Construction method and application of resident community short-term load prediction model

Abstract

The invention belongs to the technical field of short-term load prediction, and discloses a residential community short-term load prediction model construction method and application, and the method comprises the steps: employing a data set to train a load prediction base model, and obtaining a trained short-term load prediction base model; wherein each training sample in the data set comprises load data and meteorological data of a day before a prediction day, two days before and seven days before the prediction day, and meteorological data of a prediction moment of the current day of the prediction day; and carrying out online incremental training on the trained short-term load prediction base model by utilizing real load data and meteorological data of the day before the prediction day and a meteorological data forecast value of the prediction moment of the day on the prediction day and adopting online learning, and taking the model obtained after the online incremental training is finished as a final short-term load prediction model. The method is used for predicting the target resident community load. According to the invention, the accuracy of short-term load prediction of the resident community can be improved, and the real-time performance of prediction is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of short-term load forecasting, and more specifically, relates to a method for constructing a short-term load forecasting model for a residential community and its application. Background Art

[0002] As the largest and most widely distributed type of load in the social electricity load structure, residential electricity load is an important part of the demand side in the balance of power grid supply and demand. With the improvement of residents' living standards, the number of household appliances used and the continuous increase in load capacity have directly led to an increase in the proportion of residential load in the power grid load. Forecasting residential users' electricity load can improve the distribution network's ability to cope with risks, reduce the cost of electricity production and reduce environmental pollution. Residential loads have the characteristics of diverse load equipment and different electricity consumption behaviors, which makes their load curves usually volatile and difficult to predict.

[0003] At present, artificial intelligence methods for short-term load forecasting are generally based on the progress of measurement technology, taking historical load data as input, and continuously training the forecasting model to obtain a forecasting model with better performance. However, in actual application, due to the dual volatility introduced by the randomness of residential users' electricity load and the volatility of community users, feature engineering built entirely based on historical load data often has a continuously reduced forecasting performance over time, requiring frequent retraining, wasting a lot of computing and storage resources, and poor real-time performance. Summary of the invention

[0004] In view of the above defects or improvement needs of the prior art, the present invention provides a method for constructing a short-term load forecasting model for a residential community and its application, which aims to improve the accuracy of short-term load forecasting for residential communities and improve the real-time performance of the forecast.

[0005] To achieve the above object, according to a first aspect of the present invention, a method for constructing a residential community short-term load forecasting model is provided, comprising:

[0006] The load forecasting base model is trained using a data set to obtain a trained short-term load forecasting base model; wherein the data set includes the historical load data and corresponding meteorological data of users in the target residential community, and each training sample in the data set includes: the load data and meteorological data of the day before the forecast day d d-1, the two days before the forecast day d-2, and the seven days before the forecast day d-7, as well as the meteorological data at the forecast time t on the forecast day d;

[0007] Using the actual load data and meteorological data of d-1 day before the prediction day d, as well as the forecast value of meteorological data at the prediction time t on the prediction day d, online learning is adopted to perform online incremental training on the trained short-term load forecasting base model, and the model obtained after the online incremental training is completed is used as the final short-term load forecasting model. The final short-term load forecasting model is used to predict the load of the target residential community at the prediction time t on the prediction day d.

[0008] Furthermore, the load forecasting base model is a parallel ensemble learning model based on XGBoost; the parallel ensemble learning model includes n parallel XGBoost sub-models;

[0009] The sample subsets used for training each XGBoost sub-model are constructed as follows:

[0010] One sample is randomly selected from the data set with replacement each time, and this is repeated multiple times to obtain a sample subset; after n rounds of random sampling with replacement, n sample subsets are obtained, which correspond to the sample subsets used for training each XGBoost sub-model respectively;

[0011] The output of the parallel ensemble learning model is obtained by voting on n parallel XGBoost sub-models.

[0012] Further, in the process of training the short-term load forecasting base model, the load data and meteorological data of the day before the forecast day d-1, the two days before the forecast day d-2, and the seven days before the forecast day d-7 are the load data and meteorological data corresponding to the time t-1, t, and t+1 of the day before the forecast day d-1, the two days before the forecast day d-2, and the seven days before the forecast day d-7; and also include the meteorological data corresponding to the time t-1 and t+1 of the forecast day d;

[0013] In the process of online incremental training of the trained short-term load forecasting base model using online learning, the real load data and meteorological data of the day before the forecast day d d-1 are the real load data and real meteorological data corresponding to the time t-1, t, and t+1 of the day before the forecast day d d-1; and also include the meteorological data forecast values ​​corresponding to the time t-1 and t+1 of the forecast day d.

