Adaptive building hourly load forecasting method based on transfer learning

Abstract

The application discloses a self-adaptive building day-ahead load prediction method based on transfer learning and relates to the technical field of buildings and environmental protection.The application comprises the following steps: S1, data acquisition and processing, wherein original data sets are divided into small data sets of target buildings and large data sets of basic building groups, and missing values of all original data sets are filled; S2, use mode clustering; S3, source domain data screening, historical daily load curves of use modes of load target buildings are screened, and a data transfer training set and a model transfer training set are respectively constructed; S4, day-ahead load prediction model construction; and S5, self-adaptive model optimization, wherein model parameters are continuously adjusted by using Bayesian optimization, and self-adaptive load prediction of target buildings is realized.The application realizes load prediction of target buildings by combining historical data of data-sufficient building groups with the data transfer and model transfer methods of transfer learning.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of building technology, and is used for intelligent buildings, building energy management and related systems, and relates to a self-adaptive building day-ahead load forecasting method based on transfer learning. BACKGROUND

[0002] Building energy consumption accounts for an increasing proportion of total social energy consumption. With the development of technology, buildings have gradually changed from simple energy-consuming terminals to small and complex energy conversion devices. Intelligent systems such as building energy management, intelligent buildings and user behavior guidance can effectively improve building energy efficiency by reasonably allocating energy consumption and energy storage devices. The prerequisite for realizing building energy optimization management is accurate load forecasting. Deep learning algorithms, as a new model in recent years, have better nonlinear description capabilities and can effectively improve load forecasting accuracy, and have been widely studied and applied.

[0003] It is found through retrieval that load forecasting based on deep learning algorithms has become mature. Chinese invention patent CN111563610A “Building electrical load comprehensive forecasting method and system based on LSTM neural network” uses load data, weather parameters and building data of a typical building as input to construct a prediction model using LSTM. Chinese invention patent CN113836823A “Load combination forecasting method based on load decomposition and optimized bidirectional long short-term memory network” uses a variational mode decomposition method to decompose historical load according to trends, and constructs a bidirectional long short-term memory neural network model for each subsequence to improve prediction accuracy. However, deep learning algorithms often require a large amount of historical data as support, otherwise they are prone to fail to converge. However, in fact, most buildings cannot provide sufficient historical data to establish a prediction model due to short operation time, incomplete data collection system and other conditions. SUMMARY

[0004] In order to make up for the shortcomings of the prior art, the application provides a load forecasting method for data-deficient buildings, which realizes load forecasting of target buildings by data transfer and model transfer methods of transfer learning combined with historical data of data-sufficient building groups. The application is realized by the following technical solutions: the application provides a self-adaptive building day-ahead load forecasting method based on transfer learning, which includes S1 data acquisition and processing, target building data and basic building group data are obtained through a smart meter, the original data set is divided into a small data set of the target building and a large data set of the basic building group, and 1-hour historical load data is obtained; missing values in the entire original data set are filled, and the average value of the corresponding time points in the previous and subsequent two weeks is used to fill the missing values;

[0005] S2 uses energy mode clustering, and target building data is used to build daily load curves in units of days;The best K value of K-means++ (k-means++) is found through the elbow rule, and the target building daily load curve is clustered to obtain a clustering model;Each cluster in the clustering result represents an energy mode of the target building;

[0006] S3 source domain data screening, the big data set of the basic building group is composed of daily load curves in units of days, which is input as a clustering model, and the historical daily load curve of the energy mode of the target building is screened, and the data migration training set and the model migration training set are constructed respectively;

[0007] S4, the daily load prediction model is constructed, the long short-term memory neural network (LSTM network) is trained by using the data migration training set, and the data migration prediction model is constructed;The convolutional neural network-long short-term memory network (CNN-LSTM network) is trained by using the model migration training set, and the network structure is fine-tuned by using the target building load data, and the model migration prediction model is constructed, and the average of the outputs of the two models (data migration prediction model and model migration prediction model) is used as the target building daily prediction result;

[0008] S5, adaptive model optimization, in the running process of the daily load prediction model, the model parameters are continuously adjusted by using the Bayesian optimization according to the updated historical load, and the adaptive load prediction of the target building is realized.

[0009] The beneficial effects of the present application are:

[0010] The method of the present application aims at the problem that the target building data is insufficient and it is difficult to establish an accurate daily load prediction model, and the training set is expanded by transfer learning to realize the convergence of the deep learning model, and the advantages of the deep learning network are played while the model convergence is guaranteed, and accurate prediction is realized. BRIEF DESCRIPTION OF DRAWINGS

[0011] The present application will be further described below in combination with the drawings.

[0012] Figure 1 The building daily load prediction model construction flowchart of the present application. DETAILED DESCRIPTION

[0013] In order to make the above objectives, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application are described in detail below with reference to the accompanying drawings. In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the spirit of the present application, so the present application is not limited to the specific embodiments disclosed below.

[0014] The accompanying drawings are specific embodiments of the adaptive building day-ahead load forecasting method based on transfer learning of the present application. The embodiments include,

[0015] S1 data acquisition and processing, obtaining target building data and basic building group data through smart meters, decomposing the original data set into a small data set of the target building and a large data set of the basic building group, and obtaining historical load data with a step of 1 hour; filling in the missing values of the entire original data set, and filling in the missing values with the average value of the corresponding time points of the previous and next two weeks.

[0016] S2 energy use mode clustering, the target building data is divided into days to construct a daily load curve; the elbow rule is used to find the best K value of K-means++, and the target building daily load curve is clustered to obtain a clustering model; each cluster in the clustering result represents a kind of energy use mode of the target building.

