CN112734135B - A power load forecasting method, intelligent terminal and computer-readable storage medium - Google Patents

Abstract

The invention discloses a power load prediction method, an intelligent terminal and a computer readable storage medium, wherein the method comprises the following steps: acquiring prediction characteristic data corresponding to time to be predicted, wherein the prediction characteristic data comprises meteorological data and a time type; inputting the predicted characteristic data into a trained classification model and classifying the electricity utilization mode of the time to be predicted through the classification model to obtain a predicted electricity utilization mode corresponding to the time to be predicted; inputting the prediction characteristic data into a trained load prediction model and predicting the electric load of the time to be predicted through the load prediction model to obtain an initial electric load curve corresponding to the time to be predicted; and determining a predicted power load curve corresponding to the time to be predicted according to the predicted power consumption mode and the initial power load curve. The invention can accurately predict the electrical load.

Description

A power load forecasting method, intelligent terminal and computer-readable storage medium

technical field

The present invention relates to the technical field of power data analysis, and in particular, to a power load prediction method, an intelligent terminal and a computer-readable storage medium.

Background technique

Since electric energy cannot be stored, in order to maintain the frequency of the grid and ensure a balanced supply and demand relationship between power generation and power consumption, the power system must forecast the power load in advance, and formulate a power generation plan based on the forecast results. The research purpose of power system load forecasting is to improve the forecasting accuracy and provide safe and reliable power supply at the lowest possible operating cost. Since there are many power consumption factors on the user side, that is, the power consumption end, in the daily power generation process, it is necessary to ensure that some thermal power units are in a rotating standby state. If the electricity load on the user side suddenly decreases, the thermal power unit will shed load, which will greatly affect the security of the power grid and cause huge energy waste. Therefore, short-term load forecasting is an important task of the power system. The accuracy of the forecast results directly affects the stability of the power system, as well as the operating cost and power grid security of power grid companies.

However, due to many external factors such as climate change, social activities, and residents' living habits, the forecasting of power load is highly nonlinear and unpredictable. Therefore, accurate short-term power load forecasting has always been a difficult problem in the power industry. Therefore, it is of great significance to develop a short-term power load high-precision forecasting method for reducing energy waste and optimizing power system operation.

SUMMARY OF THE INVENTION

The main purpose of the present invention is to provide a power load forecasting method, an intelligent terminal and a computer-readable storage medium, aiming to solve the problem of low accuracy of short-term power consumption forecasting in the prior art.

In order to achieve the above object, the present invention provides a power load prediction method, and the power load prediction method includes the following steps:

Obtaining prediction feature data corresponding to the time to be predicted, wherein the prediction feature data includes meteorological data and time type;

Inputting the predicted feature data into a trained classification model and classifying the electricity consumption pattern of the time to be predicted by the classification model, to obtain a predicted electricity consumption pattern corresponding to the time to be predicted;

Inputting the prediction feature data into a trained load prediction model and predicting the electricity load for the to-be-predicted time by using the load prediction model, to obtain an initial electricity-use load curve corresponding to the to-be-predicted time to predict the electricity load;

The power consumption load is predicted according to the predicted power consumption mode and the initial power consumption load curve, and the predicted power consumption load curve corresponding to the to-be-predicted time is determined.

Optionally, in the power load forecasting method, the classification model includes several classification decision trees trained based on random forest algorithm; the training process of the classification decision tree specifically includes:

Obtain historical load data and historical feature data corresponding to each preset historical time;

Performing clustering processing on the historical load data to generate a target cluster set, wherein the target cluster set includes several sets of electricity consumption patterns, and the set of electricity consumption patterns includes historical load data corresponding to the same electricity consumption mode;

For each of the set of power consumption patterns, according to the historical time corresponding to each historical load data in the set of power consumption patterns, the historical feature data is marked to obtain the power consumption pattern corresponding to each of the historical feature data;

For each preset decision tree, select the training feature data in the historical feature data to input into the decision tree and split the decision tree according to the Gini index, until the power consumption corresponding to the historical feature data of each node in the decision tree The pattern is the same, and the classification decision tree is obtained.

Optionally, in the power load forecasting method, the performing clustering processing on the historical load data to generate a target cluster set specifically includes:

According to the current number of clusters, randomly determine the initial cluster center in the historical load data, wherein, when clustering is performed for the first time, the number of clusters is 2;

According to the preset fuzzy C-means clustering algorithm and the initial clustering center, calculate the k-th intermediate membership degree matrix corresponding to the historical load data, wherein k is the number of times of clustering;

When the number of clusters is less than the preset number of clusters threshold, the number of clusters is increased by one, and the kth intermediate membership matrix is repeatedly determined;

When the number of clusters is greater than or equal to the threshold of the number of clusters, determine the target membership degree matrix from the first intermediate membership degree matrix to the (K-1)th intermediate membership degree matrix, where K is the clustering degree matrix. class number threshold;

According to the target membership degree matrix and the cluster center corresponding to the target membership degree matrix, the historical load data is clustered to obtain a target cluster set.

Optionally, the power load prediction method, wherein, when the number of clusters is greater than or equal to the threshold of the number of clusters, determine the first intermediate membership degree matrix to the (K-1)th intermediate membership. The target membership degree matrix in the degree matrix, including:

For each of the intermediate membership degree matrices from the first intermediate membership degree matrix to the (K-1)th intermediate membership degree matrix, calculate the clustering validity index corresponding to the intermediate membership degree matrix;

According to the clustering effectiveness index, a set of target clusters in the first intermediate membership degree matrix to the (K-1)th intermediate membership degree matrix is determined.