[0014] Furthermore, in the process of training the short-term load forecasting base model, each training sample in the data set further includes:

[0015] Date type data of the day before the forecast day d-1, two days before d-2, and seven days before d-7, as well as date type data of the forecast day d;

[0016] The date type data is used to indicate whether the load data input on the corresponding day is working day load data or non-working day load data;

[0017] It also includes using the date type data of d-1 day before the prediction day d and the date type data of the prediction day d, adopting online learning, and performing online incremental training on the trained short-term load forecasting base model to obtain the final short-term load forecasting model.

[0018] Furthermore, before using the data set to train the load forecasting base model, the feature screening of the meteorological data in the training samples is also included, including:

[0019] The mutual information between the load data at the current moment and each meteorological feature in the corresponding meteorological data is calculated in turn; the meteorological features corresponding to the preset mutual information threshold are regarded as the meteorological feature data with greater influence, and are used to participate in the training of the load prediction base model.

[0020] According to a second aspect of the present invention, a method for short-term load forecasting in a residential community is provided, comprising:

[0021] The load data and meteorological data of the day before the forecast day d-1, two days before the forecast day d-2 and seven days before the forecast day d-7, as well as the forecast value of the meteorological data at the forecast time t on the forecast day d are input into the short-term load forecast model to obtain the load forecast value of the target residential community at the forecast time t on the forecast day d;

[0022] Wherein, the short-term load forecasting model is obtained by any of the above-mentioned methods for constructing a short-term load forecasting model for a residential community.

[0023] Furthermore, the load data and meteorological data are the load data and meteorological data corresponding to the time t-1, t, and t+1 of the day before the forecast day d-1, two days before the forecast day d-2, and seven days before the forecast day d-7; and also include the forecast values ​​of meteorological data corresponding to the time t-1 and t+1 of the forecast day d;

[0024] Or / and, the data input into the short-term load forecasting model further includes: date type data of the day d-1 before the forecast day d, the two days before d-2 and the seven days before d-7, and the date type data of the forecast day d;

[0025] The date type data is used to indicate whether the load data input on the corresponding day is working day power load data or non-working day power load data.

[0026] According to a third aspect of the present invention, there is provided an electronic device comprising a computer-readable storage medium and a processor;

[0027] The computer-readable storage medium is used to store executable instructions;

[0028] The processor is used to read the executable instructions stored in the computer-readable storage medium to execute any one of the above-mentioned methods for constructing a residential community short-term load forecasting model, or / and to execute the above-mentioned method for short-term load forecasting in a residential community.

[0029] According to a fourth aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the method for constructing a residential community short-term load forecasting model as described in any one of the above items, or / and the method for short-term load forecasting in a residential community as described in the above items.

[0030] According to a fifth aspect of the present invention, a computer program product is provided. When the computer program product is run on a computer, the computer executes any one of the above-mentioned methods for constructing a residential community short-term load forecasting model, or / and executes the above-mentioned method for short-term load forecasting in a residential community.

[0031] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0032] (1) The present invention takes into account the dual volatility introduced by the randomness of residential users' electricity load and the volatility of community users. The "fixed" model trained directly based on historical data cannot accurately reflect the current characteristics of residential community load, which leads to problems such as reduced prediction accuracy and the need for frequent retraining. Based on this, the present invention first uses the past historical data of the target residential community users to train the base model, so that the base model first learns the law of community residents' electricity load in the past historical period. After obtaining the basic model, for each forecast day, the base model is updated online using the real load data and meteorological data known one day before the forecast day and the forecast value of meteorological data on the forecast day to learn the small fluctuation characteristics of residents' electricity consumption. In this way, the final model obtained can adapt to the fluctuation characteristics of residential community load, improve the prediction accuracy of residential community load, and do not need to re-train the entire offline for the small fluctuation, which also improves the real-time performance of the prediction. Moreover, each training sample constructed by the present invention also includes the load data and meteorological data of d-1, two days before the prediction day d, d-2, and seven days before the prediction day d, as well as the meteorological data at the prediction time t on the prediction day d. The present invention believes that these data are all key data that affect the load accuracy at the prediction time on the prediction day. Based on this, the prediction model is trained to enable the prediction model to fully learn the electricity load patterns of community residents to improve the accuracy of short-term load prediction of the residential community. Experiments have also confirmed that the present invention has good prediction performance, faster prediction speed, and lower computing resources and storage requirements.

[0033] (2) As a preferred method, the load forecasting base model adopts a parallel ensemble learning model based on XGBoost, which can take advantage of the high prediction accuracy and high computational efficiency of XGBoost and improve the generalization ability of the XGBoost ensemble model. Specifically, the construction of the sample subset corresponding to the parallel ensemble sub-model makes the input data extracted each time independent of each other, and the sample data distribution of each constructed sub-model is the same. Various random situations are simulated in the process of model integration, which reduces the risk of high errors, effectively reduces the output variance of the ensemble model (base model), reduces the degree of model overfitting, and improves the generalization ability of the model.

