σ is the standard deviation of white noise, mkFor different types of zone numbers, y is the actual total air conditioning load of the building.
Step S42: the regression coefficient is obtained by the following formula:
Figure BDA0003078007850000123
wherein m islFor the number of data used to determine the sum of the historical air conditioning load predictions for each location of the regression coefficients,
Figure BDA0003078007850000124
is the l-th region of the l-th class1Sum of predicted values of load of the wheel air conditioners.
After the regression model coefficient is determined by the method, the air conditioner predicted load of the whole building is calculated, the indoor building area of the whole building is obtained, and the ratio of the air conditioner predicted load of the whole building to the indoor building area is the building space unit cold load predicted value. And then the cloud computing module replaces the neural network weight of the corresponding zone with the aggregated model weight of the different zone.
The prediction system firstly determines parameters of all prediction models according to historical data, and then carries out real-time parameter adjustment along with the operation of the heating and ventilation system and predicts the prediction value of the cold load of the building space unit.
The present embodiment is only for explaining the present invention, and it is not limited to the present invention, and those skilled in the art can make modifications without inventive contribution to the present embodiment as required after reading the present specification, but all of them are protected by patent law within the scope of the present invention.

Claims (8)

1. A building space unit cold load prediction method based on federal learning is characterized by comprising the following steps:
s1: acquiring data related to different locations of a building and prediction of cold load of a building space unit;
s2: respectively training the weight of a building space unit cold load prediction model of the corresponding zone through neural network models distributed in different zones according to the data obtained in the step S1, and calculating the predicted value of the air conditioner load of the corresponding zone and the predicted value of the air conditioner load of the unit of the corresponding zone;
s3: obtaining model weight, air conditioner load predicted value and unit air conditioner load predicted value of the zone bit with aggregation value, judging and classifying models of different zone bits, and aggregating the weight values of the models classified into one type by using a transverse federal learning method; the judgment method of the zone bit with the aggregation value comprises the following steps:
when the calculated weight update values of the neural network models deployed in different regions meet the following formula, the aggregation value is considered to be possessed:
Figure FDA0003560089100000011
Figure FDA0003560089100000012
Figure FDA0003560089100000013
wherein, wtThe weight value of the t-th iteration update of the position,
Figure FDA0003560089100000014
weight gradient for the tth iteration update of the location, mhAs the number of historical weights used for fitting, mhThe value is greater than 5, m is the total number of zone bits, alpha is the learning rate, and alpha is (0,1)]A constant value between;
the method for judging and classifying models of different locations comprises the following steps:
firstly, arranging the historical values of the weights from different zone bits according to a time sequence:
Figure FDA0003560089100000015
wherein,
Figure FDA0003560089100000016
the weight vector of the t iteration of the jth zone bit is h, the number of historical weights for judging the relevance is h, and the value of h is more than 10;
the weights are then normalized:
Figure FDA0003560089100000021
further, a differencing sequence is generated for the compared locational network weights:
Figure FDA0003560089100000022
wherein
Figure FDA0003560089100000023
Is the normalized weight vector of the compared location o;
finally, judging the correlation degree with the compared locational network weight:
Figure FDA0003560089100000024
traverse the above formula through mv-1 pass forming a square matrix of correlation degrees between weights of different locations, mvClassifying zone bit weights in the square matrix, which are higher than a relevance threshold value, into a class model weight for the number of zone bits with aggregation value;
after judging that the models are the same type, adopting the following method to aggregate the weights of the models of the same type:
Figure FDA0003560089100000025
wherein,
Figure FDA0003560089100000026
is the weight of the aggregation after the t-th iteration, msThe number of zone bits belonging to the same type of model;
s4: aggregating the network load predicted values with different weights to establish a regression prediction model by utilizing a longitudinal federal learning method, and predicting the cold load of the building space unit according to the regression prediction model; the regression prediction model establishing method comprises the following steps:
step S41: regarding the air conditioner cooling load prediction data from different models as a multiple regression independent variable of the building total air conditioner cooling load, establishing a multiple linear regression model of the air conditioner cooling load prediction data and the building total air conditioner cooling load:
Figure FDA0003560089100000027
wherein, betalRegression coefficient, x, for predicted values of class I zone bitslIs the sum of predicted values of air conditioning load in class I zone, beta0Is the coefficient of the constant term of the regression model, ε is the noise term, N is a Gaussian distribution, σ is the standard deviation of white noise, mkThe number of different types of zone bits is determined, and y is the actual total air conditioning load of the building;