Optionally, in the power load forecasting method, the load forecasting model includes several load forecasting sub-models, and the load forecasting sub-models are in one-to-one correspondence with the electricity consumption patterns.

Optionally, in the power load forecasting method, the historical load data includes a historical load curve corresponding to each historical time, and the historical load curve includes a historical load characteristic value; The training process includes specifically:

For each preset initial model, input the training curve data into the training feature data corresponding to the initial model in the historical feature data in the initial model, and classify the historical feature data of the training curve data through the initial model to obtain Each training load characteristic value corresponding to the historical characteristic data of the training curve data, wherein the training curve data is the data corresponding to the initial model in the historical characteristic data;

According to the training load characteristic value and the historical load characteristic value corresponding to the historical characteristic data, parameter optimization of the initial model is performed until the initial model converges, and a loss load prediction sub-model is obtained.

Optionally, in the power load forecasting method, wherein the initial power consumption curve prediction power consumption load includes a candidate estimated power consumption load curve corresponding to each load prediction sub-model; Input the trained load prediction model and use the load prediction model to predict the electricity load for the time to be predicted, and obtain the initial electricity load curve corresponding to the to-be-predicted time to predict the electricity load, which specifically includes:

For each of the load prediction sub-models, input the predicted characteristic data into the load prediction sub-model to obtain the predicted load characteristic value corresponding to the predicted characteristic data;

According to the set of power consumption patterns corresponding to the load forecasting sub-model and the predicted load characteristic value, a candidate power consumption curve corresponding to the load forecasting sub-model is calculated to estimate the power consumption load.

Optionally, in the power load forecasting method, the predicted power consumption mode includes an estimated power consumption mode corresponding to the to-be-predicted time obtained by classifying the predicted feature data by each of the classification decision trees; Predicting the power consumption load according to the predicted power consumption mode and the initial power consumption load curve, and determining the predicted power consumption load curve corresponding to the to-be-predicted time, specifically including:

According to all the estimated power consumption modes, calculate the mode ratio value corresponding to each power consumption mode;

For each power consumption mode, according to the mode ratio value corresponding to the power consumption mode, weighted summation is performed on the predicted load curve of the candidate power consumption curve corresponding to the power consumption mode to obtain the predicted power consumption corresponding to the to-be-predicted time. load curve.

In addition, in order to achieve the above object, the present invention also provides an intelligent terminal, wherein the intelligent terminal includes: a memory, a processor, and a power load prediction program stored in the memory and running on the processor, The electrical load forecasting program, when executed by the processor, implements the steps of the electrical load forecasting method as described above.

In addition, in order to achieve the above object, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a power load prediction program, and the power load prediction program is executed by a processor to achieve the above-mentioned The steps of the power load forecasting method.

The present invention divides the prediction of electricity consumption into mode prediction and load prediction. First, the to-be-predicted characteristic data corresponding to the to-be-predicted time is obtained. The to-be-predicted characteristic data includes meteorological data and time type, and then the corresponding predicted electricity consumption is determined according to the predicted characteristic data. At the same time, the corresponding initial power consumption curve is predicted according to the predicted characteristic data, and then the predicted power consumption load is adjusted according to the predicted power consumption mode, so as to obtain the power consumption load curve corresponding to the time to be predicted. Since the present invention adds the prediction of the power consumption mode on the basis of the conventional load prediction, and adjusts the load prediction result on the basis of the power consumption mode, the prediction result is more accurate.

Description of drawings

Fig. 1 is the flow chart of the preferred embodiment provided by the power load forecasting method of the present invention;

FIG. 2 is a schematic diagram of an operating environment of a preferred embodiment of an intelligent terminal of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer and clearer, the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

The power load prediction method according to the preferred embodiment of the present invention can be executed by an intelligent terminal, and the intelligent terminal includes terminals such as a smart TV and a smart phone. As shown in Figure 1, the power load forecasting method includes the following steps:

Step S100, obtaining prediction feature data corresponding to the time to be predicted, wherein the prediction feature data includes meteorological data and time type.

Specifically, the prediction feature data corresponding to the time to be predicted is obtained first. In this embodiment, the prediction feature data refers to data based on features related to electricity load, such as weather data and time type. Taking meteorological data as an example, when the weather is hot, every household needs fans and air conditioners, so the electricity load will increase compared with the normal temperature. The meteorological data in this embodiment includes temperature, wind, humidity and the like. The time type refers to the time characteristics related to the time to be predicted and the electricity load. For example, there is definitely a difference between the electricity load at night and the electricity load during the day, and there are also certain differences in the electricity load on weekdays, weekends, and holidays.

In addition, the time to be predicted in this embodiment is in units of days, that is, to predict the electricity load on a certain day. However, in practical applications, the time to be predicted can be divided into more detail or roughness according to actual needs. The forecast time is a certain hour of a day, or the electricity load of a certain week.

Step S200 , inputting the predicted feature data into a trained classification model, and classifying the electricity consumption pattern of the to-be-predicted time through the classification model to obtain a predicted electricity-consumption pattern corresponding to the to-be-predicted time.