[0034] (3) As a preferred embodiment, the present invention takes into account that the power load data at a certain moment does not actually have "memory", that is, it is irrelevant to the past change process and is only related to the previous moment or the next moment. Therefore, only the relevant data of the three moments adjacent to the prediction moment including the prediction moment are selected to train the model, which can greatly reduce the dimension of the input matrix, increase the model training speed, reduce the interference of irrelevant noise, and improve the prediction accuracy; and, it also includes the meteorological data forecast values ​​corresponding to the prediction day d at t-1 and t+1, which can avoid the accidental influence brought by using only the meteorological data forecast values ​​corresponding to the prediction moment.

[0035] (4) Furthermore, the present invention also takes into account the difference in electricity load between community residents on working days and non-working days, and uses date type data as one of the influencing factors to train the prediction model, thereby further improving the accuracy of short-term load prediction for the residential community.

[0036] (5) Furthermore, for the short-term power load forecasting of residential communities, meteorological data is a type of influencing factor with a large degree of influence and a large number of features. By screening the features of meteorological data, it is possible to avoid the meteorological features that do not contribute to the forecasting task interfering with the learning of the forecasting model. The use of meteorological feature data after feature screening can reduce the complexity of the data, reduce the feature dimension, reduce the computing cost, and improve the performance and interpretability of the forecasting model. When performing feature screening, the randomness of the load data sequence and the meteorological data, as well as the linear and nonlinear correlation between the load data sequence and the meteorological data, are taken into account. The mutual information calculation method is used to comprehensively analyze the feature correlation degree, which can achieve the most valuable information in the feature data while reducing the number of redundant features. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of a method for constructing a short-term load forecasting model for a residential community in an embodiment of the present invention.

[0038] Figure 2 The flowchart of XGBoost-based parallel ensemble learning training in an embodiment of the present invention.

[0039] Figure 3 This is a flowchart of an online learning update iteration in an embodiment of the present invention.

[0040] Figure 4 The figure is an overall flow chart of the method for short-term load forecasting in a residential community in an embodiment of the present invention.

[0041] Figure 5 It is a test performance diagram of the method in the embodiment of the present invention and two control groups. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0043] Example 1

[0044] like Figure 1 As shown, in an embodiment of the present invention, a method for constructing a short-term load forecasting model for a residential community is provided, comprising:

[0045] The load forecasting base model is trained using a data set to obtain a trained short-term load forecasting base model; wherein the data set includes the historical load data and corresponding meteorological data of users in the target residential community. In the embodiment of the present invention, the relevant historical data of the year before the forecast year is selected to carry out the training of the base model; each training sample in the data set includes: the load data and meteorological data of the day before the forecast day d d-1, the two days before the forecast day d-2 and the seven days before the forecast day d-7, and the meteorological data of the forecast time t on the forecast day d;

[0046] Using the real load data and meteorological data known one day before the forecast day, as well as the forecast value of meteorological data at prediction time t on prediction day d, the trained short-term load forecasting base model is incrementally trained online through online learning, and the model after online training is used as the final short-term load forecasting model. The final short-term load forecasting model is used to predict the short-term load of the target residential community at prediction time t on prediction day d.

[0047] Preferably, the load data and meteorological data of the day before the forecast day d d-1, two days before the forecast day d-2 and seven days before the forecast day d-7 in each training sample are the load data and meteorological data corresponding to the time t-1, t, and t+1 of the day before the forecast day d d-1, two days before the forecast day d-2 and seven days before the forecast day d-7; and also include the meteorological data corresponding to the time t-1 and t+1 of the forecast day d. In the process of online incremental training of the trained short-term load forecasting base model using online learning, the real load data and meteorological data of the day before the forecast day d d-1 are the real load data and real meteorological data corresponding to the day before the forecast day d d-1 at the time t-1, t, and t+1; and also include the forecast values ​​of meteorological data corresponding to the time t-1 and t+1 of the forecast day d. The present invention takes into account that the power load data at a certain moment does not actually have "memory", that is, it is irrelevant to the past change process and is only related to the previous moment or the next moment. Therefore, only selecting relevant data at three moments adjacent to the prediction moment including the prediction moment can greatly reduce the input matrix dimension, increase the model training speed, reduce the interference of irrelevant noise, and improve the prediction accuracy.