step S42: the regression coefficient is obtained by the following formula:
Figure FDA0003560089100000028
wherein,
Figure FDA0003560089100000029
is the first1Actual total air conditioner load m of buildinglFor the number of data used to determine the sum of the historical air conditioning load predictions for each location of the regression coefficients,
Figure FDA0003560089100000038
is the l-th region of the l-th class1The sum of the predicted load values of the wheel air conditioners;
s5: and updating the neural network model of the corresponding zone according to the model weight of the aggregated different zones.
2. The method for predicting the cold load of the building space unit based on the federal study as claimed in claim 1, wherein the data acquired from different regions are respectively preprocessed and processed into input data which can be directly input into a neural network model; the preprocessing process is to eliminate sampling error data, take the average value of sampling points as the representative value of the prediction period, and then make the sampling data dimensionless by a normalization method; the method for eliminating the sampling error data comprises the following steps:
first, a data point is collected every minute, and the statistic F of the sampled data in a 1-hour sampling period is obtained:
Figure FDA0003560089100000031
wherein s isiIs the ith sample point and is the last sample point,
Figure FDA0003560089100000032
is the average of the sampling points and is,
Figure FDA0003560089100000033
calculating a regression equation parameter by a least square method, wherein the regression fitting value is the regression fitting value of the ith sampling point and is obtained by a regression equation of a linear function;
obtaining statistic F and then judging the statistic F and FαSize of (1,58), Fα(1,58) represents the threshold value of the rejection region with a degree of freedom of 1 and a data point of 60, when F > Fα(1,58), eliminating data, and when F is less than or equal to Fα(1,58), removing the point with the maximum regression fitting value difference value of the sampling points, and repeating the steps asJudging;
using formulas
Figure FDA0003560089100000034
Realizing the dimensionless of the sampling points;
wherein n is1The number of sampling points after sampling error data is eliminated.
3. The method for predicting the cold load of the building space unit based on the federal learning as claimed in claim 1, wherein the neural network model has a BP neural network structure, and the weight of the BP neural network is updated in a particle swarm optimization manner; the particle fitness V is calculated by the following formula:
Figure FDA0003560089100000035
wherein n is2Is the number of sample point classes, yiIs the true value of the cooling load of the ith time sequence unit, oiThe predicted value of the cooling load of the ith time sequence unit is;
the velocity and position update formula of the particle is:
Figure FDA0003560089100000036
Figure FDA0003560089100000037
wherein,
Figure FDA0003560089100000041
respectively the speed and position of the ith particle in the t iteration, u is the inertia weight, c1And c2As an acceleration factor, c3As disturbance coefficient, c3Taking a constant between (0,1), Pi tFor the poles of the ith particle of the t-th iterationThe value of the one or more of the one,
Figure FDA0003560089100000042
is the extreme value of the total particle in the t-th iteration, r1And r2Is a random number between (0, 1).
4. The method as claimed in claim 1, wherein the method includes the step of predicting the cold load of the building space unit based on federal learning if the method is performed after a plurality of iterations
Figure FDA0003560089100000043
When the convergence is not achieved, the polymerization method is switched to:
Figure FDA0003560089100000044
wherein, λ is a speed constant, and a constant between (0,1) is taken.
5. A federal learning based building space unit cold load prediction system, comprising:
the data acquisition module is used for acquiring data related to the cold load of the building space unit predicted at different locations of the building;
the edge calculation module is used for receiving and processing data collected by corresponding zone bits, training the weight of a building space unit cold load prediction model of the zone bit through a neural network model of the zone bit, and calculating a zone bit air conditioner load prediction value and a zone bit unit air conditioner load prediction value; judging whether the model weight of the local zone bit has the aggregation value uploaded to the cloud computing module; the process of judging the zone bit aggregation value is as follows:
when the calculated weight update values of the neural network models deployed in different regions meet the following formula, the aggregation value is considered to be possessed:
Figure FDA0003560089100000045
Figure FDA0003560089100000046
Figure FDA0003560089100000047
wherein, wtThe weight value of the t-th iteration update of the position,
Figure FDA0003560089100000048
weight gradient for the tth iteration update of the location, mhAs the number of history weights used for fitting, mhThe value is more than 5, m is the total number of zone bits, alpha is the learning rate, and alpha is (0,1)]A constant value between;
the cloud computing module is used for receiving and processing the weight, the load predicted value and the unit load predicted value sent by the edge computing module of different areas of the building, judging and classifying models of the different areas, and aggregating the weight of the models classified into one type by using a transverse federal learning method; aggregating the network load predicted values with different weights to establish a regression prediction model by utilizing a longitudinal federal learning method, and predicting the cold load of the building space unit according to the regression prediction model; the model weight of the aggregated different zone bits is transmitted back to the edge calculation module of the different zone bits;