Specifically, the predicted feature data is input into a trained classification model. The classification model used in this embodiment is a supervised learning classification, and the classification model calculates a probability value that the predicted feature data is a preset power consumption mode, thereby Get the predicted electricity usage pattern. In the first method of obtaining the predicted power consumption mode in this embodiment, the power consumption mode with the largest probability value is directly used as the predicted power consumption mode.

Further, in this embodiment, the classification model includes several classification decision trees obtained by training based on the random forest algorithm; the classification decision tree is obtained by training the decision trees on the basis of the random forest algorithm, and in this embodiment, the classification is obtained. The process of decision tree is:

A10. Acquire historical load data and historical feature data corresponding to each preset historical time.

Specifically, the historical load data corresponding to each preset historical time is obtained. Since the unit of time to be predicted is taken as day in this embodiment, the historical load data obtained during training is also taken as the unit of day, which is set as X={x ₁ ,x ₂ ,…,x _N }, where N represents the total number of days the dataset contains. Further, in this embodiment, the historical load data of each historical time, that is, each day, includes the normalized value of the historical load curve of the day for representation, and the historical load curve is x _i ={x _i1 ,x _i2 ,…,x _im }, m represents the load sampling points of one day, and then divide the load value corresponding to each sampling point in the historical load curve by the maximum value in the historical load curve to achieve normalization of the historical load curve, Get historical load data.

The historical feature data is the meteorological data and time type corresponding to the acquired historical time. The set of historical feature data is expressed as Y={y ₁ , y ₂ ,..., y _N }, where N represents the total number of days included in the data set, and each The historical characteristic data corresponding to a historical time can be expressed as y _i ={y _i1 ,y _i2 ,...,y _in ,y _in+1 ,y _in+2 ,y _in+3 }, the first n values are meteorological characteristic data , the last three values are the encoding results of the time type in the present embodiment, the encoding result in the present embodiment adopts the one-hot encoding method, and the time type is only three time types of working days, weekends, and holidays three time types, But in application, it can also include time types such as Monday to Sunday, morning, afternoon, etc. For example, weekdays are [0,0,1], weekends are [0,1,0], and holidays are [1,0,0].

In this embodiment, the load data and characteristic data of city A are collected for 697 days. The sampling period in the load curve in this embodiment, that is, the interval between sampling points is 1 hour, so each load data includes 24 normalized load values, and these load values are normalized, so that Get load data. The characteristic data includes meteorological data and time type, and the meteorological data includes 5 characteristics of maximum wind speed, minimum wind speed, maximum temperature, minimum temperature and surface average pressure. The dimension of load data X is 24×730, and the dimension of feature data Y is 8×730.

In order to evaluate the accuracy of the prediction method adopted in this embodiment, in this embodiment, the load data of 486 days (486/697 is about 0.7) is randomly selected as the historical load data and its corresponding characteristics in a 7:3 manner. data as historical feature data. The remaining 211 days are used as the time to be predicted, and the corresponding load data and characteristic data are used as test load data and test characteristic data for evaluating the accuracy of the method adopted in this embodiment.

A20. Perform clustering processing on the historical load data to generate a target cluster set, where the target cluster set includes several sets of electricity consumption patterns, and the set of electricity consumption patterns includes historical loads corresponding to the same electricity consumption pattern data.

Specifically, by performing clustering processing on the historical load data, a target cluster set is generated, and the target cluster set includes several power consumption mode sets. In this embodiment, a supervised or semi-supervised clustering algorithm or an unsupervised clustering algorithm may be used for the clustering process. For example K-Means clustering, mean shift clustering algorithm, hierarchical clustering.

Further, the present embodiment adopts the improved fuzzy C-means clustering algorithm as the method for clustering the historical load data in the clustering process, and the specific process is as follows:

B10. Randomly determine the initial cluster center in the historical load data according to the current number of clusters.

Specifically, before performing clustering, the current number of clusters is obtained first, and the number of clusters increases with the increase of the number of clusters. In this embodiment, when clustering is performed for the first time, the number of clusters is: 2, that is to say, the number of initial clusters is c=2. Randomly select c samples from the historical load dataset as initial cluster centers θ _j , where j ∈ [1, c].

B20. Perform clustering on the historical load data according to a preset fuzzy C-means clustering algorithm and the initial clustering center to obtain the kth cluster set corresponding to the current number of clusters.

Specifically, the fuzzy C-means clustering (Fuzzy C-Means, FCM) algorithm is a clustering algorithm that introduces fuzzy power. The samples are classified by calculating the probability that a sample belongs to a certain class as a degree of membership. Clustering is achieved by minimizing the objective function and preset constraints. The process is:

Update the (r-1)th cluster center and the (r-1)th membership matrix to obtain the rth cluster center and the rth initial membership matrix, where the update process is:

Calculate the volume corresponding to the (r-1)th cluster center according to the (r-1)th membership degree matrix, where (r-1) is the number of iterations. When the first iteration is performed, the first cluster The center is the primary clustering center, and the first initial membership matrix is the membership matrix corresponding to the historical load data constructed with random numbers;

According to the preset cluster center update formula, update the (r-1)th cluster center to obtain the rth cluster center, and update the formula according to the preset membership value to update the (r-1)th cluster center Membership degree matrix, get the rth initial membership degree matrix;

When it is determined that the rth initial membership degree matrix complies with the preset constraint condition, the rth initial membership degree matrix is used as the kth intermediate membership degree matrix and output;

When it is determined that the rth initial membership degree matrix does not meet the constraint condition, the rth cluster center and the rth initial membership degree matrix are updated iteratively.