[0048] Furthermore, each training sample also includes date type data of the day before, two days before, and seven days before the prediction day d, as well as date type data of the prediction time t on the prediction day d; and also includes using the date type data of the day before the prediction day d-1, and the date type data on the prediction day d, to perform online incremental training on the trained short-term load forecasting base model using online learning to obtain the final short-term load forecasting model. The date type data is used as one of the influencing factors of the prediction to determine whether the input sample is a working day electricity load or a non-working day electricity load. The present invention also takes into account the difference between the working day electricity load and the non-working day electricity load of community residents, and uses the date type data as one of the influencing factors to train the prediction model, further improving the accuracy of the short-term load prediction of the residential community.

[0049] Furthermore, before the training samples are input into the load forecasting base model for training, the training samples are preprocessed, specifically including: data outlier identification and missing value supplementation.

[0050] In the embodiment of the present invention, the 3σ principle is used to identify data outliers, that is, an outlier is defined as a data point whose value deviates from the mean by more than 3 standard deviations. Specifically, descriptive statistics are performed on the training sample data (generally referring to load data), and then the mean and standard deviation of the data are calculated; data outside the range of the mean plus or minus 3 standard deviations are considered to be outliers; after the outliers are identified, they are eliminated and supplemented together when the missing values ​​are supplemented later.

[0051] The missing data value supplementation is aimed at the missing statistical values ​​of the original data and the missing values ​​after the outliers are eliminated. In the embodiment of the present invention, the cubic spline interpolation method is used to supplement the missing data.

[0052] After preprocessing the training samples, in order to prevent meteorological features that have no contribution to the prediction task from interfering with the learning of the prediction model, the meteorological data is also screened for features, and meteorological data with strong correlation with the prediction is selected as the input of the model. Specifically, it includes:

[0053] The mutual information between the current load data and each feature of the corresponding meteorological data is calculated in turn, and the meteorological data features corresponding to the mutual information values ​​not less than the preset threshold are regarded as the meteorological feature data with greater influence and participate in the subsequent model prediction tasks.

[0054] In the embodiment of the present invention, the mutual information calculation is performed on the load data at the current moment and each feature of the corresponding meteorological data in turn, and the calculation formula is:

[0055]

[0056] Among them, I(X;Y) is the mutual information of random variables X and Y, random variable X is the historical load data sequence of the target residential community users, and random variable Y is the corresponding meteorological data, including air temperature, dew point, perceived temperature, atmospheric pressure, relative humidity, visibility, cloud cover, wind speed, wind direction, precipitation, precipitation probability, weather type and other meteorological characteristic data. p(x,y) is the joint probability density distribution function of X and Y; p(x) and p(y) are the marginal probability density distribution functions of X and Y respectively.

[0057] Before comparing the MI values ​​between X and Y, they must be normalized to make their value ranges uniform; the expression of the normalized mutual information value (NMI) of the two can be defined as:

[0058]

[0059] Among them, H(X) and H(Y) are the information entropy of random variables X and Y, which can be defined as:

[0060]

[0061] In the embodiment of the present invention, the calculation results are shown in Table 1:

[0062] Table 1 Mutual information values ​​between various meteorological characteristics and community loads

[0063]

[0064] Excluding meteorological factors with very small normalized mutual information values, four-dimensional meteorological influencing factors (meteorological characteristic data) including air temperature, dew point, perceived temperature, and relative humidity were selected as the considerations for subsequent feature engineering construction. These meteorological factors are related to the daily lives of residents, thus affecting the electricity consumption behavior of residential users.

[0065] For short-term power load forecasting in residential communities, meteorological data is a type of influencing factor with a large degree of influence and a large number of features. By screening the features of meteorological data, it is possible to avoid the meteorological features that do not contribute to the forecasting task interfering with the learning of the forecasting model. The use of meteorological feature data after feature screening can reduce the complexity of the data, reduce the feature dimension, reduce the computing cost, and improve the performance and interpretability of the forecasting model. When performing feature screening, the randomness of the load data sequence and the meteorological data, as well as the linear and nonlinear correlation between the load data sequence and the meteorological data are taken into account. The mutual information calculation method is used to comprehensively analyze the degree of feature correlation, which can achieve the most valuable information in the feature data while reducing the number of redundant features.

[0066] In the embodiment of the present invention, Figure 2 As shown in the figure, the short-term load forecasting base model adopts a parallel ensemble learning model based on the XGBoost (Extreme Gradient Boosting) algorithm, which combines the gradient boosting machine (Gradient Boosting Machine, GBM) and the decision tree ensemble method, and is optimized and developed based on the gradient boosting decision tree (Gradient Boosting Decision Tree, GBDT), which greatly improves the accuracy and computational efficiency of the prediction model.

[0067] The loss function in the prediction process includes the residual between the model prediction value and the true value, as well as the sum of the complexity of each decision tree in XGBoost, which is used to limit the overfitting of the model. XGBoost uses an incremental training method, that is, the original model is maintained during the training process, and a new item (i.e. a new tree) is added to the objective function each time, and each round of new trees added needs to reduce the objective function as much as possible.

