CN114239991A - Building heat supply load prediction method, device and equipment based on data driving - Google Patents

Building heat supply load prediction method, device and equipment based on data driving Download PDF

Info

Publication number
CN114239991A
CN114239991A CN202111604996.3A CN202111604996A CN114239991A CN 114239991 A CN114239991 A CN 114239991A CN 202111604996 A CN202111604996 A CN 202111604996A CN 114239991 A CN114239991 A CN 114239991A
Authority
CN
China
Prior art keywords
building
heat supply
data
building heat
prediction 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.)
Pending
Application number
CN202111604996.3A
Other languages
Chinese (zh)
Inventor
谢海鹏
汤凌峰
王晓阳
别朝红
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.)
Xian Jiaotong University
Original Assignee
Xian Jiaotong University
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 Xian Jiaotong University filed Critical Xian Jiaotong University
Priority to CN202111604996.3A priority Critical patent/CN114239991A/en
Publication of CN114239991A publication Critical patent/CN114239991A/en
Pending legal-status Critical Current

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
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • 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
    • 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
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/08—Construction

Landscapes

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

Abstract

The invention discloses a building heat supply load prediction method, a device and equipment based on data drive, which construct a characteristic set according to a building structure, user behaviors and environmental weather, form a data sample based on the characteristic set of different time scales, extract deep data characteristics of a plurality of different time scales in the data sample, establish a building heat supply prediction model based on an input sample and building heat supply power, adopt a neural network training framework of off-line training and on-line optimization, use an L1-RDA on-line learning method to carry out on-line optimization updating on the building heat supply prediction model, and update the building heat supply prediction model in real time by using dynamic data flow of actual heat supply power, so that the prediction model can better cope with uncertainty of the environmental weather and the user behaviors, enhance the robustness and the generalization capability of the model, and improve the accuracy of short-term building heat supply prediction, the method has important significance for deeply excavating the energy-saving potential of the building and constructing a novel building with nearly zero energy consumption.

Description

Building heat supply load prediction method, device and equipment based on data driving
Technical Field
The invention belongs to the field of building heat energy regulation, and particularly relates to a data-driven building heat supply load prediction method, device and equipment.
Background
Under the aim of double carbon, the improvement of energy utilization efficiency and the realization of energy transformation become common knowledge for the development of various fields in China. According to Chinese building energy consumption research report 2020 issued by the China building energy conservation Association in 2020, the operating energy consumption and the whole process energy consumption of the buildings in China respectively account for 21.7 percent and 46.5 percent of the total energy consumption of China, and the total energy consumption of the buildings can be further improved along with the deep development of industrialization and urbanization in China. Meanwhile, the wide application of the distributed photovoltaic in the building changes the energy consumption structure of the building, the capability of a user side for consuming new energy nearby is fully utilized, and the heat storage performance of the building provides a buffer area for the formulation of a building operation scheduling strategy. Therefore, the energy-saving potential of the building needs to be deeply excavated, the operation scheduling strategy of the building needs to be optimized, and a novel building with near-zero energy consumption is built.
The accuracy and reliability of short-term prediction of building heat supply are the basis for formulating and optimizing the operation scheduling strategy of the building, and have important significance for improving the building energy efficiency. Research methods on building heating prediction can be divided into two categories: physical models and data driven models. The physical model takes a thermal balance equation as a core, a thermodynamic dynamic model of the building is described by using a thermal resistance-thermal capacity network, a conduction-transfer function and the like, and the building heat supply power is predicted by combining a prediction model of the environmental temperature, a probability model of the user behavior and the like; the data driving model designs characteristics from multiple aspects such as environmental weather, user behaviors and historical heating power, and a mapping relation between the structural characteristics and the building heating power is established by utilizing a machine learning model.
However, in the building heat supply prediction method based on the physical model, the building structure is more complicated to model, the models of different building buildings are greatly different, the model is poor in popularization and universality, and the user behavior is simpler to describe, so that the prediction accuracy is relatively low. With the rapid development of the machine learning field and the wide access of a large number of intelligent electric meters at the building side, the data driving model has a good effect in the short-term prediction of building heat supply. However, the current data-driven model lacks of feature extraction of building structures, and only uses time attribute features to depict complex user behaviors, and data information of different time scales is not fully utilized. Meanwhile, the current data-driven model is directly used for short-term heat supply power prediction of a building after offline training, and the data-driven model applied online is not updated in real time by utilizing the dynamic data flow of the building heat supply power, so that the method is difficult to adapt to the randomness and variability of user behaviors.
Disclosure of Invention
The invention aims to provide a method, a device and equipment for predicting building heat supply load based on data driving, so as to overcome the defects of the prior art.
A building heating load prediction method based on data driving comprises the following steps:
s1, constructing a feature set according to the building structure, the user behavior and the environmental weather, and forming a data sample based on the feature set with different time scales;
s2, respectively extracting deep data characteristics of a plurality of different time scales in the data sample by using a bidirectional long-short term memory network, inputting the extracted deep data characteristics and the deep data characteristics of the current moment into a multi-layer perceptron together, and establishing a building heat supply prediction model based on the input sample and the building heat supply power;
s3, updating parameters of the building heat supply prediction model by adopting historical sample data in data samples with different time scales, and then performing online optimization updating on the building heat supply prediction model by adopting a dynamically generated building heat supply actual power data stream and using an L1-RDA online learning method;
and S4, acquiring a feature set of the corresponding building to train the optimized and updated building heat supply prediction model, and realizing building heat supply load prediction by using the trained building heat supply prediction model based on the dynamic data flow of the building heat supply actual power.
Further, the building structure-based characteristics include building equivalent heat loss coefficient and building transparent envelope area ratio, the user behavior-based characteristics include time attribute of the current time, the environment weather-based characteristics include external environment temperature, average illumination intensity, relative humidity and average wind speed, and the user-environment combined equivalent heat loss coefficient based on the user behavior and the environment weather.
Further, building heating power P at historical moment i is utilized based on a heat balance equationiAnd corresponding building interior air temperature
Figure BDA0003433399810000021
And the temperature of the external environment
Figure BDA0003433399810000022
Obtaining the heat loss coefficient at the current moment, and averagely obtaining the final building equivalent heat loss coefficient kbuildThe specific calculation formula is as follows:
Figure BDA0003433399810000023
in the formula, n is the total amount of the sampled historical time.
Further, the corresponding deep features are respectively extracted by adopting Bi-LSTM, and the output of the Bi-LSTM neural network at the time t is as follows:
Figure BDA0003433399810000024
in the formula, WforAnd WbackMapping matrices for the outputs of the forward and backward layers respectively,
Figure BDA0003433399810000025
and
Figure BDA0003433399810000026
the short-term memory of the layer output is hidden at the time t for the forward layer and the backward layer respectively.
Further, by taking the characteristics corresponding to the current moment and the sequence characteristics extracted from the plurality of different time scale data as input, establishing a mapping relation between the characteristics and the building heat supply power value by using an MLP, wherein a forward propagation formula of the l-th sensing layer in the MLP is as follows:
hl＝ELU(Wlhl-1+bl)
in the formula, hlFor the output of the first sensing layer, ELU function is the nonlinear function used by the sensing layer, WlA connection weight matrix for the l-th layer sensing layer, hl-1As input to the l-th sensing layer, blThe input bias of the sensing layer of the l-th layer.
Further, a building heat supply prediction model is optimized by taking the mean square error as a loss function, and the specific calculation formula of the mean square error loss function L is as follows:
Figure BDA0003433399810000031
in the formula, X is an input matrix formed by N training samples, N is the total number of input training samples, theta is a parameter of the neural network, and yiFor the actual heating power corresponding to the ith training sample,
Figure BDA0003433399810000032
predicted heating power, x, for the ith training sampleiIs the ith training sample.
Further, the output power of a building heating system is collected at intervals, corresponding training samples are constructed, and on the basis of an online learning method, the deep characteristics of data flow are mined and the parameters of a heating prediction network are optimized in real time with the aim of minimizing accumulated online loss.
Further, an L1-RDA online learning algorithm is a method for solving the optimal parameters of the neural network online, and the optimization target of the L1-RDA algorithm at the t-th time step is as follows:
Figure BDA0003433399810000033
in the formula, thetatNeural network parameters updated for the t-th time step, GrFor the gradient at the r-th time step,<Gr,θ>is a gradient GrThe median of the integrals for θ, γ and λ are the parameters of the L1-RDA algorithm.
A data-driven building heating load prediction system, comprising:
the data sample acquisition module is used for constructing a feature set according to a building structure, user behaviors and environmental weather and forming data samples based on the feature sets with different time scales;
the building heat supply prediction module is used for respectively extracting deep data characteristics of a plurality of different time scales in the data sample based on the bidirectional long-short term memory network, inputting the extracted deep data characteristics and the deep data characteristics of the current moment into the multilayer perceptron together, and establishing a building heat supply prediction model based on the input sample and the building heat supply power;
the network optimization module is used for updating parameters of the building heat supply prediction model by adopting historical sample data in data samples with different time scales, and then performing online optimization updating on the building heat supply prediction model by adopting a dynamically generated building heat supply actual power data stream and using an L1-RDA online learning method;
and the prediction module is used for training the optimized and updated building heat supply prediction model according to the characteristic set of the corresponding building, and realizing building heat supply load prediction by using the trained building heat supply prediction model and based on the dynamic data flow of the building heat supply actual power.
A terminal device comprising a memory, a processor and a computer program stored in said memory and executable on said processor, said processor implementing the steps of a data-driven building heating load prediction method when executing said computer program.
Compared with the prior art, the invention has the following beneficial technical effects:
the invention relates to a building heat supply load prediction method based on data driving, which comprises the steps of constructing a feature set according to a building structure, user behaviors and environmental weather, and forming data samples based on the feature sets with different time scales; respectively extracting deep data characteristics of a plurality of different time scales in the data sample by using a bidirectional long-short term memory network, and the extracted deep data features and the deep data features at the current moment are input into the multi-layer perceptron together, building heat supply prediction models based on input samples and building heat supply power are established, a neural network training framework of off-line training and on-line optimization is adopted, the building heat supply prediction models are subjected to on-line optimization updating by using an L1-RDA on-line learning method, the building heat supply prediction models are updated in real time by using dynamic data flow of actual heat supply power, the prediction model can better cope with the uncertainty of environmental weather and user behavior, the robustness and generalization capability of the model are enhanced, the accuracy of short-term prediction of building heat supply is improved, the method has important significance for deeply excavating the energy-saving potential of the building and constructing a novel building with nearly zero energy consumption.
Furthermore, based on the statistical rule of the heat balance equation and the historical data, the building equivalent heat loss coefficient and the user-environment combined equivalent heat loss coefficient are used for respectively describing the physical structure of the building and the contribution of the user behavior and the environmental change to the heat supply requirement, so that the complex physical modeling of the building structure and the probability description of the uncertainty of the user behavior and the environmental change are avoided, and the method has strong popularization and applicability.
Furthermore, the Bi-LSTM network is used for analyzing sequence data at the same time within 6 hours of history and one week of history respectively, extracting the change trend characteristics of the data on different time scales and fully utilizing the autocorrelation of the history data on time. Meanwhile, the extracted deep features and the data information at the current moment are input to the multi-layer perceptron together, so that the used data information is enriched, and the accuracy of building heat supply power prediction is further improved.
Drawings
FIG. 1 is a diagram illustrating feature sets according to an embodiment of the present invention.
Fig. 2 is a schematic structural diagram of a Bi-LSTM network in an embodiment of the present invention.
Fig. 3 is a block diagram of the Bi-LSTM forward layer in an embodiment of the present invention.
FIG. 4 is a diagram of a building heating prediction model according to an embodiment of the present invention.
FIG. 5 is a schematic diagram of off-line training and on-line optimization of a building heat supply prediction model in an embodiment of the invention.
Detailed Description
The following examples are intended to further illustrate the present invention and should not be construed as limiting the scope of the invention, and other insubstantial modifications and adaptations of the invention by those skilled in the art based on the teachings herein are intended to be covered thereby.
Throughout the specification and claims, unless explicitly stated otherwise, the word "comprise", or variations such as "comprises" or "comprising", will be understood to imply the inclusion of a stated element or component but not the exclusion of any other element or component.
In order to achieve the purpose, the invention provides a building heat supply load prediction method based on data driving, which comprises the following steps:
step 1, building an equivalent heat loss coefficient of a building and a user-environment combined equivalent heat loss coefficient based on a statistical rule of a heat balance equation and historical data, building a feature set according to time attributes of the building at the same moment, the temperature of an external environment, relative humidity, average wind speed and average illumination intensity, and describing influence factors related to building heating power from three angles of a building structure, user behaviors and environmental weather. Considering the autocorrelation of historical data in time, three data samples with different time scales at the same time in the current time, the historical 6 hours and the historical week are finally formed;
step 2, respectively extracting deep data characteristics of two time scales at the same time in the historical 6 hours and the historical week by using a bidirectional long and short term memory network (Bi-LSTM), inputting the extracted deep data characteristics and the deep data characteristics of the current time into a multi-layer perceptron (MLP) together, and establishing a mapping relation between an input sample and building heat supply power to obtain a building heat supply prediction model based on the input sample and the building heat supply power;
step 3, in an off-line training stage, updating parameters of the building heat supply prediction model by adopting a batch-training Adam algorithm on the basis of the historical data set; in the online application stage, aiming at the dynamically generated building heat supply actual power data flow, an L1-RDA online learning method is used for online learning and updating a building heat supply prediction model;
and 4, collecting relevant data and constructing a training set aiming at a specific building, and training the building heat supply prediction model off line by using the training set.
And entering an online application stage after the offline training is finished, and updating the building heat supply prediction model in real time by using an online learning mode on the basis of the dynamic data stream of the building heat supply actual power.
A data-driven building heat supply load prediction method specifically comprises the steps of building a multi-time scale data set, building a building heat supply prediction model, performing off-line training and on-line optimization on the building heat supply prediction model, and testing and applying the building heat supply prediction model, wherein each part is as follows. It should be noted that the building object analyzed by the present invention is an office building or a teaching building, not a residential building, and the time interval for predicting the heat supply of the building is 15 minutes.
Construction of multi-timescale datasets
The invention constructs a feature set from three angles of a building structure, user behaviors and environmental weather, and is specifically shown in figure 1. The selected characteristics in the aspect of the building structure comprise building equivalent heat loss coefficient and building transparent envelope structure area ratio; the user behavior aspect is characterized by the time attribute (instant, week, month and quarter) at the current time; the characteristics selected in the aspect of environmental weather comprise external environment temperature, average illumination intensity, relative humidity and average wind speed, and a user-environment combined equivalent heat loss coefficient is added as the characteristics in the aspects of users and environments. The building equivalent heat loss coefficient and the user-environment combined equivalent heat loss coefficient are described below.
Building equivalent heat loss coefficient
Building self envelope has hindered the heat convection process of building inside air and outside environment air to a certain extent for building self has the heat storage nature of certain degree. Considering the complexity of building structures and the difference of different building structures, the invention uses the building equivalent heat loss coefficient to measure the heat storage capacity of the building.
The invention is based on a heat balance equation and utilizes the building heat supply power P of the historical moment iiAnd corresponding building interior air temperature
Figure BDA0003433399810000061
And the temperature of the external environment
Figure BDA0003433399810000062
Obtaining the heat loss coefficient at the current moment, and averagely obtaining the final building equivalent heat loss coefficient kbuildThe specific calculation formula is as follows:
Figure BDA0003433399810000063
in the formula, n is the total amount of the sampled historical time.
In order to avoid errors and interference caused by user behaviors and external environments to the building equivalent heat loss coefficient solving, historical data of the building from zero hour to five hours in the morning are only selected, meanwhile, the building lighting load and the external environment wind speed are monitored, if the historical data are larger than a threshold value, the user behaviors or the external environment wind speed at the moment can possibly influence the building thermal balance, and therefore the data at the moment are abandoned.
It should be noted that the building equivalent heat loss coefficient is only related to the physical structure of the building itself, and is considered to be a constant when the building is not modified. Meanwhile, in order to improve the reliability of the calculation result of the formula (1), the equivalent heat loss coefficient of the building is calculated by using historical data of a winter season.
User-environment combined equivalent heat loss coefficient
The user behavior and the environmental weather have strong randomness and variability, and have large uncertain influence on the heating power of the building. Therefore, the invention constructs the user-environment combined equivalent heat loss coefficient
Figure BDA0003433399810000064
The influence of the two is comprehensively measured, and the corresponding calculation formula is as follows.
Figure BDA0003433399810000065
It should be noted that the user-environment combined equivalent heat loss coefficient is a dynamic variable, and comprehensively describes the physiological heat dissipation of the user, the heat load of the user controlling the lighting equipment and the like, the user behavior of the user controlling the transparent building envelope structure and the like of the window and the like, and the influence of the external environment weather on the building heat balance.
After the characteristic set is constructed, the area occupation ratio of the building transparent enclosure structure, the corresponding time, week, month, quarter, external environment temperature, average illumination intensity, relative humidity, average wind speed and building equivalent heat loss coefficient at the future time are selected as the input of the building heat supply prediction model. Meanwhile, based on building heating power and a user-environment combined equivalent heat loss coefficient which correspond to the same time within 6 hours and one week in the history, sequence data of two different time scales are constructed and used as input. Therefore, the building heat supply prediction model provided by the invention comprises input data of the current time, the historical 6 hours and the same time in the historical week, namely three time scales.
Building heat supply prediction model
For history 6 hours and history within one week same timeThe present invention uses Bi-LSTM to extract the corresponding deep features. The structural diagram of Bi-LSTM is shown in fig. 2, in which,
Figure BDA0003433399810000071
indicating the LSTM cell at time t in the Bi-LSTM forward layer,
Figure BDA0003433399810000072
indicating the LSTM cell at time t in the Bi-LSTM backward layer, xtIs the input at time t of the Bi-LSTM neural network, ytIs the output of the Bi-LSTM neural network at the time t.
In the Bi-LSTM neural network, data information of a forward layer is propagated forwards along a time step, data information of a backward layer is propagated backwards along the time step, and the output of each time step depends on the output of the forward layer and the output of the backward layer. Therefore, the Bi-LSTM neural network can bidirectionally analyze the variation trend of the sequence data and better extract the dependency and correlation between the time step data information before and after extraction. The output of the Bi-LSTM neural network at time t is:
Figure BDA0003433399810000073
in the formula, WforAnd WbackMapping matrices for the outputs of the forward and backward layers respectively,
Figure BDA0003433399810000074
and
Figure BDA0003433399810000075
the short-term memory of the layer output is hidden at the time t for the forward layer and the backward layer respectively.
The forward layer unit structure of the Bi-LSTM neural network is shown in FIG. 3, and the unit structure of the backward layer is similar, and only the information propagation direction is opposite. The Bi-LSTM forward layer requires an input x according to the current time ttAnd short-term memory of the output of the hidden layer of adjacent time steps
Figure BDA0003433399810000076
Obtaining the output f of the three-door structure of the forgetting door, the input door and the output door at the moment tt for、
Figure BDA0003433399810000077
And
Figure BDA0003433399810000078
the specific calculation formula is as follows:
Figure BDA0003433399810000079
Figure BDA00034333998100000710
Figure BDA00034333998100000711
in the formula (I), the compound is shown in the specification,
Figure BDA00034333998100000712
Figure BDA00034333998100000713
and
Figure BDA00034333998100000714
respectively the connection weights of the current time step input and the three gate structures,
Figure BDA00034333998100000715
Wi forand
Figure BDA00034333998100000716
respectively the short-term memory output by the hidden layer at the adjacent time step and the connection weight of the three gate structures,
Figure BDA00034333998100000717
Figure BDA00034333998100000718
and
Figure BDA00034333998100000719
three door structure offsets respectively.
In the forward layer unit structure of Bi-LSTM neural network, the output of forgetting gate controls the long-term memory of adjacent time step
Figure BDA00034333998100000720
Degree of forgetting, output of input gate controlling long-term memory to be selected
Figure BDA00034333998100000721
The output of the output gate controls the long-term memory of the current time step
Figure BDA00034333998100000722
Retention of middle to short term memory
Figure BDA00034333998100000723
Part (c) of (a). Therefore, long-term memory
Figure BDA00034333998100000724
And short term memory
Figure BDA00034333998100000725
The specific calculation formula of the update is as follows:
Figure BDA0003433399810000081
Figure BDA0003433399810000082
Figure BDA0003433399810000083
in the formula (I), the compound is shown in the specification,
Figure BDA0003433399810000084
inputting the connection weight of the current time step and the long-term memory to be selected,
Figure BDA0003433399810000085
the connection weight of the short-term memory and the long-term memory to be selected in the adjacent time step,
Figure BDA0003433399810000086
biased for the corresponding input.
The method takes the characteristics corresponding to the current moment and the sequence characteristics extracted from two different time scale data as input, and establishes the mapping relation between the MLP and the building heat supply power value by using the MLP. The forward propagation formula of the l-th sensing layer in the MLP is as follows:
hl＝ELU(Wlhl-1+bl) (14)
in the formula, hlFor the output of the first sensing layer, ELU function is the nonlinear function used by the sensing layer, WlA connection weight matrix for the l-th layer sensing layer, hl-1As input to the l-th sensing layer, blThe input bias of the sensing layer of the l-th layer. It is noted that the last layer of the sensing layer does not use a non-linear function.
Therefore, the building heating prediction model structure provided by the invention is shown in figure 4. In consideration of the fact that an MLP model in the existing research is high in accuracy and good in performance, two characteristics of building equivalent heat loss coefficient and user-environment combined equivalent heat loss coefficient are added on the basis of keeping the basic structure of the MLP model, and deep characteristics of historical data under different time scales are extracted by using a Bi-LSTM network and used as model input.
Offline training and online optimization of building heat supply prediction model
Aiming at the parameter optimization process of the building heat supply prediction model, the invention uses the neural network learning route of off-line training-on-line optimization. In the off-line training stage, the invention uses a batch training mode and optimizes the parameters of the neural network based on fixed sample data; in the online optimization stage, the method uses an L1-RDA online learning algorithm, optimizes parameters of the neural network based on dynamic sample data flow, and improves the precision and the sparsity of the model. The overall schematic is shown in fig. 5.
Offline training of building heat supply prediction model
In the off-line training stage, historical data related to building buildings are collected, a data set is constructed, and the data set is divided into an off-line training set and an off-line testing set according to the proportion of eight to two. Based on an offline training set, the method takes the mean square error as a loss function to optimize a building heat supply prediction model. The specific calculation formula of the mean square error loss function L is:
Figure BDA0003433399810000087
in the formula, X is an input matrix formed by N training samples, N is the total number of input training samples, theta is a parameter of the neural network, and yiFor the actual heating power corresponding to the ith training sample,
Figure BDA0003433399810000088
predicted heating power, x, for the ith training sampleiIs the ith training sample.
Considering that the samples in the data set are static during off-line training, the method is based on a batch training mode, and the Adam algorithm is used for updating the parameters of the heat supply prediction neural network, and the specific process is as follows.
(1) Setting the number of samples of each batch training as N, and setting hyper-parameters alpha and beta of Adam algorithm1And beta2While initializing the parameter m0、v0And t0Is 0, epsilon is 10-8；
(2) And (3) calculating the corresponding loss function value of the training sample of the jth batch by using an equation (11), and obtaining the gradient of the loss function to the neural network parameter theta by using an error back propagation method
Figure BDA0003433399810000091
(3) Biased first moment estimate m in Adam algorithmj-1And biased second moment estimate vj-1Updating, wherein the calculation formula is as follows:
Figure BDA0003433399810000092
Figure BDA0003433399810000093
(4) m obtained according to the updatejAnd vjParameter θ to neural networkj-1Updating:
Figure BDA0003433399810000094
and (4) the steps (2) to (4) are executed circularly, and the building heat supply prediction network can be trained and optimized. And when the loss function value of the off-line test set does not drop any more, ending the off-line training stage of the building heat supply prediction network.
On-line optimization of building heat supply prediction model
In the on-line application stage, the output power of the building heating system is collected at intervals of 15 minutes, and a corresponding training sample is constructed. Since the samples of the online phase are generated continuously at 15-minute intervals, the batch training algorithm used in the offline training phase has difficulty updating the parameters of the heating prediction network with a single sample. Therefore, the method is based on online learning (online learning), aims at minimizing accumulated online loss, mines deep features of data flow, and optimizes parameters of a heat supply prediction network in real time.
The sparsity of the neural network model is considered to play a role in feature selection, and meanwhile, the calculation complexity of the heat supply power prediction process can be greatly reduced. Therefore, the invention uses the L1-RDA online learning algorithm to update the building heating prediction network in real time.
The L1-RDA online learning algorithm is essentially a method for solving the optimal parameters of the neural network online. The optimization goal of the L1-RDA algorithm at the t time step is as follows:
Figure BDA0003433399810000095
in the formula, thetatNeural network parameters updated for the t-th time step, GrFor the gradient at the r-th time step,<Gr,θ>is a gradient GrThe median of the integrals for θ, γ and λ are the parameters of the L1-RDA algorithm.
Equation (15) is decomposed into a plurality of independent optimization problems according to each characteristic dimension. The optimization objective of the ith dimension is as follows:
Figure BDA0003433399810000101
in the formula (I), the compound is shown in the specification,
Figure BDA0003433399810000102
for the updated neural network parameters of the ith dimension in the t-th time step,
Figure BDA0003433399810000103
is a gradient GrComponent of the ith feature dimension.
Due to | thetaiThe term is in thetaiNot conducting at 0, so xi is defined as | θiThe second derivative of | thus:
Figure BDA0003433399810000104
calculating a second derivative of the optimization target in the formula (16) and making the second derivative be 0, then judging whether the formula (16) can be 0 by taking the formula (17) as a reference, and finally updating the weight of the available L1-RDA online learning algorithm in each dimension in a way that:
Figure BDA0003433399810000105
in the formula, sgn is a sign function.
The specific steps of the L1-RDA online learning algorithm are set forth below.
(1) Setting parameters gamma and lambda of an L1-RDA algorithm, and initializing an accumulative gradient G0Is 0 and predicts the network parameter theta in the off-line phase when the training is completed0Is an initial parameter;
(2) for the samples generated at the t-th time step, the cumulative gradient value G will bet-1Is updated to GtMore particularly, the method comprises the following steps:
Figure BDA0003433399810000106
(3) the parameter values for each dimension of the heating prediction network are updated separately using equation (18).
And (4) circularly executing the step (2) and the step (3), and updating the parameters of the building heat supply prediction model by using the dynamically generated data stream, so that the capability of the prediction model for learning the uncertainty of the environmental weather and the user behavior on line is enhanced, and the accuracy of the prediction model is improved.
Test and application of building heat supply prediction model
Historical data such as heating power of a specific building are collected, and a data sample set which is characterized by multiple time scales is constructed. And respectively processing sequence data corresponding to different time scales by adopting two Bi-LSTM networks, then inputting the obtained deep features and the features of the current moment into an MLP network together, and predicting the building heat supply power value with the time interval of 15 minutes. In an off-line stage, according to the constructed data sample set, Adam algorithm based on batch training is adopted to update the heat supply prediction network parameters. After training is completed, in an online optimization stage, the invention uses an L1-RDA online learning algorithm to extract the information of the heating power data flow in real time and optimize a heating prediction network. The invention adopts the mean square error index MSE to evaluate the performance of the building heat supply prediction model, and the specific calculation formula is as follows:
Figure BDA0003433399810000111
in the formula (I), the compound is shown in the specification,
Figure BDA0003433399810000112
total heating power value, y, predicted for on-line phasesiIs the actual value of the ith heating power,
Figure BDA0003433399810000113
the predicted value of the ith heating power is.
In one embodiment of the present invention, a terminal device is provided that includes a processor and a memory, the memory storing a computer program comprising program instructions, the processor executing the program instructions stored by the computer storage medium. The processor is a Central Processing Unit (CPU), or other general purpose processor, Digital Signal Processor (DSP), Application Specific Integrated Circuit (ASIC), ready-made programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc., which is a computing core and a control core of the terminal, and is adapted to implement one or more instructions, and in particular, to load and execute one or more instructions to implement a corresponding method flow or a corresponding function; the processor provided by the embodiment of the invention can be used for the operation of the post-disaster circuit-circuit cooperative repair method.
A data-driven building heating load prediction system, comprising:
the data sample acquisition module is used for constructing a feature set according to a building structure, user behaviors and environmental weather and forming data samples based on the feature sets with different time scales;
the building heat supply prediction module is used for respectively extracting deep data characteristics of a plurality of different time scales in the data sample based on the bidirectional long-short term memory network, inputting the extracted deep data characteristics and the deep data characteristics of the current moment into the multilayer perceptron together, and establishing a building heat supply prediction model based on the input sample and the building heat supply power;
the network optimization module is used for updating parameters of the building heat supply prediction model by adopting historical sample data in data samples with different time scales, and then performing online optimization updating on the building heat supply prediction model by adopting a dynamically generated building heat supply actual power data stream and using an L1-RDA online learning method;
and the prediction module is used for training the optimized and updated building heat supply prediction model according to the characteristic set of the corresponding building, and realizing building heat supply load prediction by using the trained building heat supply prediction model and based on the dynamic data flow of the building heat supply actual power.
In still another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory), which is a Memory device in the terminal device and is used for storing programs and data. The computer-readable storage medium includes a built-in storage medium in the terminal device, provides a storage space, stores an operating system of the terminal, and may also include an extended storage medium supported by the terminal device. Also, one or more instructions, which may be one or more computer programs (including program code), are stored in the memory space and are adapted to be loaded and executed by the processor. It should be noted that the computer-readable storage medium may be a high-speed RAM memory, or may be a Non-volatile memory (Non-volatile memory), such as at least one disk memory. One or more instructions stored in a computer-readable storage medium may be loaded and executed by a processor to implement the corresponding steps of the above-described embodiments that may be used in a data-driven building heating load prediction method.
Based on the statistical law of the heat balance equation and the historical data, the building equivalent heat loss coefficient and the user-environment combined equivalent heat loss coefficient are used for respectively describing the physical structure of the building and the contribution of the user behavior and the environmental change to the heat supply requirement, so that the complex physical modeling of the building structure and the probability description of the uncertainty of the user behavior and the environmental change are avoided, and the method has strong popularization and applicability.
The invention uses the Bi-LSTM network to respectively analyze the sequence data at the same time within 6 hours of history and one week of history, extracts the change trend characteristics of the data on different time scales and fully utilizes the autocorrelation of the history data on time. Meanwhile, the extracted deep features and the data information at the current moment are input to the multi-layer perceptron together, so that the used data information is enriched, and the accuracy of building heat supply power prediction is further improved.
The invention uses a neural network training framework of off-line training-on-line optimization. In the on-line application stage, a building heat supply prediction model trained in the off-line stage is used as a basis, and the dynamic data stream of the actual heat supply power is used for updating the building heat supply prediction model in real time, so that the prediction model can better cope with the uncertainty of environmental weather and user behaviors, and the robustness and the generalization capability of the model are enhanced.
The embodiments of the present invention have been described in detail, but the embodiments are only a part of the embodiments of the present invention, and the scope of the present invention is not limited thereto. Non-inventive modifications, equivalents and variations of the embodiments of the invention herein disclosed may occur to persons skilled in the art and are to be included within the scope of the invention as defined by the appended claims.

Claims (10)

1. A building heating load prediction method based on data driving is characterized by comprising the following steps:
s1, constructing a feature set according to the building structure, the user behavior and the environmental weather, and forming a data sample based on the feature set with different time scales;
s2, respectively extracting deep data characteristics of a plurality of different time scales in the data sample by using a bidirectional long-short term memory network, inputting the extracted deep data characteristics and the deep data characteristics of the current moment into a multi-layer perceptron together, and establishing a building heat supply prediction model based on the input sample and the building heat supply power;
s3, updating parameters of the building heat supply prediction model by adopting historical sample data in data samples with different time scales, and then performing online optimization updating on the building heat supply prediction model by adopting a dynamically generated building heat supply actual power data stream and using an L1-RDA online learning method;
and S4, acquiring a feature set of the corresponding building to train the optimized and updated building heat supply prediction model, and realizing building heat supply load prediction by using the trained building heat supply prediction model based on the dynamic data flow of the building heat supply actual power.
2. The method as claimed in claim 1, wherein the building structure-based characteristics include building equivalent heat loss coefficient and building transparent enclosure area ratio, the user behavior-based characteristics include time attribute of the current time, the environmental weather-based characteristics include external ambient temperature, average illumination intensity, relative humidity and average wind speed, and the user-environment combined equivalent heat loss coefficient based on user behavior and environmental weather.
3. The method as claimed in claim 2, wherein the building heating power P at the historical time i is used based on the heat balance equationiAnd corresponding building interior air temperature
Figure FDA0003433399800000011
And the temperature of the external environment
Figure FDA0003433399800000012
Obtaining the heat loss coefficient at the current moment, and averagely obtaining the final building equivalent heat loss coefficient kbuildThe specific calculation formula is as follows:
Figure FDA0003433399800000013
in the formula, n is the total amount of the sampled historical time.
4. The method for predicting building heating load based on data driving as claimed in claim 1, wherein the Bi-LSTM is adopted to extract corresponding deep features thereof, and the output of the Bi-LSTM neural network at time t is:
Figure FDA0003433399800000014
in the formula, WforAnd WbackMapping matrices for the outputs of the forward and backward layers respectively,
Figure FDA0003433399800000015
and
Figure FDA0003433399800000016
the short-term memory of the layer output is hidden at the time t for the forward layer and the backward layer respectively.
5. The method as claimed in claim 1, wherein the method for predicting building heating load based on data driving is characterized in that MLP is used to establish a mapping relation between MLP and building heating power value by taking the characteristics corresponding to the current moment and the sequence characteristics extracted from a plurality of different time scale data as input, and the forward propagation formula of the l-th sensing layer in MLP is as follows:
hl＝ELU(Wlhl-1+bl)
in the formula, hlFor the output of the first sensing layer, ELU function is the nonlinear function used by the sensing layer, WlA connection weight matrix for the l-th layer sensing layer, hl-1As input to the l-th sensing layer, blThe input bias of the sensing layer of the l-th layer.
6. The building heating load prediction method based on data driving as claimed in claim 1, wherein the building heating prediction model is optimized by taking the mean square error as the loss function, and the specific calculation formula of the mean square error loss function L is as follows:
Figure FDA0003433399800000021
in the formula, X is an input matrix formed by N training samples, N is the total number of input training samples, theta is a parameter of the neural network, and yiFor the actual heating power corresponding to the ith training sample,
Figure FDA0003433399800000022
predicted heating power, x, for the ith training sampleiIs the ith training sample.
7. The method as claimed in claim 1, wherein the output power of the building heating system is collected at intervals and constructed into corresponding training samples, and the deep features of the data stream are mined based on the online learning method with the aim of minimizing the accumulated online loss, so as to optimize the parameters of the heating prediction network in real time.
8. The method for predicting building heating load based on data driving as claimed in claim 1, wherein the L1-RDA online learning algorithm is a method for solving the optimal parameters of the neural network online, and the optimization target of the L1-RDA algorithm at the t time step is as follows:
Figure FDA0003433399800000023
in the formula, thetatNeural network parameters updated for the t-th time step, GrFor the gradient at the r-th time step,<Gr,θ>is a gradient GrThe median of integral over theta, gamma and lambda are parameters of the L1-RDA algorithm。
9. A building heating load prediction system based on data driving is characterized by comprising:
the data sample acquisition module is used for constructing a feature set according to a building structure, user behaviors and environmental weather and forming data samples based on the feature sets with different time scales;
the building heat supply prediction module is used for respectively extracting deep data characteristics of a plurality of different time scales in the data sample based on the bidirectional long-short term memory network, inputting the extracted deep data characteristics and the deep data characteristics of the current moment into the multilayer perceptron together, and establishing a building heat supply prediction model based on the input sample and the building heat supply power;
the network optimization module is used for updating parameters of the building heat supply prediction model by adopting historical sample data in data samples with different time scales, and then performing online optimization updating on the building heat supply prediction model by adopting a dynamically generated building heat supply actual power data stream and using an L1-RDA online learning method;
and the prediction module is used for training the optimized and updated building heat supply prediction model according to the characteristic set of the corresponding building, and realizing building heat supply load prediction by using the trained building heat supply prediction model and based on the dynamic data flow of the building heat supply actual power.
10. A terminal device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the steps of the method of any of claims 1 to 8 are implemented when the computer program is executed by the processor.
CN202111604996.3A 2021-12-24 2021-12-24 Building heat supply load prediction method, device and equipment based on data driving Pending CN114239991A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111604996.3A CN114239991A (en) 2021-12-24 2021-12-24 Building heat supply load prediction method, device and equipment based on data driving

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111604996.3A CN114239991A (en) 2021-12-24 2021-12-24 Building heat supply load prediction method, device and equipment based on data driving

Publications (1)

Publication Number Publication Date
CN114239991A true CN114239991A (en) 2022-03-25

Family

ID=80762944

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111604996.3A Pending CN114239991A (en) 2021-12-24 2021-12-24 Building heat supply load prediction method, device and equipment based on data driving

Country Status (1)

Country Link
CN (1) CN114239991A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115754875A (en) * 2022-11-03 2023-03-07 国网安徽省电力有限公司营销服务中心 Voltage transformer error evaluation method and device
CN116579506A (en) * 2023-07-13 2023-08-11 陕西通信规划设计研究院有限公司 Building energy consumption data intelligent management method and system based on big data
CN117132006A (en) * 2023-10-27 2023-11-28 中国铁塔股份有限公司吉林省分公司 Energy consumption prediction method and system based on energy management system

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115754875A (en) * 2022-11-03 2023-03-07 国网安徽省电力有限公司营销服务中心 Voltage transformer error evaluation method and device
CN115754875B (en) * 2022-11-03 2023-08-11 国网安徽省电力有限公司营销服务中心 Voltage transformer error assessment method and device
CN116579506A (en) * 2023-07-13 2023-08-11 陕西通信规划设计研究院有限公司 Building energy consumption data intelligent management method and system based on big data
CN116579506B (en) * 2023-07-13 2023-09-19 陕西通信规划设计研究院有限公司 Building energy consumption data intelligent management method and system based on big data
CN117132006A (en) * 2023-10-27 2023-11-28 中国铁塔股份有限公司吉林省分公司 Energy consumption prediction method and system based on energy management system
CN117132006B (en) * 2023-10-27 2024-01-30 中国铁塔股份有限公司吉林省分公司 Energy consumption prediction method and system based on energy management system

Similar Documents

Publication Publication Date Title
CN113962364B (en) Multi-factor power load prediction method based on deep learning
CN110705743B (en) New energy consumption electric quantity prediction method based on long-term and short-term memory neural network
CN114239991A (en) Building heat supply load prediction method, device and equipment based on data driving
CN113128793A (en) Photovoltaic power combination prediction method and system based on multi-source data fusion
CN103117546B (en) A kind of Ultrashort-term slide prediction method for wind power
CN111260136A (en) Building short-term load prediction method based on ARIMA-LSTM combined model
CN113112077B (en) HVAC control system based on multi-step prediction deep reinforcement learning algorithm
CN112434787B (en) Terminal space energy consumption prediction method, medium and equipment based on total energy consumption of building
CN113554466A (en) Short-term power consumption prediction model construction method, prediction method and device
CN115374995A (en) Distributed photovoltaic and small wind power station power prediction method
CN115471362A (en) Comprehensive energy source-load prediction method for depth feature-guided two-stage transfer learning
CN112418495A (en) Building energy consumption prediction method based on longicorn stigma optimization algorithm and neural network
CN113591368A (en) Comprehensive energy system multi-energy load prediction method and system
CN116796141A (en) GBDT regression model-based office building energy consumption prediction method
Dong et al. Short-term building cooling load prediction model based on DwdAdam-ILSTM algorithm: A case study of a commercial building
CN116345555A (en) CNN-ISCA-LSTM model-based short-term photovoltaic power generation power prediction method
CN115238854A (en) Short-term load prediction method based on TCN-LSTM-AM
Kumar et al. Forecasting indoor temperature for smart buildings with ARIMA, SARIMAX, and LSTM: A fusion approach
CN114611757A (en) Electric power system short-term load prediction method based on genetic algorithm and improved depth residual error network
CN115310727B (en) Building cooling, heating and power load prediction method and system based on transfer learning
CN115936236A (en) Method, system, equipment and medium for predicting energy consumption of cigarette factory
CN115310355A (en) Multi-energy coupling-considered multi-load prediction method and system for comprehensive energy system
CN114861555A (en) Regional comprehensive energy system short-term load prediction method based on Copula theory
CN114862023A (en) Distributed photovoltaic power prediction method and system based on four-dimensional point-by-point meteorological forecast
Roh et al. TFE-NET: Time and Feature focus Embedding Network for Multivariate-to-Multivariate Time Series Forecasting

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