In this embodiment, taking (r-1) as 1 as an example for description, after several initial cluster centers are determined according to the number of clusters c, a first initial membership degree matrix U _N×c is constructed with random numbers ={u _ij },i=1,...,N,j=1,...,c, where u _ij is the membership value of the i-th historical load data to the j-th cluster center, random The number is a random number in the [0,1] interval, and N is the data volume of the historical load data corresponding to all historical times. According to the membership value between each historical load data in the membership matrix and the initial cluster center, the corresponding volume of each first cluster center can be calculated. The calculation formula of the volume is:

q为预设的模糊度值。同时，根据体量可评估根据第一初始隶属度矩阵中的隶属度值的可靠性，因此可根据计算得到的体量，对第一隶属度值进行更新，得到第二隶属度值，隶属度值更新公式为

其中DTW(x_i,θ_j)是动态时间规整距离。Since the initial cluster center is randomly selected, and the first membership value is obtained by random assignment, which is far from the real situation, it is necessary to update the initial cluster center to obtain the second cluster center, among which, the cluster center The update formula is

q is the preset blur value. At the same time, the reliability of the membership degree value in the first initial membership degree matrix can be evaluated according to the volume, so the first membership degree value can be updated according to the calculated volume to obtain the second membership degree value, the membership degree The value update formula is

where DTW( _xi , θ _j ) is the dynamic time warping distance.

After obtaining the second initial membership degree matrix, it is judged whether the second initial membership degree matrix satisfies the preset constraint rule. If it is satisfied, the second initial membership degree matrix is taken as the kth intermediate membership degree matrix and output; if not, the membership degree value and the cluster center will continue to be updated until the rth initial membership degree matrix satisfies the constraint rule, and Take the rth initial membership matrix as the kth intermediate membership matrix and output it.

或者r≥R，其中，ε为预设的迭代误差阈值，R为迭代次数阈值。Further, the constraints adopted in this embodiment are

Or r≥R, where ε is the preset iteration error threshold, and R is the iteration number threshold.

B30. When the number of clusters is less than a preset threshold for the number of clusters, add one to the number of clusters, and repeatedly determine the kth intermediate membership matrix.

Specifically, a threshold K of the number of clusters is preset, and when the number of clusters is less than the preset threshold of the number of clusters, the number of clusters is increased by one, and the current number of clusters is 2 as an example, the number of clusters after adding one is 3, and then re-confirm the initial cluster center in the historical load data with the number of clusters as 3, and calculate the second intermediate membership corresponding to when the number of clusters is 3 degree matrix.

B40. When the number of clusters is greater than or equal to the threshold of the number of clusters, determine target membership degree matrices from the first intermediate membership degree matrix to the (K-1)th intermediate membership degree matrix.

Specifically, when the number of clusters is greater than or equal to the threshold of the number of clusters, then iteratively increases all kth intermediate membership degree matrices in the process from the number of clusters being 2 to the number of clusters being K Confirm the target membership matrix in . Since each time the number of clusters is increased, an intermediate membership matrix will be output, and each intermediate membership matrix must be different, so the most accurate intermediate membership matrix is selected among them, that is, the target membership matrix. In this embodiment, K is 10, so the first to ninth membership degree matrices are obtained.

In this embodiment, the method of determining the target membership degree matrix is:

For each of the intermediate membership degree matrices from the first intermediate membership degree matrix to the (K-1)th intermediate membership degree matrix, calculate the clustering validity index corresponding to the intermediate membership degree matrix;

According to the clustering effectiveness index, a set of target clusters in the first intermediate membership degree matrix to the (K-1)th intermediate membership degree matrix is determined.

其中，c为聚类个数，

为该中间隶属度矩阵的对应的聚类中心的动态时间规整距离的最小值。然后选择所有的CVI中最小值所对应的中间隶属度矩阵作为目标隶属度矩阵，将目标隶属度矩阵对应的聚类个数记为c*。本实施例中的c*＝5，即经过历史负荷数据训练得到的用电模式有五种。由于聚类中心θ_j是某一个天对应的历史负荷数据，本实施例中的历史负荷数据为24小时内每一个小时采点并归一化的负荷值，因此，聚类中心θ_j可标识为由24个数值组合的矢量。下表表1为基于历史负荷数据确定的5种用电模式的聚类中心，其中，每一列表示该用电模式对应的聚类中心。表2为历史负荷数据对应不同的用电模式的聚类中心的隶属度值。Specifically, for each cluster set, calculate its corresponding cluster validity index (Cluster Validity Index, CVI). In this embodiment, the calculation formula of the cluster validity index is:

where c is the number of clusters,

is the minimum value of the dynamic time warping distance of the corresponding cluster centers of the intermediate membership matrix. Then, the intermediate membership matrix corresponding to the minimum value in all CVIs is selected as the target membership matrix, and the number of clusters corresponding to the target membership matrix is denoted as c*. In this embodiment, c*=5, that is, there are five power consumption modes obtained through the training of historical load data. Since the cluster center θ _j is the historical load data corresponding to a certain day, the historical load data in this embodiment is the load value of points collected and normalized in every hour within 24 hours. Therefore, the cluster center θ _j can be identified as is a vector composed of 24 values. Table 1 below shows the cluster centers of the five electricity consumption patterns determined based on historical load data, wherein each column represents the cluster center corresponding to the electricity consumption pattern. Table 2 shows the membership values of the cluster centers of the historical load data corresponding to different electricity consumption patterns.