[0068] Although XGBoost is an algorithm with strong learning ability and high precision, and has certain generalization ability and anti-overfitting ability. However, the serial iteration process of XGBoost causes the entire tree model to overgrow, resulting in a large degree of sample dependence, and there is still a risk of overfitting. In addition, the XGBoost algorithm lacks randomness, and there may be a risk of high error when changing the data source. Therefore, in an embodiment of the present invention, an improved load prediction base model based on the XGBoost algorithm is provided, which specifically includes:

[0069] The load forecasting base model includes n parallel sub-models, each of which uses the XGBoost algorithm; the input of each sub-model is the corresponding sample subset; wherein the construction method of the sample subset corresponding to each sub-model includes:

[0070] The random samples are obtained by random sampling. Specifically, one sample is randomly selected with replacement from the training sample set after preprocessing and feature screening, and this is repeated m times to obtain a sample subset containing m samples.

[0071] After n rounds of random sampling with replacement, n sample subsets are obtained, each of which contains m samples; the n sample subsets serve as the input of the corresponding sub-model.

[0072] During the training process, n sub-models are obtained based on the n sample subsets, which are XGBoost_1 to XGBoost_n in the embodiment of the present invention. Finally, the n prediction sub-models are integrated and combined through a voting combination strategy.

[0073] Through this parallel ensemble learning improvement, the input data extracted each time are independent of each other, and the sample data distribution of each constructed sub-model is the same. Various random situations are simulated in the process of model integration, which reduces the risk of high errors, effectively reduces the output variance of the integrated model (base model) and reduces the degree of model overfitting, improves the generalization ability of the model, and avoids the poor performance of the prediction model on new users after training with partial user data.

[0074] The traditional test application method of offline learning, also known as batch learning, refers to training independent data and using the trained model for prediction tasks. Once the data changes, retraining is required. However, considering the short-term load forecasting of residential communities, there is a double volatility introduced by the randomness of residential users' electricity consumption and the volatility of community users. This "fixed" model trained based on historical data cannot accurately reflect the current characteristics of residential community loads, resulting in reduced prediction accuracy and the need for frequent retraining.

[0075] Therefore, if Figure 3 As shown, in the embodiment of the present invention, after the training of the short-term load forecasting base model is completed, the real load data and influencing factor data (meteorological data) known one day before the forecast day are used to perform online incremental training on the basis of the base model, and the new model updated after online learning is used as the final short-term load forecasting model to predict the residential community load on the forecast day. When the base model is trained online, the difference between the output predicted by the model and the true value is minimized, and the model parameters and structure are adjusted. When convergence occurs, a new model updated after online learning is obtained.

[0076] Generally speaking, online learning training is mostly used to train models with a large sample size of input data, while offline training is mostly used for models with a small amount of data to obtain higher prediction accuracy. However, for the training of short-term load forecasting models for residential communities with relatively small input data, considering the randomness of residential users' electricity load and the volatility of community users, after obtaining the base model, direct offline training requires frequent retraining, which wastes a lot of computing and storage resources, and the real-time performance is not good. Therefore, the present invention applies online learning training to short-term prediction of residential community load, processes constantly changing data through online learning, and allows the model to be updated in real time to adapt to changes in residential community load demand. When emergencies or seasonal changes occur, online learning will show greater advantages, giving full play to the real-time and good dynamic adjustment capabilities of online learning. For example, with changes in residents' lifestyles, the introduction of new equipment, or changes in the migration of users in the community, the model can quickly adapt and adjust.

[0077] The model can be updated online using the real load data and meteorological data known one day before the forecast day, without the need to retrain the entire model. This means that the forecast accuracy can be gradually improved without requiring a large amount of computing resources, thereby reducing the demand for computing and storage resources.

[0078] The method of the present invention is described below with reference to specific examples.

[0079] In the embodiment of the present invention, the intelligent load measurement database of the MUELLER green energy community under the Pecan Street energy project in the United States is used. The residential community is located at about 30° north latitude, with an average summer temperature of about 33°C, and uses refrigeration equipment such as air conditioners. The time span of the database is from January 1, 2014 to March 30, 2015, and the number of users involved is large, but the data of some users is missing. After a certain amount of data screening and sorting, this article finally uses the data of 71 residential users as the object for subsequent related analysis. Among them, the load data is collected through non-invasive load measurement devices at the interfaces of the entire house and various household electrical sockets, and the collection frequency is 15 minutes. In addition to the total load of each household user, the load of important household appliances and the load of public sockets in some rooms are also recorded in the database. In addition, the database also contains the meteorological data of the MUELLER green energy community within the corresponding date range, and the collection frequency of the meteorological data is 1 hour.