the model judgment and classification process of different regions is as follows:
firstly, arranging the historical values of the weights from different zone bits according to a time sequence:
Figure FDA0003560089100000051
wherein,
Figure FDA0003560089100000052
for the t iteration of the jth locationA weight vector, h is the number of historical weights used for judging relevance, and the value of h is more than 10;
the weights are then normalized:
Figure FDA0003560089100000053
further, a differencing sequence is generated for the compared locational network weights:
Figure FDA0003560089100000054
wherein
Figure FDA0003560089100000055
Is the normalized weight vector of the compared location o;
finally, judging the relevance degree of the compared locational network weight:
Figure FDA0003560089100000056
traverse the above formula through mv1 pass formation of a matrix of correlations between weights of different locations, mvClassifying zone bit weights above a relevance threshold in the square matrix as class model weights for the number of zone bits having aggregation value;
after judging the models to be the same type, adopting the following method to aggregate the weights of the models of the same type:
Figure FDA0003560089100000057
wherein,
Figure FDA0003560089100000058
is the weight of the aggregation after the t-th iteration, msThe number of zone bits belonging to the same type of model;
the regression prediction model is established as follows:
regarding the air-conditioning cold load prediction data from different models as a multiple regression independent variable of the total air-conditioning cold load of the building, establishing a multiple linear regression model of the air-conditioning cold load prediction data and the building total air-conditioning cold load:
Figure FDA0003560089100000059
wherein, betalRegression coefficient, x, for predicted values of class I zone bitslIs the sum of predicted values of air conditioner load in class I zone0Is the constant term coefficient of the regression model, ε is the noise term, N is a Gaussian distribution, σ is the standard deviation of white noise, mkThe number of different types of zone bits, and y is the actual total air conditioning load of the building;
and (3) solving a regression coefficient, wherein the formula is as follows:
Figure FDA0003560089100000061
wherein,
Figure FDA0003560089100000062
is the first1Actual total air conditioner load m of buildinglFor the number of data used to determine the sum of the historical air conditioning load predictions for each location of the regression coefficients,
Figure FDA0003560089100000063
is the l-th region of the l-th class1The sum of the predicted load values of the wheel air conditioners;
each zone of the building corresponds to one edge calculation module, each edge calculation module is connected with a plurality of data acquisition modules, and the plurality of edge calculation modules are connected with one cloud calculation module.
6. The federal learning based building space unit cooling load prediction system as claimed in claim 5, wherein: after the data are collected by the data collection modules in a certain zone and converted into digital quantity, model input source data are sent to the edge calculation module, the edge calculation module preprocesses the received data after receiving the data sent by the data collection module and trains or predicts the data, and after the edge calculation module processes the data, the cloud calculation module aggregates processing results of different zones and reversely sends the aggregated information to the edge calculation module.
7. The federal learning based building space unit cooling load prediction system as claimed in claim 5, wherein: the edge calculation module comprises a data preprocessing unit, a neural network model and a weight uploading judgment unit; after the edge calculation module receives the source data sent by the data acquisition module, preprocessing the data, and using the preprocessed data as input data of a neural network model so as to optimize parameters of the neural network; after the judgment processing of the aggregation value of the weight uploading judgment unit, the edge calculation module sends the network weight with the aggregation value and the load predicted value to the cloud calculation module, and receives weight adjustment data sent reversely by the cloud calculation module.
8. The federal learning based building space unit cooling load prediction system as claimed in claim 5, wherein: the cloud computing module comprises a model similarity distinguishing unit, a model aggregation unit and a multiple regression model; after the cloud computing module receives the data sent by the edge computing module, judging the zone bits of the prediction models of the same category according to the received updated weights, and aggregating the updated weights and reversely sending the aggregated weights to the edge computing unit by using a transverse federal learning method aiming at the prediction models of the same category; and aiming at the load predicted values of different types of prediction models, determining the contribution degrees of the predicted values of the different zone models to the total air-conditioning load by establishing a regression prediction model, and obtaining the total building area and calculating the predicted value of the cold load of the building space unit after the total air-conditioning load is calculated by the cloud computing module.
CN202110562321.0A 2021-05-21 2021-05-21 Building space unit cold load prediction method and system based on federal learning Active CN113240184B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110562321.0A CN113240184B (en) 2021-05-21 2021-05-21 Building space unit cold load prediction method and system based on federal learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110562321.0A CN113240184B (en) 2021-05-21 2021-05-21 Building space unit cold load prediction method and system based on federal learning