[0017] The sum of squared distances of the sample points in each cluster to the cluster center is taken as the performance measure, and the smaller the performance measure is, the more convergent each cluster is. At the same time, extreme cases are avoided, in which all sample points are considered as a cluster (SSE is 0), and the elbow rule is used to seek the balance point between the number of clusters (i.e. K value) and SSE.

[0018] The elbow rule includes the following specific implementation steps: specifying an i value, which is the maximum possible number of clusters; then increasing the number of clusters from 1 to i, calculating i SSEs; according to the potential patterns of the data, when the number of set clusters continuously approaches the true number of clusters, SSE shows a rapid downward trend, and when the number of set clusters exceeds the true number of clusters, SSE will continue to decrease, and when the decrease slows down, the K-SSE curve is drawn to find the inflection point in the decrease, and the K value is determined.

[0019] S3 source domain data screening, the big data set of the basic building group is composed of daily load curves in units of one day, which is input to the clustering model, the historical daily load curves of the load target building energy consumption mode are screened, and the data migration training set and the model migration training set are constructed respectively. (1) Data migration training set: the selected historical data and the target building historical data are merged to construct the data migration training set. (2) Model migration training set: select the building with the highest proportion of historical data that meets the target building energy consumption mode as the most similar building, and select its historical data as the model migration training set.

[0020] S4 day-ahead load prediction model construction, the long short-term memory neural network (LSTM network) is trained by using the data migration training set to construct the data migration prediction model; the convolutional neural network-long short-term memory network (CNN-LSTM network) is trained by using the model migration training set, and then the network structure is fine-tuned by using the target building load data to construct the model migration prediction model. The mean of the outputs of the two models (data migration prediction model and model migration prediction model) is taken as the day-ahead prediction result of the target building.

[0021] S5 adaptive model optimization, in the running process of the day-ahead load prediction model, the model parameters are continuously adjusted by using Bayesian optimization according to the updated historical load, and the adaptive load prediction of the target building is realized. In this embodiment, Bayesian optimization is adopted, the selection of the hyperparameters of the deep learning network is often inaccurate and unreliable by using the empirical method, Bayesian optimization can find a better combination of hyperparameters with very few steps, and generally the derivative of the hyperparameters of the neural network is difficult to obtain, but Bayesian optimization does not require the derivative, therefore, the hyperparameters of the deep learning network are obtained by using Bayesian optimization.

[0022] The application is based on transfer learning and deep learning technology, and proposes an adaptive building day-ahead load forecasting method based on transfer learning for the case of lacking target building historical load data. Day-ahead refers to obtaining the load curve of the whole day on the target prediction day the day before, which belongs to multi-point prediction. First, the K-means++ clustering target building daily load data is used to obtain the target building energy mode and source domain data screening clustering model. The historical load of the building group is screened by using the clustering model, the data meeting the target building energy mode is combined with the target building daily load data to train the long short-term memory neural network (LSTM) to obtain the data transfer prediction model. At the same time, the building with the highest similarity to the target building is screened out, and the convolutional neural network-long short-term memory network (CNN-LSTM) is trained by using its historical load, and then the CNN-LSTM model is fine-tuned by using the target load data, so as to obtain the model transfer prediction model. In the running process of the day-ahead load forecasting model, the corresponding data set is updated continuously, the prediction model is adjusted by using the Bayesian optimization, and the adaptive load forecasting of the target building is realized. The prediction results of the target building load are obtained by comprehensively considering the prediction models of the two transfer modes, the load forecasting of the data-lacking target building is realized, and the prediction accuracy is improved.

[0023] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and they should be covered in the scope of the claims and the specification of the present application.

Claims

1. A method for adaptive building day-ahead load forecasting based on transfer learning, characterized in that, Comprising the following steps: S1 data acquisition and processing, obtaining target building data and basic building group data through smart meters, decomposing the original data set into small data set of target building and large data set of basic building group, obtaining historical load data with step of 1 hour; filling in the missing values of all original data sets, filling in the missing values with the average value of the corresponding time points of the previous and next two weeks; S2 energy use mode clustering, the target building data is divided into days, and the daily load curve is constructed; the best K value of K-means++ is found by elbow rule, and the target building daily load curve is clustered to obtain a clustering model; each cluster in the clustering result represents a kind of energy use mode of the target building; S3 source domain data screening, the large data set of the basic building group is divided into daily load curves with one day as a unit, which is input into the clustering model to screen the historical daily load curves of the target building energy use mode, and data migration training set and model migration training set are constructed respectively; S4 construction of day-ahead load forecasting model, long short-term memory neural network is trained by using data migration training set to construct data migration prediction model; convolutional neural network-long short-term memory network is trained by using model migration training set, and network structure is fine-tuned by using target building load data to construct model migration prediction model, and the average value of the outputs of data migration prediction model and model migration prediction model is taken as the day-ahead prediction result of target building; S5 adaptive model optimization, in the running process of day-ahead load forecasting model, the model parameters are continuously adjusted by using Bayesian optimization according to the updated historical load, so as to realize adaptive load forecasting of target building; The construction of data migration training set and model migration training set is as follows: (1) data migration training set, the selected historical data is combined with the target building historical data to construct data migration training set; (2) model migration training set, the building with the highest proportion of historical data conforming to the energy use mode of the target building is selected as the most similar building, and its historical data is selected as the model migration training set.

2. The adaptive building day-ahead load forecasting method based on transfer learning according to claim 1, characterized in that, The specific steps of elbow rule include: specifying an i value, which is the maximum possible cluster number; then increasing the cluster number from 1 to i, and calculating i SSE; according to the potential mode of data, when the set cluster number continuously approaches the true cluster number, SSE presents a rapid downward trend, and when the set cluster number exceeds the true cluster number, SSE will continue to decline, when the decline slows down, by drawing K-SSE curve, the inflection point in the decline is found, that is, the K value is determined.

Images