[0080] A corresponding prediction model is established for each prediction time point of the prediction day, and every 15 minutes is regarded as a time point. For a prediction day, a total of 96 short-term load prediction models are established to predict the load at the corresponding time point. For each short-term load prediction model, the method in the above embodiment is used to construct it, such as Figure 4 As shown in . Each sample of the model input is the load, weather and date type data corresponding to three consecutive moments (t-1, t, t+1) adjacent to the prediction time t 1, 2, and 7 days before the prediction date, as well as the forecast value of the weather data and the date type value at the prediction time on the prediction day, as shown in the following vector a. In particular, when the prediction time is the first time point of the day, the previous time point of the model input is postponed to the last time point of the previous day, as shown in the following vector b; or, when the prediction time is the last time point of the day, the last time point of the model input is postponed to the first time point of the next day, as shown in the following vector c.

[0081] a=[y d-1,t-1 ,y d-1,t ,y d-1,t+1 ,w d-1,t-1 ,w d-1,t ,w d-1,t+1 ,date d-1 ...

[0082] y d-2,t-1 ,y d-2,t ,y d-2,t+1 ,w d-2,t-1 ,w d-2,t ,w d-2,t+1 ,date d-2 ...

[0083] y d-7,t-1 ,y d-7,t ,y d-7,t+1 ,w d-7,t-1 ,w d-7,t ,w d-7,t+1 ,date d-7 ...

[0084] w d,t-1 ,w d,t ,w d,t+1 ,date d ]

[0085] b=[y d-2,24 ,y d-1,1 ,y d-1,2 ,w d-2,24 ,w d-1,1 ,w d-1,2 ,date d-1 ...

[0086] y d-3,24 ,y d-2,1 ,y d-2,2 ,w d-3,24 ,w d-2,1 ,w d-2,2 ,date d-2 ...

[0087] y d-8,24 ,y d-7,1 ,y d-7,2 ,w d-8,24 ,w d -7,1,w d-7,2 ,date d-7 ...

[0088] w d-1,24 ,w d,1 ,w d,2 ,date d ]

[0089] c=[y d-1,23 ,y d-1,24 ,y d-1,1 ,w d-1,23 ,w d-1,24 ,w d,1 ,date d-1 ...

[0090] y d-2,23 ,y d-2,24 ,y d-1,1 ,w d-2,23 ,w d-2,24 ,w d-1,1 ,date d-2 ...

[0091] y d-7,23 ,y d-7,24 ,y d-6,1 ,w d-7,23 ,w d-7,24 ,w d-6,1 ,date d-7 ...

[0092] w d,23 ,w d,24 ,w d+1,1 ,date d ]

[0093] Among them, d is the forecast day, t is the forecast time, y is the load data, with a subscript indicating the load data of the corresponding day and time, w is the meteorological data, with a subscript indicating the meteorological data of the corresponding day and time, and date is the date type data, with a subscript indicating the date type data of the corresponding day and time.

[0094] The model output is the load value y at the predicted time d,tThe model sets the training set time length to be the data of the whole year of 2014, and the test set to be the data from January to March 2015. 71 residential users are taken as the research objects of the residential community, and the loads of 71 households are superimposed to obtain the total load curve of the community to complete the data acquisition. After the above preprocessing and feature screening (as shown in Table 1), the processed data set is obtained to construct the subsequent n parallel XGBoost sub-models, and the n parallel XGBoost sub-models constitute the above short-term load forecasting base model. Then online learning is performed to obtain the trained short-term load forecasting model.

[0095] In order to fully illustrate the advantages of the method of the present invention, two control groups are set up to compare with the method proposed in the present invention on the test set. Control group 1 directly uses the short-term load forecasting base model to make predictions on the test set, and re-selects the historical data of the year before the current month for offline training every month, which is used for the load prediction of the current month. Control group 2 reconstructs the historical data of the year before the current day for re-training every day, which is used for the load prediction of the current day, that is, a new model is trained offline every day. The present invention is based on the short-term load forecasting base model. After obtaining the real load data of each new day, only the real data of this day is used before the prediction of the next day to perform incremental training and update on the base model, which is used for the load prediction of the next day.

[0096] In addition, in order to simulate random situations such as user changes in residential communities, the present invention randomly removes 20 users from the total community load in the last 40 days as interference in the test set of 82 days, and then tests the model performance obtained in the above three ways to illustrate the dynamic adaptability of different models. The mean absolute percentage error (MAPE) of the test results is shown in Table 2, and the MAPE of each day before and after the test set is interfered and for all 82 days is averaged. The specific prediction performance is shown in Figure 5 shown.

[0097] Table 2. MAPE performance of prediction model

[0098] The present invention Control group 1 Control group 2 42 days before the disruption 13.53 12.70 12.66 40 days after the disturbance 21.66 45.33 21.69 Total 82 days 17.49 28.62 17.06

[0099] The time consumed and memory requirements of the three models for prediction are ranked as control group 2> present invention> control group 1. It can be seen that under the ideal situation that there is no change in users in the residential community, the prediction results of the present invention and the other two control models are similar. However, in actual residential communities, there must be fluctuations in electricity users such as renting and terminating, business trips, traveling and visiting relatives. Under the interference of simulated user changes, the present invention shows good prediction performance. There is no need to re-train a whole year of data offline every day, and large-scale matrix operations are avoided. Compared with the control group 2, there is still a certain improvement in prediction performance, and the prediction speed of the present invention is faster, and the computing resources and storage requirements are lower, showing good prediction performance.

[0100] The present invention uses big data technology to collect data on the electricity consumption characteristics of residential communities and influencing factors such as meteorology, and then uses intelligent algorithms to predict future residential community loads. Parallel ensemble learning and online learning are introduced to improve the generalization performance of artificial intelligence prediction models, improve the dynamic adaptability of prediction models, and solve the dual uncertainty problem in short-term load prediction of residential communities. It has good practical application capabilities. After comparative testing, this method can effectively respond to community load disturbances, greatly save computing resources and storage space, and improve prediction speed, which is of great significance for the dispatching management of distribution networks and the participation of residential users in demand response.

[0101] Example 2

[0102] An embodiment of the present invention provides a method for short-term load forecasting in a residential community, comprising:

[0103] The load data and meteorological data of the day before, two days before, and seven days before the forecast day d, as well as the forecast value of meteorological data at the forecast time t on the forecast day d, are input into the short-term load forecast model to obtain the short-term load forecast value of the target residential community at the forecast time t on the forecast day d; wherein the short-term load forecast model is constructed by the residential community short-term load forecast model construction method in Example 1. For the relevant technical solutions, please refer to the description in Example 1, which will not be repeated here.

[0104] Preferably, the load data and meteorological data of the day, two days and seven days before the forecast day d are the load data and meteorological data corresponding to the time t-1, t and t+1 of the day, two days and seven days before the forecast day d; and also include the meteorological data forecast values ​​corresponding to the time t-1 and t+1 of the forecast day d.

[0105] Furthermore, the data input into the short-term load forecasting model also includes date type data of the day before, two days before and seven days before the forecast day d, and date type data of the forecast time t on the forecast day d; the date type data is used as one of the influencing factors of the forecast to determine whether the input load data is the working day electricity load or the non-working day electricity load.

[0106] Example 3

[0107] An embodiment of the present invention provides an electronic device, including a computer-readable storage medium and a processor;

[0108] Computer-readable storage media for storing executable instructions;

[0109] The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the residential community short-term load forecasting model construction method in the above-mentioned embodiment 1, or / and, to execute the residential community short-term load forecasting method in embodiment 2. For related technical solutions, please refer to the corresponding descriptions in embodiments 1 and 2, which will not be repeated here.

[0110] Example 4

[0111] The embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the program is executed by a processor, the method for constructing a residential community short-term load forecasting model as in the above-mentioned embodiment 1 is implemented, or / and the method for forecasting residential community short-term load as in the above-mentioned embodiment 2 is implemented. For the relevant technical solutions, please refer to the corresponding descriptions in embodiments 1 and 2, which will not be repeated here.

[0112] Example 5

[0113] The embodiment of the present invention provides a computer program product. When the computer program product is run on a computer, the computer executes the method for constructing a residential community short-term load forecasting model in the above embodiment 1, or / and executes the method for forecasting residential community short-term load in the above embodiment 2. For the relevant technical solutions, please refer to the corresponding descriptions in Embodiment 1 and Embodiment 2, which will not be repeated here.

[0114] In summary, the present invention utilizes the advantages of real-time increments of online learning to train the short-term load forecasting base model to adapt to the fluctuation characteristics of residential electricity consumption, giving full play to the powerful nonlinear fitting ability of artificial intelligence algorithms; and utilizes parallel integrated learning methods to enhance the generalization ability of the XGBoost model, thereby improving the prediction accuracy and practical application capabilities of residential community load forecasting, and providing reliable technical guidance for residential users to participate in demand response and distribution network dispatching management.

[0115] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing a short-term load forecasting model for a residential community, characterized in that: include: The load forecasting base model is trained using a data set to obtain a trained short-term load forecasting base model; wherein the data set includes the historical load data and corresponding meteorological data of users in the target residential community, and each training sample in the data set includes: the load data and meteorological data of the day before the forecast day d d-1, the two days before the forecast day d-2, and the seven days before the forecast day d-7, as well as the meteorological data at the forecast time t on the forecast day d; Using the actual load data and meteorological data of d-1 day before the prediction day d, as well as the forecast value of meteorological data at the prediction time t on the prediction day d, online learning is adopted to perform online incremental training on the trained short-term load forecasting base model, and the model obtained after the online incremental training is completed is used as the final short-term load forecasting model. The final short-term load forecasting model is used to predict the load of the target residential community at the prediction time t on the prediction day d.

2. The method for constructing a residential community short-term load forecasting model according to claim 1, characterized in that: The load forecasting base model is a parallel ensemble learning model based on XGBoost; the parallel ensemble learning model includes n parallel XGBoost sub-models; The sample subsets used for training each XGBoost sub-model are constructed as follows: One sample is randomly selected from the data set with replacement each time, and this is repeated multiple times to obtain a sample subset; after n rounds of random sampling with replacement, n sample subsets are obtained, which correspond to the sample subsets used for training each XGBoost sub-model respectively; The output of the parallel ensemble learning model is obtained by voting on n parallel XGBoost sub-models.

3. The method for constructing a residential community short-term load forecasting model according to claim 1, characterized in that: In the process of training the short-term load forecasting base model, the load data and meteorological data of the day before the forecast day d-1, the two days before the forecast day d-2 and the seven days before the forecast day d-7 are the load data and meteorological data corresponding to the time t-1, t, and t+1 of the day before the forecast day d-1, the two days before the forecast day d-2 and the seven days before the forecast day d-7; and also include the meteorological data corresponding to the time t-1 and t+1 of the forecast day d; In the process of online incremental training of the trained short-term load forecasting base model using online learning, the real load data and meteorological data of the day before the forecast day d d-1 are the real load data and real meteorological data corresponding to the time t-1, t, and t+1 of the day before the forecast day d d-1; and also include the meteorological data forecast values ​​corresponding to the time t-1 and t+1 of the forecast day d.

4. The method for constructing a residential community short-term load forecasting model according to any one of claims 1 to 3, characterized in that: In the process of training the short-term load forecasting base model, each training sample in the data set further includes: Date type data of the day before the forecast day d-1, two days before d-2, and seven days before d-7, as well as date type data of the forecast day d; The date type data is used to indicate whether the load data input on the corresponding day is working day load data or non-working day load data; It also includes using the date type data of d-1 day before the prediction day d and the date type data of the prediction day d, adopting online learning, and performing online incremental training on the trained short-term load forecasting base model to obtain the final short-term load forecasting model.

5. The method for constructing a residential community short-term load forecasting model according to claim 4, characterized in that: Before using the data set to train the load forecasting base model, the meteorological data in the training samples are also screened for features, including: The mutual information between the load data at the current moment and each meteorological feature in the corresponding meteorological data is calculated in turn; the meteorological features corresponding to the preset mutual information threshold are regarded as the meteorological feature data with greater influence, and are used to participate in the training of the load prediction base model.

6. A method for short-term load forecasting in a residential community, characterized in that: include: The load data and meteorological data of the day before the forecast day d-1, two days before the forecast day d-2 and seven days before the forecast day d-7, as well as the forecast value of the meteorological data at the forecast time t on the forecast day d are input into the short-term load forecast model to obtain the load forecast value of the target residential community at the forecast time t on the forecast day d; Wherein, the short-term load forecasting model is obtained by the method for constructing a short-term load forecasting model for a residential community as described in any one of claims 1-5.

7. The method for short-term load forecasting in residential communities according to claim 6, characterized in that: The load data and meteorological data are the load data and meteorological data corresponding to the time t-1, t, and t+1 of the day before the forecast day d-1, two days before the forecast day d-2, and seven days before the forecast day d-7; and also include the forecast values ​​of meteorological data corresponding to the time t-1 and t+1 of the forecast day d; Or / and, the data input into the short-term load forecasting model further includes: date type data of the day d-1 before the forecast day d, the two days before d-2 and the seven days before d-7, and the date type data of the forecast day d; The date type data is used to indicate whether the load data input on the corresponding day is working day power load data or non-working day power load data.

8. An electronic device, characterized in that: comprising a computer readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the method for constructing a residential community short-term load forecasting model as described in any one of claims 1-5, or / and to execute the method for short-term load forecasting in a residential community as described in claim 6 or 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the method for constructing a residential community short-term load forecasting model as described in any one of claims 1-5, or / and, it implements the method for forecasting residential community short-term load as described in claim 6 or 7.

10. A computer program product, characterized in that When the computer program product runs on a computer, the computer executes the method for constructing a residential community short-term load forecasting model as described in any one of claims 1 to 5, or / and executes the method for forecasting residential community short-term load as described in claim 6 or 7.

Images