Publications (2)

Publication Number Publication Date
CN113240184A CN113240184A (en) 2021-08-10
CN113240184B true CN113240184B (en) 2022-06-24

Family

ID=77138305

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110562321.0A Active CN113240184B (en) 2021-05-21 2021-05-21 Building space unit cold load prediction method and system based on federal learning

Country Status (1)

Country Link
CN (1) CN113240184B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114331761B (en) * 2022-03-15 2022-07-08 浙江万胜智能科技股份有限公司 Equipment parameter analysis and adjustment method and system for special transformer acquisition terminal
CN115037618B (en) * 2022-06-06 2023-11-07 电子科技大学 Lightweight edge intelligent collaborative federal learning platform based on KubeEdge
CN115773562A (en) * 2022-11-24 2023-03-10 杭州经纬信息技术股份有限公司 Unified heating ventilation air-conditioning system fault detection method based on federal learning
CN116187598B (en) * 2023-04-28 2023-07-04 深圳市科中云技术有限公司 Building-based virtual power plant load prediction method

Citations (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106874581A (en) * 2016-12-30 2017-06-20 浙江大学 A kind of energy consumption of air conditioning system in buildings Forecasting Methodology based on BP neural network model
CN107480815A (en) * 2017-08-03 2017-12-15 国网河北省电力公司保定供电分公司 A kind of power system taiwan area load forecasting method
CN107704875A (en) * 2017-09-30 2018-02-16 山东建筑大学 Based on the building load Forecasting Methodology and device for improving IHCMAC neutral nets
CN111148066A (en) * 2018-11-06 2020-05-12 丰田自动车株式会社 Wireless communication assurance for networked vehicles in high network load scenarios
CN111401658A (en) * 2020-04-08 2020-07-10 西安建筑科技大学 Parallel cold load prediction method based on building space unit
WO2020192896A1 (en) * 2019-03-26 2020-10-01 Huawei Technologies Co., Ltd. Apparatus and method for hyperparameter optimization of a machine learning model in a federated learning system
CN112000988A (en) * 2020-08-28 2020-11-27 深圳前海微众银行股份有限公司 Factorization machine regression model construction method and device and readable storage medium
CN112183730A (en) * 2020-10-14 2021-01-05 浙江大学 Neural network model training method based on shared learning
CN112232608A (en) * 2020-12-16 2021-01-15 浙江大学 Regional building air conditioner short-term power consumption prediction method based on data driving
CN112611080A (en) * 2020-12-10 2021-04-06 浙江大学 Intelligent air conditioner control system and method based on federal learning
CN112634027A (en) * 2020-12-30 2021-04-09 杭州趣链科技有限公司 Self-adaptive federal parameter aggregation method for credit assessment of small and micro enterprises
CN112686456A (en) * 2020-12-31 2021-04-20 广东电网有限责任公司 Power load prediction system and method combining edge calculation and energy consumption identification
CN112800461A (en) * 2021-01-28 2021-05-14 深圳供电局有限公司 Network intrusion detection method for electric power metering system based on federal learning framework

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11651231B2 (en) * 2019-03-01 2023-05-16 Government Of The United States Of America, As Represented By The Secretary Of Commerce Quasi-systolic processor and quasi-systolic array
US20210042628A1 (en) * 2019-08-09 2021-02-11 International Business Machines Corporation Building a federated learning framework

