A data-driven building heating load prediction method, device and equipment

Abstract

The present invention discloses a data-driven building heating load prediction method, device and equipment, which constructs a feature set according to building structure, user behavior and environmental weather, forms data samples based on the feature sets of different time scales, extracts deep data features of multiple different time scales in the data samples, establishes a building heating prediction model based on the input samples and the building heating power, adopts an offline training and online optimization neural network training framework, uses the L1-RDA online learning method to perform online optimization and update on the building heating prediction model, and uses the dynamic data stream of the actual heating power to perform real-time update on it, so that the prediction model can better cope with the uncertainty of environmental weather and user behavior, enhances the robustness and generalization ability of the model, improves the accuracy of short-term prediction of building heating, and has important significance for deeply exploring the energy-saving potential of buildings and constructing new near-zero energy consumption buildings.

Description

Technical Field

[0001] The present invention belongs to the field of building thermal energy regulation, and particularly relates to a data-driven building heating load prediction method, device and equipment. Background Art

[0002] Under the dual-carbon goal, improving energy utilization efficiency and realizing energy transformation have become the consensus in the development of various fields in China. According to the "Research Report on Building Energy Consumption in China (2020)" released by the China Building Energy Conservation Association in 2020, the operating energy consumption and the total energy consumption of the whole process of buildings in China account for 21.7% and 46.5% of the total national energy consumption respectively. With the in-depth development of industrialization and urbanization in China, the total building energy consumption will further increase. At the same time, the wide application of distributed photovoltaics in buildings has changed the energy consumption structure of buildings, making full use of the ability of users to consume new energy nearby, and the heat storage property of buildings themselves also provides a buffer for the formulation of building operation scheduling strategies. Therefore, it is necessary to deeply explore the energy-saving potential of buildings, optimize the building operation scheduling strategy, and build a new type of near-zero energy consumption building.

[0003] The accuracy and reliability of short-term building heating prediction are the basis for the formulation and optimization of building operation scheduling strategies, and are of great significance for improving building energy efficiency. The research methods for building heating prediction can be divided into two categories: physical models and data-driven models. In physical models, the heat balance equation is the core, and thermodynamic dynamic models of buildings are described using heat resistance-capacitance networks, conduction-transfer functions, etc., and combined with prediction models of ambient temperature, probability models of user behavior, etc. to predict the building heating power; data-driven models design features from multiple aspects such as environmental weather, user behavior, historical heating power, etc., and use machine learning models to establish the mapping relationship between the constructed features and the building heating power.

[0004] However, in the building heating prediction method based on physical models, the modeling of building structures is relatively complex, there are large differences in models for different buildings, the generalization and universality of the models are poor, and the description of user behavior is relatively simple. Therefore, the prediction accuracy is relatively low. With the rapid development of the machine learning field and the wide access of a large number of smart meters on the building side, data-driven models have achieved good results in short-term building heating prediction. However, the current data-driven models lack feature extraction of building structures, and only use time attribute features to describe complex user behavior, without making full use of data information at different time scales. At the same time, the current data-driven models are directly used for short-term heating power prediction of buildings after offline training, without using the dynamic data stream of building heating power to update the data-driven models for online applications in real time, and it is difficult to adapt to the randomness and variability of user behavior. Summary of the Invention

[0005] The purpose of the present invention is to provide a data-driven building heating load prediction method, device and equipment to overcome the deficiencies of the prior art. The present invention can improve the accuracy of short-term prediction of building heating, which is of great significance for deeply exploring the energy-saving potential of buildings and constructing new near-zero energy consumption buildings.

[0006] A data-driven building heating load prediction method includes the following steps:

[0007] S1, construct a feature set based on building structure, user behavior, and environmental weather, and form data samples based on the feature sets at different time scales;

[0008] S2, use a bidirectional long short-term memory network to extract deep data features at multiple different time scales in the data samples respectively, and input the extracted deep data features and the deep data features at the current moment into a multi-layer perceptron together to establish a building heating prediction model based on the input samples and the building heating power;

[0009] S3, update the parameters of the building heating prediction model using the historical sample data in the data samples at different time scales, and then use the dynamically generated actual building heating power data stream and the L1-RDA online learning method to perform online optimization and update on the building heating prediction model;

[0010] S4, collect the feature set corresponding to the building to train the optimized and updated building heating prediction model, and use the trained building heating prediction model to realize building heating load prediction based on the dynamic data stream of the actual building heating power.

[0011] Further, the features based on the building structure include the equivalent heat loss coefficient of the building and the proportion of the area of the building's transparent envelope structure, the features based on user behavior include the time attribute at the current moment, the features based on environmental weather include the external environmental temperature, average light intensity, relative humidity, and average wind speed, and the user-environment joint equivalent heat loss coefficient based on user behavior and environmental weather.

[0012] Further, based on the heat balance equation, using the building heating power P i at the historical moment i, and the corresponding indoor air temperature and the external environmental temperature to obtain the heat loss coefficient at the current moment, and averaging to obtain the final equivalent heat loss coefficient k build of the building, and the specific calculation formula is as follows:

[0013]

[0014] In the formula, n is the total number of sampled historical moments.

[0015] Furthermore, Bi-LSTM is used to extract its corresponding deep features respectively. The output of the Bi-LSTM neural network at time t is:

[0016]

[0017] In the formula, W for and W back are the output mapping matrices of the forward layer and the backward layer respectively, and are the short-term memories of the hidden layer outputs of the forward layer and the backward layer at time t respectively.

[0018] Furthermore, taking the features corresponding to the current moment and the sequence features extracted from multiple data at different time scales as inputs, an MLP is used to establish the mapping relationship with the heating power value of the building. The forward propagation formula of the l-th perceptron layer in the MLP is:

[0019] h l = ELU(W l h l-1 + b l )

[0020] In the formula, h l is the output of the l-th perceptron layer, the ELU function is the non-linear function used by the perceptron layer, W l is the connection weight matrix of the l-th perceptron layer, h l-1 is the input of the l-th perceptron layer, and b l is the input bias of the l-th perceptron layer.

[0021] Furthermore, taking the mean squared error as the loss function to optimize the building heating prediction model, the specific calculation formula of the mean squared error loss function L is:

[0022]

[0023] In the formula, X is the input matrix composed of N training samples, N is the total number of input training samples, θ is the parameter of the neural network, y i is the actual heating power corresponding to the i-th training sample, is the predicted heating power corresponding to the i-th training sample, and x i is the i-th training sample.

[0024] Furthermore, the output power of the building heating system is collected at intervals and constructed into corresponding training samples. Based on the online learning method, with the goal of minimizing the cumulative online loss, the deep features of the data stream are mined, and the parameters of the heating prediction network are optimized in real time.

[0025] Furthermore, the L1-RDA online learning algorithm is a method for online solving the optimal parameters of a neural network. The optimization objective of the L1-RDA algorithm at the t-th time step is:

[0026]

[0027] where θ t is the neural network parameter updated at the t-th time step, G r is the gradient at the r-th time step, <G r , θ> is the integral median value of the gradient G r with respect to θ, and γ and λ are the parameters of the L1-RDA algorithm.

[0028] A data-driven building heating load prediction system includes:

[0029] A data sample collection module, configured to construct a feature set according to building structure, user behavior, and environmental weather, and form data samples based on the feature sets of different time scales;

[0030] A building heating prediction module, which respectively extracts deep data features of multiple different time scales in the data samples based on a bidirectional long short-term memory network, and inputs the extracted deep data features and the deep data features at the current moment into a multi-layer perceptron together to establish a building heating prediction model based on the input samples and the building heating power;

[0031] A network optimization module, configured to update the parameters of the building heating prediction model using the historical sample data in the data samples of different time scales, and then use the dynamically generated actual building heating power data stream to perform online optimization and update on the building heating prediction model using the L1-RDA online learning method;

[0032] A prediction module, which trains the optimized and updated building heating prediction model according to the feature set of the corresponding building, and uses the trained building heating prediction model to realize building heating load prediction based on the dynamic data stream of the actual building heating power.

[0033] A terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the data-driven building heating load prediction method.

[0034] Compared with the prior art, the present invention has the following beneficial technical effects:

[0035] A building heating load prediction method based on data-driven constructs a feature set according to building structure, user behavior, and environmental weather, and forms data samples based on the feature sets of different time scales; uses a bidirectional long short-term memory network to extract deep data features of multiple different time scales in the data samples respectively, and inputs the extracted deep data features and the deep data features at the current moment into a multi-layer perceptron together to establish a building heating prediction model based on the input samples and the heating power of the building. An offline training and online optimization neural network training framework is adopted, and the L1-RDA online learning method is used to optimize and update the building heating prediction model online, and its dynamic data stream of actual heating power is used to update it in real time, so that the prediction model can better cope with the uncertainties of environmental weather and user behavior, enhance the robustness and generalization ability of the model, improve the accuracy of short-term prediction of building heating, and is of great significance for deeply exploring the energy-saving potential of buildings and constructing new near-zero energy consumption buildings.

[0036] Furthermore, based on the heat balance equation and the statistical law of historical data, the equivalent heat loss coefficient of the building and the user-environment joint equivalent heat loss coefficient are used to describe the physical structure of the building and the contributions of user behavior and environmental changes to heating demand respectively, avoiding complex physical modeling of the building structure and probabilistic description of the uncertainties of user behavior and environmental changes, and having strong popularization and applicability.

[0037] Furthermore, the Bi-LSTM network is used to analyze the sequence data at the same time of the historical 6 hours and the historical one week respectively, and extract the change trend features of the data on different time scales, making full use of the autocorrelation of historical data in time. At the same time, the extracted deep features and the data information at the current moment are input into the multi-layer perceptron together, enriching the data information used and further improving the accuracy of the prediction of the heating power of the building. Description of the Drawings

[0038] Figure 1 It is a schematic diagram of the feature set in the embodiment of the present invention.

[0039] Figure 2 It is a schematic diagram of the structure of the Bi-LSTM network in the embodiment of the present invention.

[0040] Figure 3 It is a unit structure diagram of the forward layer of the Bi-LSTM in the embodiment of the present invention.

[0041] Figure 4 It is a structure diagram of the building heating prediction model in the embodiment of the present invention.

[0042] Figure 5 It is a schematic diagram of the offline training and online optimization of the building heating prediction model in the embodiment of the present invention. Specific Embodiments

[0043] The embodiments given below are intended to further illustrate the present invention, but should not be construed as limiting the protection scope of the present invention. Some non-essential improvements and adjustments made by those skilled in the art to the present invention based on the content of the present invention still fall within the protection scope of the present invention.

[0044] Unless otherwise clearly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "having" etc. will be understood to include the stated elements or components, without excluding other elements or other components.

[0045] To achieve the above object, the present invention provides a data-driven building heating load prediction method, including the following steps:

[0046] Step 1: Based on the heat balance equation and the statistical law of historical data, construct the equivalent heat loss coefficient of the building and the combined equivalent heat loss coefficient of the user-environment, and construct a feature set according to the time attribute, temperature, relative humidity, average wind speed, and average light intensity of the external environment at the same moment of the building. Describe the influencing factors related to the heating power of the building from three perspectives: building structure, user behavior, and environmental weather. Considering the autocorrelation of historical data in time, finally, data samples at three different time scales are formed: the current moment, the historical 6 hours, and the same moment within the historical week.

[0047] Step 2: Use a bidirectional long short term memory network (Bi-LSTM) to extract the deep data features at two time scales of the historical 6 hours and the same moment within the historical week respectively, and input the extracted deep data features and the deep data features of the current moment into a multi-layer perceptron (MLP) together to establish a mapping relationship between the input sample and the heating power of the building, and then a building heating prediction model based on the input sample and the heating power of the building can be obtained.

[0048] Step 3: In the offline training stage, based on the historical data set, use the batch training Adam algorithm to update the parameters of the building heating prediction model; in the online application stage, for the dynamically generated actual power data stream of building heating, use the L1-RDA online learning method to perform online learning and update on the building heating prediction model.

[0049] Step 4: For a specific building, collect relevant data and construct a training set, and use this training set to offline train the building heating prediction model.

[0050] After the offline training is completed, it enters the online application stage. Based on the dynamic data stream of the actual power of building heating, the building heating prediction model is updated in real time using the online learning method.

[0051] A data-driven building heating load prediction method specifically includes the construction of a multi-time scale data set, the establishment of a building heating prediction model, the offline training and online optimization of the building heating prediction model, and the testing and application of the building heating prediction model. The specific details of each part are as follows. It should be noted that the building objects analyzed in the present invention are office buildings or teaching buildings, rather than residential buildings, and the time interval for building heating prediction in the present invention is 15 minutes.

[0052] Construction of a multi-time scale data set

[0053] The present invention constructs a feature set from three perspectives: building structure, user behavior, and environmental weather, specifically as Figure 1 shown. Among them, the features selected in terms of building structure include the equivalent heat loss coefficient of the building and the proportion of the area of the building's transparent envelope structure; the feature selected in terms of user behavior is the time attribute at the current moment (instant, week, month, and quarter); the features selected in terms of environmental weather include the external environmental temperature, average light intensity, relative humidity, and average wind speed. The user-environment joint equivalent heat loss coefficient is added as a feature in terms of users and the environment. The equivalent heat loss coefficient of the building and the user-environment joint equivalent heat loss coefficient are described separately below.

[0054] Equivalent heat loss coefficient of the building

[0055] The envelope structure of the building itself hinders the heat convection process between the air inside the building and the air in the external environment to a certain extent, making the building itself have a certain degree of heat storage. Considering the complexity of the building structure and the differences in different building structures, the present invention uses the equivalent heat loss coefficient of the building to measure the heat storage capacity of the building itself.

[0056] Based on the heat balance equation, the present invention uses the building heating power P i at the historical moment i, and the corresponding indoor air temperature and the external environmental temperature to obtain the heat loss coefficient at the current moment, and the final equivalent heat loss coefficient k build of the building is obtained by averaging. The specific calculation formula is as follows:

[0057]

[0058] In the formula, n is the total number of historical moments sampled.

[0059] To avoid the errors and interferences caused by user behavior and external environment in the solution of the equivalent heat loss coefficient of building, the present invention only selects the historical data of the building from 0:00 to 5:00 in the early morning, and monitors the building lighting load and the external environment wind speed at the same time. If it is greater than the threshold value, it indicates that the user behavior or the environmental wind speed at this time may affect the building heat balance. Therefore, the data at this moment is discarded.

[0060] It should be noted that the equivalent heat loss coefficient of building is only related to the physical structure of the building itself, and it is considered a constant when the building is not transformed. At the same time, in order to improve the credibility of the calculation result of formula (1), the present invention uses the historical data of a winter quarter to calculate the equivalent heat loss coefficient of building.

[0061] User-environment combined equivalent heat loss coefficient

[0062] User behavior and environmental weather have strong randomness and variability, and have a great uncertainty impact on the heating power of buildings. Therefore, the present invention constructs a user-environment combined equivalent heat loss coefficient to comprehensively measure the impacts of both, and the corresponding calculation formula is as follows.

[0063]

[0064] It should be noted that the user-environment combined equivalent heat loss coefficient is a dynamic variable, which comprehensively describes the user's physiological heat dissipation, the heat loads such as the user's control of lighting equipment, the user's control of transparent envelope structures such as windows, and the impacts of external environmental weather on the building heat balance.

[0065] After the feature set is constructed, the present invention selects the proportion of the building transparent envelope structure area, the corresponding hour, week, month, quarter, external environmental temperature, average light intensity, relative humidity, average wind speed and the equivalent heat loss coefficient of building at the future moment as the inputs of the building heating prediction model. At the same time, based on the building heating power and the user-environment combined equivalent heat loss coefficient corresponding to the same moment within the historical 6 hours and the historical one week, two sequence data with different time scales are constructed and used as inputs. Therefore, the building heating prediction model proposed by the present invention includes the input data at the current moment, the historical 6 hours and the same moment within the historical one week - three time scales.

[0066] Establishment of the building heating prediction model

[0067] For the sequence data at the historical 6 hours and the same moment within the historical one week, the present invention uses Bi-LSTM to extract their corresponding deep features respectively. The structural schematic diagram of Bi-LSTM is as Figure 2 shown, in the figure, Denote the LSTM cell at time step \(t\) in the forward layer of the Bi-LSTM, Denote the LSTM cell at time step \(t\) in the backward layer of the Bi-LSTM, \(x\) t Is the input of the Bi-LSTM neural network at time step \(t\), \(y\) t Is the output of the Bi-LSTM neural network at time step \(t\).

[0068] In the Bi-LSTM neural network, the data information in the forward layer propagates forward along the time steps, and the data information in the backward layer propagates backward along the time steps. Moreover, the output at each time step depends on the output of the forward layer and the output of the backward layer at that time step. Therefore, the Bi-LSTM neural network can analyze the change trend of sequence data bidirectionally and better extract the dependence and correlation between the data information of the previous and subsequent time steps. The output of the Bi-LSTM neural network at time step \(t\) is:

[0069]

[0070] Where \(W\) for And \(W\) back Are the output mapping matrices of the forward layer and the backward layer respectively, And Are the short-term memories of the hidden layer outputs of the forward layer and the backward layer at time step \(t\) respectively.

[0071] The unit structure of the forward layer of the Bi-LSTM neural network is as Figure 3 Shown. The unit structure of the backward layer is similar, only the direction of information propagation is opposite. The forward layer of the Bi-LSTM needs to obtain the outputs \(f\) t Of the forget gate, input gate, and output gate at time step \(t\) according to the input \(x\) At the current time step \(t\), and the short-term memory t for Of the hidden layer output at the adjacent time steps And The specific calculation formulas are:

[0072]

[0073]

[0074]

[0075] Where And Are the connection weights between the input at the current time step and the three gate structures respectively, \(W\) i for And They are the connection weights of the short-term memory output by the hidden layer at adjacent time steps with the three gate structures respectively, and the biases of the three gate structures respectively.

[0076] In the forward layer unit structure of the Bi-LSTM neural network, the output of the forget gate controls the forgetting degree of the long-term memory at adjacent time steps and the output of the input gate controls the retention degree of the candidate long-term memory while the output of the output gate controls the part of the long-term memory at the current time step that is retained to the short-term memory Therefore, the specific calculation formulas for updating the long-term memory and the short-term memory are as follows:

[0077]

[0078]

[0079]

[0080] In the formula, is the connection weight between the input at the current time step and the candidate long-term memory, is the connection weight between the short-term memory at adjacent time steps and the candidate long-term memory, and

[0081] is the corresponding input bias.

[0082] h l = ELU(W l h l-1 + b l ) (14)

[0083] In the formula, h l is the output of the l-th perceptron layer, the ELU function is the non-linear function used by the perceptron layer, W l is the connection weight matrix of the l-th perceptron layer, h l-1 is the input of the l-th perceptron layer, and b l is the input bias of the l-th perceptron layer. It should be noted that the last perceptron layer does not use a non-linear function.

[0084] Therefore, the structure of the building heating prediction model proposed by the present invention is shown in Figure 4. Considering the high accuracy and good performance of the MLP model in existing research, the present invention adds two features, namely the equivalent heat loss coefficient of the building and the combined equivalent heat loss coefficient of the user and the environment, while retaining the basic structure of the MLP model, and uses a Bi-LSTM network to extract deep features of historical data at different time scales as the model input.

[0085] Offline Training and Online Optimization of Building Heating Prediction Model

[0086] For the parameter optimization process of the building heating prediction model, the present invention uses a neural network learning route of "offline training - online optimization". In the offline training stage, the present invention uses batch training to optimize the parameters of the neural network based on fixed sample data; in the online optimization stage, the present invention uses the L1-RDA online learning algorithm to optimize the parameters of the neural network based on dynamic sample data streams, improving the accuracy and sparsity of the model. The overall schematic is as Figure 5 shown.

[0087] Offline Training of Building Heating Prediction Model

[0088] In the offline training stage, the present invention collects historical data related to the building and constructs a dataset, which is divided into an offline training set and an offline test set in an 8:2 ratio. Based on the offline training set, the present invention optimizes the building heating prediction model with the mean square error as the loss function. The specific calculation formula of the mean square error loss function L is:

[0089]

[0090] In the formula, X is the input matrix composed of N training samples, N is the total number of input training samples, θ is the parameter of the neural network, y i is the actual heating power corresponding to the i-th training sample, is the predicted heating power corresponding to the i-th training sample, x i is the i-th training sample.

[0091] Considering that the samples in the dataset are static during offline training, the present invention updates the parameters of the heating prediction neural network using the Adam algorithm based on batch training, and the specific process is as follows.

[0092] (1) Set the number of samples for each batch training as N, set the hyperparameters α, β1, and β2 of the Adam algorithm, and initialize the parameters m0, v0, and t0 to 0, and ε to 10 -8 ;

[0093] (2) For the training samples of the j-th batch, calculate the corresponding loss function value using Equation (11), and use the error backpropagation method to obtain the gradient of the loss function with respect to the neural network parameter θ

[0094] (3) Update the biased first - moment estimate \(m\) in the Adam algorithm j-1 and the biased second - moment estimate \(v\) j-1 The calculation formula is as follows:

[0095]

[0096]

[0097] (4) According to the updated \(m\) j and \(v\) j Update the parameters \(\theta\) of the neural network j-1 as follows:

[0098]

[0099] By repeatedly executing steps (2) to (4), the building heating prediction network can be trained and optimized. When the loss function value of the offline test set no longer decreases, the offline training phase of the building heating prediction network ends.

[0100] Online optimization of the building heating prediction model

[0101] In the online application stage, the present invention collects the output power of the building heating system at 15 - minute intervals and constructs corresponding training samples. Since the samples in the online stage are continuously generated at 15 - minute intervals, the batch training algorithm used in the offline training stage is difficult to use a single sample to update the parameters of the heating prediction network. Therefore, based on the online learning method, with the goal of minimizing the cumulative online loss, the present invention mines the deep features of the data stream and optimizes the parameters of the heating prediction network in real - time.

[0102] Considering that the sparsity of the neural network model can not only play a role in feature selection, but also greatly reduce the computational complexity of the heating power prediction process. Therefore, the present invention uses the L1 - RDA online learning algorithm to update the building heating prediction network in real - time.

[0103] The L1 - RDA online learning algorithm is essentially a method for online solving the optimal parameters of the neural network. The optimization objective of the L1 - RDA algorithm at the \(t\) - th time step is:

[0104]

[0105] where \(\theta\) t is the updated neural network parameter at the \(t\) - th time step, \(G\) r is the gradient at the \(r\) - th time step, \(\langle G r ,\theta\rangle\) is the gradient \(G\)r The integral mean value of θ, where γ and λ are parameters of the L1-RDA algorithm.

[0106] Decompose Equation (15) into multiple independent optimization problems according to each feature dimension. The optimization objective for the i-th dimension is:

[0107]

[0108] where is the updated neural network parameter for the i-th dimension at the t-th time step, is the gradient G r component of the i-th feature dimension.

[0109] Since the term |θ i | is non-differentiable at θ i = 0, define ξ as the sub-derivative of |θ i |, so:

[0110]

[0111] Take the sub-derivative of the optimization objective in Equation (16) and set it to 0, then judge whether Equation (16) can be 0 based on Equation (17). Finally, the weight update method of the L1-RDA online learning algorithm in each dimension can be obtained as:

[0112]

[0113] where sgn is the sign function.

[0114] The specific steps of the L1-RDA online learning algorithm are described as follows.

[0115] (1) Set the parameters γ and λ of the L1-RDA algorithm, initialize the cumulative gradient G0 to 0, and use the heating prediction network parameter θ0 at the end of the offline training stage as the initial parameter;

[0116] (2) For the samples generated at the t-th time step, update the cumulative gradient value G t-1 to G t , and the specific update formula is:

[0117]

[0118] (3) Use Equation (18) to update the parameter values of each dimension of the heating prediction network respectively.

[0119] Loop through steps (2) and (3) to update the parameters of the building heating prediction model using the dynamically generated data stream, thereby enhancing the ability of the prediction model to learn the uncertainty of the online learning environment weather and user behavior, and improving the accuracy of the prediction model.

[0120] Testing and Application of Building Heating Prediction Model

[0121] For a specific building, historical data such as its heating power is collected, and a data sample set characterized by multiple time scales is constructed. Two Bi-LSTM networks are used to process the sequence data corresponding to different time scales respectively. Then, the obtained deep features and the features at the current moment are jointly input into the MLP network to predict the building heating power value with a time interval of 15 minutes. In the offline stage, according to the constructed data sample set, the present invention uses the Adam algorithm based on batch training to update the parameters of the heating prediction network. After training, in the online optimization stage, the present invention uses the L1-RDA online learning algorithm to extract the information of the heating power data stream in real time and optimize the heating prediction network. The present invention adopts the mean square error index MSE to evaluate the performance of the building heating prediction model. The specific calculation formula is as follows:

[0122]

[0123] In the formula, is the total number of predicted heating power values in the online stage, y i is the actual value of the heating power for the i-th time, is the predicted value of the heating power for the i-th time.

[0124] In an embodiment of the present invention, a terminal device is provided. The terminal device includes a processor and a memory. The memory is used to store a computer program. The computer program includes program instructions. The processor is used to execute the program instructions stored in the computer storage medium. The processor uses a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used for the operation of the post-disaster electric-circuit collaborative repair method.

[0125] A data-driven building heating load prediction system includes:

[0126] A data sample acquisition module, configured to construct a feature set according to building structure, user behavior, and environmental weather, and form data samples based on the feature sets with different time scales;

[0127] The building heating prediction module extracts deep data features of multiple different time scales in the data samples based on a bidirectional long short-term memory network, and inputs the extracted deep data features and the deep data features at the current moment into a multi-layer perceptron together to establish a building heating prediction model based on the input samples and the building heating power;

[0128] The network optimization module is used to update the parameters of the building heating prediction model with the historical sample data in the data samples of different time scales, and then uses the dynamically generated actual building heating power data stream to perform online optimization and update on the building heating prediction model using the L1-RDA online learning method;

[0129] The prediction module trains the optimized and updated building heating prediction model according to the feature set of the corresponding building, and uses the trained building heating prediction model to realize the building heating load prediction based on the dynamic data stream of the actual building heating power.

[0130] In another embodiment of the present invention, the present invention also provides a storage medium, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is the memory device in the terminal device and is used to store programs and data. The computer-readable storage medium includes the built-in storage medium in the terminal device, which provides storage space and stores the operating system of the terminal. It may also include the expandable storage medium supported by the terminal device. Moreover, one or more instructions suitable for being loaded and executed by the processor are stored in this storage space. These instructions may be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here may be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. One or more instructions stored in the computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the method for data-driven building heating load prediction in the above embodiments.

[0131] Based on the heat balance equation and the statistical law of historical data, the present invention uses the equivalent heat loss coefficient of the building and the combined equivalent heat loss coefficient of the user-environment to respectively describe the physical structure of the building, as well as the contributions of user behavior and environmental changes to the heating demand, avoiding complex physical modeling of the building structure and probabilistic description of the uncertainties of user behavior and environmental changes, and having strong popularization and applicability.

[0132] The present invention uses a Bi-LSTM network to analyze the sequence data at the same time of the historical 6 hours and the historical one week respectively, extracts the change trend characteristics of the data on different time scales, and makes full use of the autocorrelation of the historical data in time. At the same time, the extracted deep features and the data information at the current moment are jointly input into a multi-layer perceptron, enriching the data information used and further improving the accuracy of the heating power prediction of the building.

[0133] The present invention uses a neural network training framework of "offline training - online optimization". In the online application stage, based on the building heating prediction model trained in the offline stage, and using the dynamic data stream of the actual heating power to update it in real time, so that the prediction model can better cope with the uncertainty of the environmental weather and user behavior, and enhance the robustness and generalization ability of the model.

[0134] The above has described the embodiments of the present invention in detail, but the above are only a part of the embodiments of the present invention, and the protection scope of the present invention is not limited thereto. Based on the embodiments of the present invention, the non-creative improvements, equivalent replacements and changes made by those skilled in the art shall all fall within the patent coverage scope of the present invention.

Claims

1. A data-driven building heating load prediction method, characterized in that, It includes the following steps: S1. Construct a feature set according to building structure, user behavior, and environmental weather, and form data samples based on the feature sets at different time scales; S2. Use a bidirectional long short-term memory network to extract deep data features at multiple different time scales in the data samples respectively, and input the extracted deep data features and the deep data features at the current moment into a multi-layer perceptron together to establish a building heating prediction model based on the input samples and the heating power of the building; S3. Update the parameters of the building heating prediction model using the historical sample data in the data samples at different time scales. Specifically, optimize the building heating prediction model with the mean square error as the loss function. The specific calculation formula of the mean square error loss function L is: Where X is the input matrix composed of N training samples, N is the total number of input training samples, θ is the parameter of the neural network, and y i is the actual heating power corresponding to the i-th training sample, is the predicted heating power corresponding to the i-th training sample, and x i is the i-th training sample; Then, use the dynamically generated actual power data stream of building heating and the L1-RDA online learning method to perform online optimization and update on the building heating prediction model; S4. Collect the feature set corresponding to the building to train the optimized and updated building heating prediction model. Use the trained building heating prediction model to realize the building heating load prediction based on the dynamic data stream of the actual power of building heating; The features based on the building structure include the equivalent heat loss coefficient of the building and the proportion of the area of the transparent envelope structure of the building. The features based on user behavior include the time attribute at the current moment. The features based on environmental weather include the external environmental temperature, average light intensity, relative humidity, and average wind speed, and the user-environment joint equivalent heat loss coefficient based on user behavior and environmental weather; User-environment combined equivalent heat loss coefficient Comprehensively measuring the influence of user behavior and environmental weather, the corresponding calculation formula is as follows: Based on the heat balance equation, using the building heating power P at historical moment i i , and the corresponding internal air temperature T of the building i in and the external ambient temperature T i out Obtain the heat loss coefficient at the current moment and average it to get the final equivalent heat loss coefficient k of the building build , and the specific calculation formula is as follows: In the formula, n is the total number of sampled historical moments.

2. The method for predicting building heating load based on data driving according to claim 1, wherein, Use Bi-LSTM to extract its corresponding deep features respectively. The output of the Bi-LSTM neural network at time t is: where, W for and W back are the output mapping matrices of the forward layer and the backward layer respectively, and are the short-term memories of the hidden layer outputs of the forward layer and the backward layer at time t respectively.

3. A data-driven building heating load prediction method according to claim 1, characterized in that Using the features corresponding to the current moment and the sequence features extracted from the data at multiple different time scales as inputs, use MLP to establish the mapping relationship between them and the heating power value of the building. The forward propagation formula of the l-th perceptron layer in MLP is: h l = ELU(W l h l-1 + b l ) where h l is the output of the l-th perception layer, the ELU function is the non-linear function used by the perception layer, W l is the connection weight matrix of the l-th perception layer, h l-1 is the input of the l-th perception layer, and b l is the input bias of the l-th perception layer.

4. A data-driven building heating load prediction method according to claim 1, characterized in that, Intermittently collect the output power of the building heating system and construct the corresponding training samples. Based on the L1-RDA online learning method, with the goal of minimizing the cumulative online loss, mine the deep features of the data stream and optimize the parameters of the heating prediction network in real time.

5. A data-driven building heating load prediction method according to claim 1, characterized in that, The L1-RDA online learning algorithm is a method for online solving the optimal parameters of a neural network. The optimization goal of the L1-RDA algorithm at the t-th time step is: where θ t is the neural network parameter updated at the t-th time step, G r is the gradient at the r-th time step, <G r , θ> is the integral median value of the gradient G r with respect to θ, and γ and λ are the parameters of the L1-RDA algorithm.

6. A data-driven building heating load prediction system based on the method according to claim 1, characterized in that, It includes: A data sample collection module, which is used to construct a feature set according to building structure, user behavior, and environmental weather, and form data samples based on the feature sets at different time scales; A building heating prediction module, which extracts deep data features at multiple different time scales in the data samples respectively based on a bidirectional long short-term memory network, and inputs the extracted deep data features and the deep data features at the current moment into a multi-layer perceptron together to establish a building heating prediction model based on the input samples and the heating power of the building; A network optimization module, which is used to update the parameters of the building heating prediction model by using the historical sample data in the data samples of different time scales, and then uses the dynamically generated actual power data stream of building heating and the L1-RDA online learning method to perform online optimization and update on the building heating prediction model; A prediction module, which trains the optimized and updated building heating prediction model according to the feature set of the corresponding building, and uses the trained building heating prediction model to realize the building heating load prediction based on the dynamic data stream of the actual power of building heating.

7. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

Images