Table 1

Table 2

B50. Perform clustering on the historical load data according to the target membership degree matrix and the cluster center corresponding to the target membership degree matrix to obtain a target cluster set.

Specifically, according to the target membership degree matrix and the cluster center corresponding to the target membership degree matrix, the historical load data is defuzzified, and for each historical load data, the membership value corresponding to the historical load data is used. The cluster center corresponding to the maximum value in is the center point, so that each historical load data is clustered to obtain the target cluster set. For example, the membership values of historical load data to the cluster centers A, B, C, D and E corresponding to the target membership matrix are 0.1, 0.2, 0.1, 0.5 and 0.1, respectively, so the historical load data A is classified as a cluster The category in which center D is located. At the end of clustering, the category corresponding to each cluster center is regarded as an electricity consumption pattern, thereby obtaining c* electricity consumption pattern sets.

A30. For each power consumption mode set, mark the historical feature data according to the historical time corresponding to each historical load data in the power consumption mode set, to obtain the power consumption mode corresponding to each historical feature data.

Specifically, for each power consumption mode set, such as power consumption mode set A, mark the historical time corresponding to the historical load data in the power consumption mode set A, for example, on March 3 and March 5, mark In order for the electricity consumption mode to be electricity consumption mode A, each historical characteristic data is marked to obtain the electricity consumption mode corresponding to each of the historical characteristic data, that is, the historical characteristic data corresponding to March 3rd and March 5th are obtained. It is marked as power consumption mode A.

A40. For each preset decision tree, select the training feature data in the historical feature data to input the decision tree and split the decision tree according to the Gini index, until the historical feature data of each node in the decision tree corresponds to The power consumption patterns are the same, and the classification decision tree is obtained.

Specifically, this embodiment uses the random forest package of Python to build a random forest model, and sets the parameter: n_trees=200, that is, a preset number of 200 decision trees, wherein the number of decision trees can be adjusted according to actual needs. Then, multiple training feature data are randomly and replaced from all historical feature data to obtain the same number of training feature data sets as decision trees. Each training feature dataset is used to train a decision tree. In this embodiment, each training feature data set includes 400 training feature data.

For each decision tree, start splitting down from the root node, and select the best feature for each split according to the Gini index, until the training feature data of all nodes come from the same class, that is, the training feature data of the node. Corresponding to the same power consumption mode, stop splitting and get a classification decision tree.

Step S300, inputting the prediction feature data into a trained load prediction model, and using the load prediction model to predict the electricity load for the to-be-predicted time to obtain an initial electricity-consumption load curve corresponding to the to-be-predicted time.

Specifically, the acquired predictive feature data is input into a trained load prediction model, and the load prediction model is used to calculate an initial electricity load curve corresponding to the predictive feature data according to the input predictive feature data. The initial model before the training of the load prediction model is preferably a neural network model.

In this embodiment, the load prediction model includes several load prediction sub-models, and the load prediction sub-models are in one-to-one correspondence with the power consumption patterns. That is to say, each power consumption mode corresponds to a load prediction sub-model. Thereby, it is avoided that the accuracy of prediction is reduced due to the difference of electricity consumption patterns. In an implementation manner of obtaining the predicted power consumption load in this embodiment, after the predicted power consumption mode is obtained, the predicted characteristic data is input into the load prediction model corresponding to the predicted power consumption mode, so as to obtain the corresponding initial power consumption load curve.

In this embodiment, the training process of the load prediction model is:

For each preset initial model, input the training curve data into the initial model and classify the training curve data through the initial model to obtain the training load characteristic value corresponding to each of the historical characteristic data, wherein the The training curve data is the data corresponding to the initial model in the historical feature data;

According to the training load characteristic value and the historical load characteristic value corresponding to the historical characteristic data, parameter optimization of the initial model is performed until the initial model converges, and a load prediction sub-model is obtained.

Specifically, in this embodiment, the number of neural network models equal to the number of power consumption modes is preset, that is, five, each neural network model includes 5 layers, the first layer is the input layer, and the number of neurons is equal to 8; The layer is the hidden layer, each layer contains 10 neurons; the last layer is the output layer, which consists of two neurons. Among them, the number of neurons in the input layer is determined by the number of types of predicted feature data, that is, m+3. In addition, the training-related parameters learning rate γ, batch size batch_size, and training target minimum error ε are also preset. In this embodiment, the learning rate γ=0.001, the batch size batch_size=64, and the minimum error of the training target ε=1e-5;

For each preset neural network model, the training curve data corresponding to the initial model in the historical feature data is input into the initial model. For the power consumption mode, determine the corresponding training curve data. For example, when the power consumption mode A is used to train the corresponding initial model, only the historical feature data of the application power mode A is used as the training curve data corresponding to the initial model. In the second implementation manner of this embodiment, the training curve data corresponding to the initial model is determined according to the membership value corresponding to each historical feature data. For example, a certain historical feature data A has a membership value of 0.99 for the cluster center of the applied electrical model A, while the membership value corresponding to the historical feature data B is 0.1, then the historical feature data A has a high probability to be selected as the initial model The corresponding training curve data, while the historical feature data B is selected as the training curve data corresponding to the initial model with a lower probability.

The initial model classifies the input training curve data to obtain its corresponding training load curve. Since the historical load data includes a historical load curve corresponding to each of the historical times, it is difficult to predict the curve independently. Therefore, in this embodiment, the historical load characteristic value is used as a value for describing the training load curve. The historical load characteristic value refers to a value that can describe the characteristics of the historical load curve, such as peak value, valley value, average value, and the like. This embodiment is described by taking both the peak value and the valley value as the historical load characteristic value as an example. The training load characteristic value refers to the load characteristic value obtained after classifying the input training characteristic data.

次，N为历史负荷数据的数量。本实施例中的收敛规则为损失值小于训练目标最小误差ε＝1e-5。Then, based on the preset loss function, according to the loss value between the training load curve and the historical load data corresponding to the actually input training curve data, the loss function in this embodiment is to calculate the root mean square error value of the peak value and the valley value. formula. Then according to the loss value, the stochastic gradient descent method is used to optimize the parameters of the neural network model until the neural network model converges, and the load forecasting sub-model is obtained. In each iteration, the above process is repeated

times, N is the number of historical load data. The convergence rule in this embodiment is that the loss value is smaller than the training target minimum error ε=1e-5.

Further, in the second implementation manner of obtaining the predicted power consumption load in this embodiment, the predicted power consumption load includes the estimated power consumption load corresponding to each load prediction sub-model, and the process of obtaining the estimated power consumption load is as follows: :

For each of the load prediction sub-models, input the predicted characteristic data into the load prediction sub-model to obtain the predicted load characteristic value corresponding to the predicted characteristic data;

According to the power consumption pattern set corresponding to the load forecasting sub-model and the predicted load characteristic value, a candidate power consumption load curve corresponding to the load forecasting sub-model is calculated.

Specifically, for each of the load forecasting sub-models, such as the load forecasting sub-model corresponding to the electricity consumption mode A, the predicted characteristic data is input into the load forecasting sub-model, and the forecasted load sub-model performs electricity load forecasting on the predicted characteristic data. to obtain the predicted load characteristic value corresponding to the predicted characteristic data. In this embodiment, it is the predicted peak value and the predicted valley value corresponding to the time to be predicted.

Since the historical electricity consumption trends in the electricity consumption pattern set corresponding to the load forecasting sub-model have the same electricity consumption trend and electricity consumption trend, according to the electricity consumption pattern set and the obtained predicted peak value and predicted valley value, The candidate power consumption load curve corresponding to the load prediction sub-model can be calculated. A candidate power consumption load curve corresponding to each of the load prediction sub-models is obtained, thereby obtaining an initial power consumption load curve.

其中，j＝1,2,...,c*，peak_j表示第j个负荷预测子模型所得到的预测峰值，valley_j表示第j个负荷预测子模型所得到的预测谷值。Further, in this embodiment, the formula for calculating the candidate electricity load curve corresponding to the load prediction sub-model is:

Among them, j=1,2,...,c*, peak _j represents the predicted peak value obtained by the jth load forecasting sub-model, and valley _j represents the predicted valley value obtained by the jth load forecasting sub-model.

Step S400, predicting the power consumption load according to the predicted power consumption mode and the initial power consumption load curve, and determining the predicted power consumption load curve corresponding to the to-be-predicted time.

Specifically, in this embodiment, after the predicted power consumption pattern and the initial power consumption load curve are obtained, the initial power consumption load curve is adjusted according to the predicted power consumption mode, and the overall trend and trend of the power load curve is used to predict the power consumption. The historical load data in the model is closer, so that the predicted electricity load curve corresponding to the time to be predicted can be obtained.

其中

n_j为第j个用电模式对应的模式占比值。在本实施例中，总共得到200个分类决策树的预估用电模式，5种用电模式对应的模式占比值分别为：0.3195,0.0589,0.4729,0.0076和0.1411。Further, in this embodiment, the predicted power consumption mode includes an estimated power consumption mode corresponding to the to-be-predicted time obtained by classifying the predicted feature data by each of the classification decision trees. First, according to all the estimated power consumption modes, the mode ratio value corresponding to each power consumption mode is calculated. For example, the predicted power consumption mode L corresponding to the time to be predicted is L={l ₁ , l ₂ , . . . , l _{n_trees} }, where l _{n_trees} is the predicted power consumption mode corresponding to the nth classification decision tree. Statistically predicting the mode ratio of each power consumption mode in the power consumption mode

in

n _j is the mode ratio value corresponding to the jth power consumption mode. In this embodiment, a total of 200 estimated power consumption modes of the classification decision tree are obtained, and the mode ratio values corresponding to the five power consumption modes are: 0.3195, 0.0589, 0.4729, 0.0076, and 0.1411, respectively.

For each of the power consumption modes, according to the mode ratio value corresponding to the power consumption mode, the candidate power consumption load curve corresponding to the power consumption mode is weighted and summed to obtain the candidate power consumption load curve corresponding to the to-be-predicted time, The calculation formula is

This embodiment takes 211 days to be forecasted as an example, and calculates the mean absolute percentage error (Mean Absolute Percentage Error, MAPE) between the predicted power consumption curve and the real power consumption curve. The 211 days to be forecasted The mean value of the corresponding MAPE is 3.1%.

Further, as shown in FIG. 2 , based on the above-mentioned power load prediction method, the present invention also provides an intelligent terminal correspondingly, and the intelligent terminal includes a processor 10 , a memory 20 and a display 30 . FIG. 2 only shows some components of the smart terminal, but it should be understood that it is not required to implement all the shown components, and more or less components may be implemented instead.

In some embodiments, the memory 20 may be an internal storage unit of the smart terminal, such as a hard disk or a memory of the smart terminal. In other embodiments, the memory 20 may also be an external storage device of the smart terminal, for example, a plug-in hard disk equipped on the smart terminal, a smart memory card (Smart Media Card, SMC), a secure digital (Secure) Digital, SD) card, flash memory card (Flash Card), etc. Further, the memory 20 may also include both an internal storage unit of the smart terminal and an external storage device. The memory 20 is used to store application software and various types of data installed in the smart terminal, such as program codes for installing the smart terminal. The memory 20 can also be used to temporarily store data that has been output or is to be output. In one embodiment, the memory 20 stores a power load forecasting program 40, and the power load forecasting program 40 can be executed by the processor 10, thereby implementing the power load forecasting method in the present application.

In some embodiments, the processor 10 may be a central processing unit (Central Processing Unit, CPU), a microprocessor or other data processing chips, which are used to execute program codes or process data stored in the memory 20, such as The power load prediction method and the like are executed.

In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode, organic light-emitting diode) touch device, and the like. The display 30 is used for displaying information on the smart terminal and for displaying a visual user interface. The components 10-30 of the intelligent terminal communicate with each other through the system bus.

In one embodiment, when the processor 10 executes the power load prediction program 40 in the memory 20, the following steps are implemented:

Obtaining prediction feature data corresponding to the time to be predicted, wherein the prediction feature data includes meteorological data and time type;

Inputting the predicted feature data into a trained classification model and classifying the electricity consumption pattern of the time to be predicted by the classification model, to obtain a predicted electricity consumption pattern corresponding to the time to be predicted;

Inputting the prediction feature data into a trained load prediction model and predicting the electricity load for the to-be-predicted time by using the load prediction model, to obtain an initial electricity-use load curve corresponding to the to-be-predicted time to predict the electricity load;

The power consumption load is predicted according to the predicted power consumption mode and the initial power consumption load curve, and the predicted power consumption load curve corresponding to the to-be-predicted time is determined.

The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a power load forecasting program, and when the power load forecasting program is executed by a processor, implements the steps of the above-mentioned power load forecasting method .

Of course, those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing relevant hardware (such as processors, controllers, etc.) through a computer program, and the programs can be stored in a In a computer-readable computer-readable storage medium, the program, when executed, may include the processes of the foregoing method embodiments. The computer-readable storage medium may be a memory, a magnetic disk, an optical disk, or the like.

It should be understood that the application of the present invention is not limited to the above examples. For those of ordinary skill in the art, improvements or transformations can be made according to the above descriptions, and all these improvements and transformations should belong to the protection scope of the appended claims of the present invention.

Claims (7)

1. A power load forecasting method, wherein the power load forecasting method comprises: Obtaining prediction feature data corresponding to the time to be predicted, wherein the prediction feature data includes meteorological data and time type; Inputting the predicted feature data into a trained classification model and classifying the electricity consumption pattern of the time to be predicted by the classification model, to obtain a predicted electricity consumption pattern corresponding to the time to be predicted; Inputting the prediction feature data into a trained load prediction model and predicting the electricity load for the to-be-predicted time through the load prediction model, to obtain an initial electricity-consumption load curve corresponding to the to-be-predicted time; determining a predicted power consumption curve corresponding to the to-be-predicted time according to the predicted power consumption pattern and the initial power consumption load curve; The classification model includes several classification decision trees trained based on the random forest algorithm; the training process of the classification decision tree specifically includes: Obtain historical load data and historical feature data corresponding to each preset historical time; Performing clustering processing on the historical load data to generate a target cluster set, wherein the target cluster set includes several sets of electricity consumption patterns, and the set of electricity consumption patterns includes historical load data corresponding to the same electricity consumption mode; For each of the set of power consumption patterns, according to the historical time corresponding to each historical load data in the set of power consumption patterns, the historical feature data is marked to obtain the power consumption pattern corresponding to each of the historical feature data; For each preset decision tree, select the training feature data in the historical feature data to input into the decision tree and split the decision tree according to the Gini index, until the power consumption corresponding to the historical feature data of each node in the decision tree The patterns are the same, and the classification decision tree is obtained; The performing clustering processing on the historical load data to generate a target clustering set specifically includes: According to the current number of clusters, randomly determine the initial cluster center in the historical load data, wherein, when clustering is performed for the first time, the number of clusters is 2; According to the preset fuzzy C-means clustering algorithm and the initial clustering center, calculate the k-th intermediate membership degree matrix corresponding to the historical load data, wherein k is the number of times of clustering; Update the (r-1)th cluster center and the (r-1)th membership matrix to obtain the rth cluster center and the rth initial membership matrix, where the update process is: Calculate the volume corresponding to the (r-1)th cluster center according to the (r-1)th membership degree matrix, where (r-1) is the number of iterations. When the first iteration is performed, the first cluster The center is the primary clustering center, and the first initial membership matrix is the membership matrix corresponding to the historical load data constructed with random numbers; According to the preset cluster center update formula, update the (r-1)th cluster center to obtain the rth cluster center, and update the formula according to the preset membership value to update the (r-1)th cluster center Membership degree matrix, get the rth initial membership degree matrix; When it is determined that the rth initial membership degree matrix complies with the preset constraint conditions, the rth initial membership degree matrix is used as the kth intermediate membership degree matrix and output; When it is determined that the rth initial membership degree matrix does not meet the constraint condition, iteratively updates the rth cluster center and the rth initial membership degree matrix; Determine several initial cluster centers according to the number of clusters c, and construct a first initial membership degree matrix U _N×c ={u _ij }, i=1,...,N; j=1,... ., c, where u _ij is the membership value of the i-th historical load data to the j-th cluster center, the random number is a random number in the [0, 1] interval, and N is the value of the historical load data corresponding to all historical times. The amount of data; According to the membership value between each historical load data in the membership matrix and the initial cluster center, the volume corresponding to each first cluster center can be calculated, and the calculation formula of the volume is:

Update the initial cluster center to obtain the second cluster center, where the cluster center update formula is:

Wherein, q is the preset ambiguity value, and x _i ={x _i1 , x _i2 , . . . , x _im } is the historical load curve; According to the volume, the first membership value is updated to obtain the second membership value. The membership value update formula is:

where DTW(x _i , θ _j ) is the dynamic time warping distance, f _k is the volume corresponding to the kth cluster center, and θ _z is the cluster center update formula when j is Z; When the number of clusters is less than the preset number of clusters threshold, the number of clusters is increased by one, and the kth intermediate membership degree matrix is repeatedly determined; When the number of clusters is greater than or equal to the threshold of the number of clusters, determine the target membership degree matrix from the first intermediate membership degree matrix to the (K-1)th intermediate membership degree matrix, where K is the clustering degree matrix. class number threshold; According to the target membership degree matrix and the cluster center corresponding to the target membership degree matrix, the historical load data is clustered to obtain a target cluster set; The initial power consumption load curve includes a candidate power consumption load curve corresponding to each load prediction sub-model; the predicted characteristic data is input into the trained load prediction model, and the to-be-predicted time is determined by the load prediction model. Carry out electricity load prediction, and obtain the initial electricity load curve corresponding to the to-be-predicted time, which specifically includes: For each of the load prediction sub-models, input the predicted feature data into the load prediction sub-model to obtain the predicted load feature value corresponding to the predicted feature data; According to the power consumption pattern set corresponding to the load forecasting sub-model and the predicted load characteristic value, calculate the candidate power consumption load curve corresponding to the load forecasting sub-model; The predicted load characteristic value corresponding to the predicted characteristic data is the predicted peak value and the predicted valley value corresponding to the time to be predicted; The formula for calculating the candidate electricity load curve corresponding to the load forecasting sub-model is:

Among them, _j ⁼ 1, ₂ , . 2 . The power load forecasting method according to claim 1 , wherein when the number of clusters is greater than or equal to the threshold of the number of clusters, the first intermediate membership degree matrix to (K-th) is determined. 3 . 1) The target membership degree matrix in the intermediate membership degree matrix, which specifically includes: For each of the intermediate membership degree matrices from the first intermediate membership degree matrix to the (K-1)th intermediate membership degree matrix, calculate the clustering validity index corresponding to the intermediate membership degree matrix; According to the clustering effectiveness index, a set of target clusters in the first intermediate membership degree matrix to the (K-1)th intermediate membership degree matrix is determined. 3 . The power load forecasting method according to claim 1 , wherein the load forecasting model comprises several load forecasting sub-models, and the load forecasting sub-models correspond to the electricity consumption patterns one-to-one. 4 . 4 . The power load forecasting method according to claim 3 , wherein the historical load data comprises a historical load curve corresponding to each historical time, and the historical load curve comprises historical load characteristic values; the load The training process of the prediction sub-model specifically includes: For each preset initial model, input the training curve data into the initial model and classify the training curve data through the initial model to obtain the training load characteristic value corresponding to each of the historical characteristic data, wherein the The training curve data is the data corresponding to the initial model in the historical feature data; According to the training load characteristic value and the historical load characteristic value corresponding to the historical characteristic data, parameter optimization of the initial model is performed until the initial model converges, and a loss load prediction sub-model is obtained. 5 . The power load forecasting method according to claim 1 , wherein the predicted power consumption pattern comprises an estimated function corresponding to the to-be-predicted time obtained by classifying the predicted feature data by each of the classification decision trees. 6 . The electricity model; the predicting the electricity load according to the predicted electricity consumption mode and the initial electricity consumption load curve, and determining the predicted electricity consumption load curve corresponding to the to-be-predicted time, specifically includes: According to all the estimated power consumption modes, calculate the mode ratio value corresponding to each power consumption mode; For each of the power consumption modes, according to the mode ratio value corresponding to the power consumption mode, weighted summation is performed on the candidate power consumption load curves corresponding to the power consumption mode to obtain a predicted power consumption load curve corresponding to the to-be-predicted time. 6. An intelligent terminal, characterized in that the intelligent terminal comprises: a memory, a processor, and a power load forecasting program stored on the memory and running on the processor, the power load forecasting program being The processor implements the steps of the power load prediction method according to any one of claims 1-5 when executed. 7. A computer-readable storage medium, wherein the computer-readable storage medium stores a power load forecasting program, and when the power load forecasting program is executed by a processor, the power load forecasting program according to any one of claims 1-5 is implemented. The steps of the power load forecasting method described above.

Images