Patent Citations (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106874581A (en) * 2016-12-30 2017-06-20 浙江大学 A kind of energy consumption of air conditioning system in buildings Forecasting Methodology based on BP neural network model
CN107480815A (en) * 2017-08-03 2017-12-15 国网河北省电力公司保定供电分公司 A kind of power system taiwan area load forecasting method
CN107704875A (en) * 2017-09-30 2018-02-16 山东建筑大学 Based on the building load Forecasting Methodology and device for improving IHCMAC neutral nets
CN111148066A (en) * 2018-11-06 2020-05-12 丰田自动车株式会社 Wireless communication assurance for networked vehicles in high network load scenarios
WO2020192896A1 (en) * 2019-03-26 2020-10-01 Huawei Technologies Co., Ltd. Apparatus and method for hyperparameter optimization of a machine learning model in a federated learning system
CN111401658A (en) * 2020-04-08 2020-07-10 西安建筑科技大学 Parallel cold load prediction method based on building space unit
CN112000988A (en) * 2020-08-28 2020-11-27 深圳前海微众银行股份有限公司 Factorization machine regression model construction method and device and readable storage medium
CN112183730A (en) * 2020-10-14 2021-01-05 浙江大学 Neural network model training method based on shared learning
CN112611080A (en) * 2020-12-10 2021-04-06 浙江大学 Intelligent air conditioner control system and method based on federal learning
CN112232608A (en) * 2020-12-16 2021-01-15 浙江大学 Regional building air conditioner short-term power consumption prediction method based on data driving
CN112634027A (en) * 2020-12-30 2021-04-09 杭州趣链科技有限公司 Self-adaptive federal parameter aggregation method for credit assessment of small and micro enterprises
CN112686456A (en) * 2020-12-31 2021-04-20 广东电网有限责任公司 Power load prediction system and method combining edge calculation and energy consumption identification
CN112800461A (en) * 2021-01-28 2021-05-14 深圳供电局有限公司 Network intrusion detection method for electric power metering system based on federal learning framework

Also Published As

Publication number Publication date
CN113240184A (en) 2021-08-10

Similar Documents

Publication Publication Date Title
CN113240184B (en) Building space unit cold load prediction method and system based on federal learning
CN102705957B (en) Method and system for predicting hourly cooling load of central air-conditioner in office building on line
CN106920006B (en) Subway station air conditioning system energy consumption prediction method based on ISOA-LSSVM
CN103912966B (en) A kind of earth source heat pump refrigeration system optimal control method
CN110332647B (en) Load prediction method for air conditioning system of underground station of subway and air conditioning system
CN102779228B (en) Method and system for online prediction on cooling load of central air conditioner in marketplace buildings
CN107392368B (en) Meteorological forecast-based office building dynamic heat load combined prediction method
CN111486554B (en) Air conditioner temperature non-sensitive control method based on online learning
Alamin et al. An Artificial Neural Network (ANN) model to predict the electric load profile for an HVAC system
CN111461921B (en) Load modeling typical user database updating method based on machine learning
CN115758912A (en) Air conditioner energy consumption optimizing system
CN112288157A (en) Wind power plant power prediction method based on fuzzy clustering and deep reinforcement learning
CN116205508A (en) Distributed photovoltaic power generation abnormality diagnosis method and system
CN111815039A (en) Weekly scale wind power probability prediction method and system based on weather classification
CN113887833A (en) Distributed energy user side time-by-time load prediction method and system
CN114200839A (en) Office building energy consumption intelligent control model for dynamic monitoring of coupled environmental behaviors
CN117553404A (en) Method and system for improving energy efficiency of large water-cooling central air conditioning system
Das et al. A study on the application of artificial intelligence techniques for predicting the heating and cooling loads of buildings
Yu et al. Predicting indoor temperature from smart thermostat and weather forecast data
Xiang et al. Prediction model of household appliance energy consumption based on machine learning
CN115563848A (en) Distributed photovoltaic total radiation prediction method and system based on deep learning
CN113177675A (en) Air conditioner cold load prediction method based on optimization neural network of longicorn group algorithm
CN115510767B (en) Regional air temperature prediction method based on depth space-time network
CN115879190B (en) Model construction method and device and building load prediction method and device
Chen et al. Integrated attention mechanism for GBDT building energy consumption prediction algorithm

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant