CN115689001A - Short-term Load Forecasting Method Based on Pattern Matching - Google Patents

Abstract

The invention relates to a short-term load forecasting method based on pattern matching, belonging to the technical field of load forecasting of an electric power system. And selecting a prediction model with the best verification effect from the multiple prediction models aiming at each type of load mode so as to improve the prediction precision of each type of load mode. The method realizes the self-adaptive matching of the user load and the load mode; the matching of the load mode and the optimal prediction model is realized, the advantages of each prediction model are fully exerted, and the overall prediction precision can be effectively improved.

Description

Short-term Load Forecasting Method Based on Pattern Matching

technical field

The invention belongs to the technical field of power system load forecasting, and in particular relates to a short-term load forecasting method based on pattern matching.

Background technique

Load forecasting is one of the key links in formulating power dispatching plan. Accurate load forecasting is helpful for the system to reasonably arrange the start and stop of units, improve the utilization rate of new energy, automatic power generation control and safety maintenance, and play an important role in the safety and stability of the system and economic dispatch. With the continuous deepening of the construction of new power systems in my country and the establishment of advanced measurement systems on the user side, a large amount of electricity consumption data on the user side has been accumulated in this process. Due to large volatility and randomness, accurate prediction of user cluster load is an important problem to be solved at present.

In recent years, many scholars at home and abroad have done a lot of research on short-term load forecasting, and proposed many forecasting models and methods, which can be roughly divided into two categories, one is the traditional forecasting method represented by the time series method, and the other is the Based on a data-driven approach to artificial intelligence. Auto regression (auto regression, AR) and auto regression moving average (autoregression moving average, ARMA) mathematical models are commonly used models in time series methods. Learning ability, high requirements for data stability. In order to solve the problems of low prediction accuracy and low adaptability of traditional forecasting methods, more and more data-driven artificial intelligence methods are involved in load forecasting, such as artificial neural network (ANN), which supports Vector machines (support vector machines, SVM), random forests, deep belief networks, and more. At present, the deep belief network is used for substation load forecasting, and the optimal parameters of the network are obtained by using the adaptive matrix estimation algorithm to realize the adaptability of the network, but the calculation is relatively complicated. There are also K-means clustering on the cells first to construct an SVM prediction model to realize space load prediction, but the SVM prediction model is greatly affected by the correlation coefficient of the kernel function and the penalty coefficient c. Another proposed a short-term power load forecasting method based on Bayesian optimization based on convolutional neural network (CNN) bidirectional gate recurrent unit (BiGRU). However, neural network training is prone to "overfitting" problems, which affects the generalization ability of its model and reduces the prediction accuracy. Others use the C-means fuzzy clustering algorithm to cluster historical samples, and then build a random forest regression model for similar data to predict. However, this method needs to create many decision trees, and the space and time required for training will be large.

However, with the popularization of new energy power generation systems on the user side, the randomness and volatility of user cluster loads increase, and the application of the above single prediction method may lead to poor generalization performance due to randomness.

Contents of the invention

The purpose of the present invention is to provide a short-term load forecasting method based on pattern matching, which is used to solve the technical problems in the above-mentioned prior art, realize adaptive matching of user load and load pattern; also realize the matching of load pattern and optimal forecasting model, Giving full play to the advantages of each forecasting model can effectively improve the overall forecasting accuracy.

For realizing the above object, technical scheme of the present invention is:

A short-term load forecasting method based on pattern matching, including the following steps:

S1. Using the hierarchical K-means clustering algorithm based on the Pearson correlation coefficient, the user's historical load data is optimally divided into various monthly load patterns;

S2. For each monthly load pattern, select the pattern matching algorithm with the best verification effect from feedback neural network, support vector machine regression and linear regression;

S3. According to the similarity between the user's latest monthly load data and the monthly load pattern, the user's electricity consumption data is adaptively matched with the monthly load pattern, and the prediction results of each pattern are added to obtain the total prediction result.

Further, step S1 is specifically as follows:

Clustering similarity measure function based on Pearson correlation coefficient:

Suppose x _i ＝(x _i1 , x _i2 ,…,x _in ), x _j ＝(x _j1 ,x _j2 ,…,x _jn ) are two load curves, and their Pearson correlation coefficients are:

和

分别为x_i和x_j的均值；In the formula:

and

are the mean values of x _i and x _j respectively;

Hierarchical K-means clustering method principle:

Repeat K-means clustering with the same K value multiple times, and record each cluster center;

Perform hierarchical clustering on all recorded cluster centers;

The cluster center obtained by hierarchical clustering is used as the initial cluster center, and then K-means clustering is performed again to obtain the final clustering result;

Selection of clustering indicators:

The evaluation indicators are:

In the formula: K is the number of clusters, v _c is the cluster center of each cluster; n _K is the number of samples contained in the current K-th cluster.

Further, step S2 is specifically as follows:

BP neural network algorithm mechanism:

Among them: P is the input vector, b is the bias constant, w is the weight vector from the input signal to the neuron, f is the activation function, and y is the output signal of the neuron; the output of the neuron is

y=f(wP+b) (14)

The BP neural network consists of three parts: the input layer, the hidden layer and the output layer, and the neurons between each layer are fully connected;

The BP neural network algorithm is composed of two parts: one is the forward propagation of the signal, the other is to calculate the error between the output value and the real value, and propagate the error back, and continuously correct the parameters of each neuron through the L-M algorithm; Carry out training until the training result reaches the error requirement or the maximum number of training times;

Support vector machine regression algorithm mechanism:

The input vector is mapped to the high-dimensional feature space through nonlinear mapping, and a regression hyperplane is obtained by calculation, so that the distance sum of all sample points in the set to the hyperplane is the smallest; nonlinear mapping is also called a kernel function; set the input The sample of is M={( _xi ,y _i ),i=1,2,…n}, x _i ∈ R ^d , y _i ∈ R, the hyperplane function expression obtained by SVR is:

f(x)=ωΦ(x)+b (15)

In the formula: Φ( ) is the kernel function, b is the threshold, and ω is the weight vector;

The SVR loss function is:

In the formula: ε is the distance from the tolerance sample point to the hyperplane, that is, when the distance between the sample point and the hyperplane is less than or equal to ε, the loss is 0;

Linear regression algorithm mechanism:

Linear regression is an analytical method that uses a regression equation to model the relationship between one or more independent variables x and dependent variables y; its model is:

f(x)＝w ₀ +w ₁ x ₁ +w ₂ x ₂ +...+w _n x _n (17)

y=f(x)+δ (18)

In the formula: w is the weight coefficient, δ is the residual error; the main purpose of linear regression is to obtain the coefficient w so that for all sample points in the set M={( _xi ,y _i ),i=1,2,…n } The distance from f(x) is the closest; linear regression uses the least square method to find its equation;

ω ₁ ,...,ω _n ＝(X ^T Y) ^-1 X ^T Y (20)

Using linear regression to do load forecasting requires a small amount of calculation, and the relationship between input features and forecast information can be directly obtained.

Further, step S3 is specifically as follows:

Input feature parameters:

To predict the load P _t at time t, it is necessary to input characteristic parameters to the model that are highly correlated with the load at time t; the input characteristic parameters are:

x＝{P _th ,P _th-1 ,P _th-2 ,P _t-h+1 ,P _t-h+2 ,P _t-2h ,P _t-2h-1 ,P _t-2h+1 ,

P _t-3h ,P _t-7h ,weekday,hour}

In the formula, the first seven are historical load data, h is the number of hours in a day when the load data is collected, weekday and hour represent the week and hour respectively;

Electric load pattern matching based on Pearson correlation coefficient:

Set a month to be calculated as 28. When predicting the load on day q+1 (q=0, 1, ... 27) of a certain month, it is necessary to select the load data of the 28 days before the forecast date for pattern matching, which includes the q The data data1 of the day, and the load data data2 of 28-q after the previous month; in order to make the pattern matching time consistent, the data of data1 is moved before data2, and spliced into a user's completed monthly load data data3 from the beginning of the month to the end of the month; Calculate the Pearson correlation coefficient between each user data3 and each monthly load pattern, and classify each user into the most similar monthly load pattern with its lowest value;

The overall process of load forecasting:

(1) Preprocessing the user load data;

(2) Taking the monthly load of all users as a sample, perform multiple K-means clustering based on Pearson's coefficient on the sample, calculate the clustering index V, and obtain the best clustering number K _b ;

(3) Select the number of clusters as K _b , perform hierarchical K-means clustering based on the Pearson correlation coefficient, and obtain the monthly load of the K _b category, and use the cluster center of each category of monthly load as the monthly load pattern;

(4) Build BPNN, SVR and LR models for each monthly load pattern; select the model with the best test effect to match the monthly load pattern;

(5) Match the load of the month before each user's forecast with the monthly load that is most similar to it;

(6) Sum the monthly loads of users matched to the same pattern, and calculate the matching degree with the monthly load pattern, that is, the Pearson correlation coefficient;

(7) Carry out load forecasting with the best forecasting model of the monthly load mode, and add up the forecasting results of each mode to obtain the final load forecasting result;

则使用BPNN预测模型，否则使用该月负荷模式的次优预测模型；(8) In (7), if the best prediction model of the monthly load pattern is BPNN, and the matching degree between the obtained user load data and the monthly load pattern is less than a certain threshold

Then use the BPNN forecasting model, otherwise use the suboptimal forecasting model of the monthly load pattern;

(9) Use MAPE and RMSE to evaluate the user cluster prediction effect;

In the formula: e is the actual value, o is the predicted value, and N is the predicted number.

Compared with prior art, the beneficial effect that the present invention has is:

One of the beneficial effects of this solution is that, firstly, fully excavate the morphological characteristics of the user's historical load data, cluster the load pattern with strong fluctuation regularity, and combine the user load data with the load pattern according to the similarity between the latest user load data and the load pattern. Pattern matching to automatically classify data with more consistent fluctuations into one category. For each type of load mode, the prediction model with the best verification effect is selected from a variety of prediction models to improve the prediction accuracy of each type of load mode. The above method realizes the adaptive matching of user load and load pattern; also realizes the matching of load pattern and optimal forecasting model, gives full play to the advantages of each forecasting model, and can effectively improve the overall forecasting accuracy. Carry out morphological clustering based on the Pearson correlation coefficient for the user's electricity load, and adaptively match the user's electricity load pattern. Take the best forecasting model for each load pattern to get the best forecasting results. The calculation example shows that the algorithm based on pattern adaptive matching can effectively improve the prediction accuracy.

Description of drawings

Fig. 1 is a comparison diagram of the shape similarity of three load curves in a specific embodiment of the present invention.

Fig. 2 is a neuron model diagram of a specific embodiment of the present invention.

FIG. 3 is a structural diagram of a three-layer BP network in a specific embodiment of the present invention.

Fig. 4 is a schematic diagram of pattern matching in a specific embodiment of the present invention.

Fig. 5 is a prediction flow chart of a specific embodiment of the present invention.

Fig. 6 is a clustering index diagram of a specific embodiment of the present invention.

Fig. 7 is a monthly load pattern diagram of a specific embodiment of the present invention.

FIG. 8 is a diagram of user cluster prediction results in a specific embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with accompanying drawings 1 to 8 of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Example:

A short-term load forecasting method based on pattern matching is provided. Based on the load data of 300 households with photovoltaic power generation systems, the Pearson correlation coefficient is used to measure the morphological distance of the user's monthly load, and the hierarchical K-means clustering method is used to optimally divide different monthly load patterns. For each monthly load pattern, the feedback neural network (backpropagation neural network, BPNN), support vector regression (support vector regression, SVR) and linear regression (linear regression, LR) model training and verification are carried out. When carrying out load forecasting, all users are matched to the load pattern with the closest morphological distance according to the user's electricity consumption data in the month before the forecast date, and the forecasting model with the best test effect above is used for forecasting. Finally, using mean absolute percentage error (MAPE) and root mean square error (RMSE) as error measurements, the effectiveness of the proposed user cluster prediction method is verified by comparing other prediction methods.

1. Division and matching of monthly load patterns based on hierarchical K-means clustering algorithm

Hierarchical K-means clustering:

The main goal of cluster analysis is to divide the samples with high similarity into one class and the samples with low similarity into different classes by comparing the similarity of all samples. The commonly used sample similarity measurement method is to normalize the samples and calculate the Euclidean distance between the normalized samples, but the normalization will cause the compression and loss of data information, and the measurement method based on the Euclidean distance will be due to the sum of squared distances Calculate the overall difference between two samples, but ignore the difference in sample morphology, resulting in unsatisfactory clustering effect. As a classic clustering algorithm, the K-means algorithm is widely used in the cluster analysis of electric power load, but its initial cluster center is random, and the K value of the cluster number needs to be determined manually, which may lead to deviations in the clustering effect. In order to solve the above problems, a hierarchical K-means clustering algorithm based on Pearson's correlation coefficient is adopted.

Clustering similarity measure function based on Pearson correlation coefficient:

The Pearson coefficient is often used to measure the similarity of curve shapes. Suppose x _i ＝(x _i1 , x _i2 ,…,x _in ), x _j ＝(x _j1 ,x _j2 ,…,x _jn ) are two load curves, and their Pearson correlation coefficients are:

和

分别为x_i和x_j的均值。从式中可以看出，计算时样本减去了样本均值，消除了曲线幅值对相似度的影响。两条曲线形态上越相似，皮尔逊相关系数越低，当两条曲线形态上完全一致时，皮尔逊相关系数为0。如图1所示，图中三条曲线分别为三个用户某天的用电负荷曲线Q₁，Q₂，Q₃。计算Q₂、Q₃与Q₁的欧式距离d和皮尔逊相关系数r，d(Q₁,Q₂)＝133.2，d(Q₂,Q₃)＝346.4，d(Q₁,Q₃)＝363.7，r(Q₁,Q₂)＝0.2524，r(Q₁,Q₃)＝0.2524，r(Q₂,Q₃)＝0。用欧式距离计算，Q₁，Q₂更为接近，但实际上Q₂，Q₃形态上完全一致，计算出来的皮尔逊系数为0。除此之外，现有用余弦相似度作为形态相似性度量，但余弦相似度受曲线幅值影响。因此，选用皮尔逊相关系作为负荷曲线形态相似性的度量。In the formula:

and

are the mean values of x _i and x _j respectively. It can be seen from the formula that the sample mean is subtracted from the sample during calculation, which eliminates the influence of the curve amplitude on the similarity. The more similar the two curves are in shape, the lower the Pearson correlation coefficient is. When the two curves are completely consistent in shape, the Pearson correlation coefficient is 0. As shown in Figure 1, the three curves in the figure are respectively the electricity load curves Q ₁ , Q ₂ , and Q ₃ of three users on a certain day. Calculate the Euclidean distance d and Pearson correlation coefficient r between Q ₂ , Q ₃ and Q ₁ , d(Q ₁ ,Q ₂ )=133.2, d(Q ₂ ,Q ₃ )=346.4, d(Q ₁ ,Q ₃ ) =363.7, r(Q ₁ ,Q ₂ )=0.2524, r(Q ₁ ,Q ₃ )=0.2524, r(Q ₂ ,Q ₃ )=0. Calculated by Euclidean distance, Q ₁ and Q ₂ are closer, but in fact Q ₂ and Q ₃ are completely consistent in shape, and the calculated Pearson coefficient is 0. In addition, cosine similarity is currently used as a morphological similarity measure, but cosine similarity is affected by the amplitude of the curve. Therefore, the Pearson correlation is chosen as the measure of the similarity of load curve shape.

Hierarchical K-means clustering method principle:

The K-means algorithm is a classic partition algorithm, and its clustering result depends on the initial cluster center and the number K of clusters. If it is not selected properly, it will directly lead to unsatisfactory clustering effect. Hierarchical clustering algorithm does not need to select the initial clustering center, and the clustering effect is stable, but its computational complexity is high, and it is not suitable for large-scale data clustering. Therefore, combining the two to establish a hierarchical K-means clustering algorithm can reduce computational complexity and improve clustering stability. The steps of the hierarchical K-means clustering algorithm are:

(1) Repeat K-means clustering with the same K value multiple times, and record each cluster center;

(2) Perform hierarchical clustering for all recorded cluster centers;

(3) The cluster center obtained by hierarchical clustering is used as the initial cluster center, and then K-means clustering is performed to obtain the final clustering result.

This algorithm can solve the problem of randomness of the initial cluster centers of K-means clustering, but the number of clusters K still needs to be determined manually. The selection of clustering index can determine the optimal number of clusters.

Selection of clustering indicators:

The optimal clustering number K can be determined by selecting an appropriate clustering quality evaluation index and verifying the clustering effectiveness. Based on the existing graph-based effectiveness index, a clustering evaluation index of cosine similarity is constructed. This method is also suitable for cluster analysis based on Pearson's correlation coefficient. The evaluation indicators are:

In the formula: K is the number of clusters, v _c is the cluster center of each cluster; n _K is the number of samples contained in the current K-th cluster. This indicator represents the sum of the distances from various samples to their corresponding cluster centers. V decreases with the increase of the number of clusters. When the downward trend of V value tends to be flat, the corresponding number of clusters is the optimal cluster. number.

2. Load forecasting method and process

Load Forecasting Method:

Different forecasting methods have different forecasting effects for different load patterns. In order to improve the forecasting accuracy, the best forecasting model suitable for the load pattern is selected from the three forecasting methods of BPNN, SVR and LR.

BP neural network algorithm mechanism:

Artificial neural network has self-learning, self-organization, self-adaptation and strong nonlinear function approximation ability, and has strong fault tolerance. Figure 2 shows the neuron model structure.

In the figure: P is the input vector, b is the bias constant, w is the weight vector from the input signal to the neuron, f is the activation function, and y is the output signal of the neuron. The output of the neuron is

y=f(wP+b) (25)

As a commonly used artificial neural network, BP neural network consists of three parts: input layer, hidden layer and output layer. The neurons between each layer are fully connected. Its structure is shown in Figure 3.

The BP neural network algorithm is mainly composed of two parts: one is the forward propagation of the signal, the other is to calculate the error between the output value and the real value, and then propagate the error back, and continuously correct the parameters of each neuron through the L-M (Levenberg-Marquardt) algorithm . After the parameters are corrected, train again until the training result reaches the error requirement or the maximum number of training times. The BP neural network model has no specific mathematical expression, and its model is completely characterized by the network structure and parameters.

Support vector machine regression algorithm mechanism:

Support vector regression is an important application branch of support vector machine (SVM). It maps the input vector to a high-dimensional feature space through nonlinear mapping, and obtains a regression hyperplane through calculation, so that the distance sum of all sample points in the set to the hyperplane is the smallest. Nonlinear mapping is also called kernel function, and commonly used kernel functions include linear function, polynomial function, Gaussian function, radial basis function, etc. Set the input sample as M={(xi _, y _i ),i=1,2,…n}, x _i ∈ R ^d , y _i ∈ R, the hyperplane function expression obtained by SVR is:

f(x)=ωΦ(x)+b (26)

Where: Φ( ) is the kernel function, b is the threshold, and ω is the weight vector.

The SVR loss function is:

In the formula: ε is the distance between the sample point and the hyperplane that can be tolerated, that is, when the distance between the sample point and the hyperplane is less than or equal to ε, the loss is 0. The solution of the model has already been described in many existing technologies, and will not be repeated here.

Linear regression algorithm mechanism:

Linear regression (linear regression, LR) is an analysis method that uses a regression equation (function) to model the relationship between one or more independent variables x and dependent variables y. Its model is:

f(x)＝w ₀ +w ₁ x ₁ +w ₂ x ₂ +...+w _n x _n (28)

y=f(x)+δ (29)

In the formula: w is the weight coefficient, and δ is the residual. The main purpose of linear regression is to obtain the coefficient w so that for all sample points M={( _xi ,y _i ),i=1,2,n} in the set, the distance from f(x) is the shortest. Linear regression generally uses the least squares method to find its equation.

ω ₁ ,...,ω _n ＝(X ^T Y) ^-1 X ^T Y (31)

The use of linear regression for load forecasting requires less calculation, and the relationship between input features and forecast information can be directly obtained, which has good interpretability.

3. Load forecasting process based on pattern matching

Enter the feature parameter section:

To predict the load Pt at time t, it is necessary to input characteristic parameters that are highly correlated with the load at time t to the model. Historical load, weekdays, weekends and the time of day have a great influence on the power load. The selected input characteristic parameters are:

x＝{P _th ,P _th-1 ,P _th-2 ,P _t-h+1 ,P _t-h+2, P _t-2h ,P _t-2h-1 ,P _t-2h+1 ,

P _t-3h ,P _t-7h ,weekday,hour}

In the formula, the first seven are historical load data, h is the number of hours in a day when the load data is collected, weekday and hour represent the week and hour respectively.

The matching part of electricity load pattern based on Pearson correlation coefficient:

Set a month to be calculated as 28 (the same below). When predicting the load on the q+1th (q=0, 1, ... 27) day of a certain month, it is necessary to select the load data of the 28 days before the forecast date for pattern matching, which includes The data data1 of the first q days of this month, and the load data data2 of the 28-q after the previous month. In order to make the pattern matching time consistent, data1 data is moved before data2, and spliced into a user's completed monthly load data data3 from the beginning of the month to the end of the month. Calculate the Pearson correlation coefficient between each user data3 and each monthly load pattern, and classify each user into the most similar monthly load pattern with its lowest value. Pattern matching is shown in Figure 4.

The overall process part of load forecasting:

The overall process of the proposed user cluster prediction method based on pattern adaptive matching is shown in Figure 5, and its specific steps are:

(1) Preprocessing the user load data.

(2) Taking the load of all users in one month (calculated according to 28 days) as a sample, perform multiple K-means clustering based on Pearson's coefficient on the sample, calculate the clustering index V, and obtain the optimal clustering number K _b .

(3) Select the number of clusters as K _b , and perform hierarchical K-means clustering based on Pearson correlation coefficient to obtain K _b category monthly loads, and take the cluster center of each category of monthly loads as the monthly load pattern.

(4) Construct BPNN, SVR and LR models for each monthly load pattern. The best tested model is selected to match the monthly load pattern.

(5) Match the load of the month before each user's forecast date with the monthly load that is most similar to it.

(6) Sum the monthly loads of users matched to the same mode, and calculate the matching degree with the monthly load mode, that is, the Pearson correlation coefficient.

(7) Carry out load forecasting with the best forecasting model of monthly load mode, and add up the forecasting results of each mode to obtain the final load forecasting result.

(设

)，方可使用BPNN预测模型，否则使用该月负荷模式的次优预测模型，以解决BPNN模型训练时出现的“过拟合”，导致泛化能力差，预测结果不好的问题。(8) In step (7), if the best prediction model of the monthly load pattern is BPNN, and the matching degree between the obtained user load data and the monthly load pattern is less than a certain threshold

(set up

), the BPNN forecasting model can be used, otherwise, the suboptimal forecasting model of the monthly load pattern is used to solve the problem of "overfitting" that occurs during BPNN model training, resulting in poor generalization ability and poor forecasting results.

(9) Use MAPE and RMSE to evaluate the effect of user cluster prediction.

In the formula: e is the actual value, o is the predicted value, and N is the predicted number.

Calculation analysis:

The grid-side power consumption data of 300 households with photovoltaic power generation systems in Australia were selected for analysis. The data length is from July 1, 2010 to June 30, 2011. The temporal resolution is 30 minutes.

The result of monthly load division:

Select the data of 300 households in the 10 months before July 1, 2010, a total of 3000 monthly load data samples. According to the above step (2), the clustering index is obtained as shown in Figure 6.

It can be seen from the figure that when the value of K is between 9 and 12, the downward trend of the evaluation index V slows down, and K _b =9 is selected as the optimal number of clusters. Then perform hierarchical K-means clustering to obtain 9 types of monthly load patterns. As shown in Figure 7.

It can be seen from the figure that the fluctuation patterns of the nine monthly load patterns are different, and the daily load fluctuations of the first, second, third, fifth, eighth, and ninth patterns are relatively stable and regular. The daily load amplitude of the fourth type of mode has a tendency to increase. The load fluctuations of the 6th and 7th modes are more random.

There are a total of 1344 points in the monthly load pattern. If the data of the monthly load pattern are connected end to end, the data at the end of the monthly load can be used to predict the head-end load of the monthly load, and 1344 training samples can be constructed, and the ones that are uniform and have replacement can be selected 1344 samples were used for forecasting model prediction training, about 63% of the samples were used for training, and the remaining samples that were not drawn were used for testing. The test results obtained are shown in Table 1 below.

Table 1 Test error MAPE of different models for various monthly load modes

Tab 1MAPE of different model test errors for various monthly load patterns

The bold font in the table is the test error of the best model suitable for this type of monthly load pattern. For example: the SVR prediction model for the first type of monthly load mode has the best effect, and the test error is 0.087.

Experimental object setup and result analysis:

In order to verify the effectiveness of the proposed method, the following four schemes are set up for load forecasting.

Option 1, without pattern division, use BPNN to predict directly.

Scheme 2, without pattern division, use SVR to predict directly.

Option 3, without pattern division, use LR to directly predict.

Scheme 4, pattern division and matching are carried out, and each pattern is predicted according to the best prediction method.

The load forecast is carried out on the first 7 days of November, and the forecast results are shown in Figure 8 and Table 2:

Table 2 User cluster prediction results of different experimental schemes

Tab 2Consumer cluster prediction results of different experimental schemes

Comparing the above results, it is found that compared with other methods using a single model, the proposed adaptive pattern matching prediction method can achieve the best prediction effect, and can improve the prediction accuracy by about 0.6%.

Select the pattern matching results on the 7th day of the 11th month as shown in Table 3 below:

Table 3 Pattern matching results on the 7th day of November

Tab 3Pattern matching results on November 7

It can be concluded from Table 3 that the fewer users matched by the model, the higher the matching degree, that is, the lower the similarity between the user load data and the monthly load pattern, which will make the prediction effect worse to a certain extent, but because The number of users is small, and the impact on the overall load forecasting is small.

The above are the preferred embodiments of the present invention, and all changes made according to the technical solution of the present invention, when the functional effect produced does not exceed the scope of the technical solution of the present invention, all belong to the protection scope of the present invention.

Claims (4)

1. The short-term load forecasting method based on pattern matching is characterized in that, comprises the following steps: S1. Using the hierarchical K-means clustering algorithm based on the Pearson correlation coefficient, the user's historical load data is optimally divided into various monthly load patterns; S2. For each monthly load pattern, select the pattern matching algorithm with the best verification effect from feedback neural network, support vector machine regression and linear regression; S3. According to the similarity between the user's latest monthly load data and the monthly load pattern, the user's electricity consumption data is adaptively matched with the monthly load pattern, and the prediction results of each pattern are added to obtain the total prediction result. 2. the short-term load forecasting method based on pattern matching as claimed in claim 1, is characterized in that, step S1 is specifically as follows: Clustering similarity measure function based on Pearson correlation coefficient: Suppose x _i ＝(x _i1 , x _i2 ,…,x _in ), x _j ＝(x _j1 ,x _j2 ,…,x _jn ) are two load curves, and their Pearson correlation coefficients are:

In the formula:

and

are the mean values of x _i and x _j respectively; Hierarchical K-means clustering method principle: Repeat K-means clustering with the same K value multiple times, and record each cluster center; Perform hierarchical clustering on all recorded cluster centers; The cluster center obtained by hierarchical clustering is used as the initial cluster center, and then K-means clustering is performed again to obtain the final clustering result; Selection of clustering indicators: The evaluation indicators are:

In the formula: K is the number of clusters, v _c is the cluster center of each cluster; n _K is the number of samples contained in the current K-th cluster. 3. the short-term load forecasting method based on pattern matching as claimed in claim 2, is characterized in that, step S2 is specifically as follows: The mechanism formula of BP neural network algorithm is as follows:

Among them: P is the input vector, b is the bias constant, w is the weight vector from the input signal to the neuron, f is the activation function, and y is the output signal of the neuron; the output of the neuron is y=f(wP+b) (3) The BP neural network consists of three parts: the input layer, the hidden layer and the output layer, and the neurons between each layer are fully connected; The BP neural network algorithm is composed of two parts: one is the forward propagation of the signal, the other is to calculate the error between the output value and the real value, and propagate the error back, and continuously correct the parameters of each neuron through the L-M algorithm; Carry out training until the training result reaches the error requirement or the maximum number of training times; Support vector machine regression algorithm mechanism: The input vector is mapped to the high-dimensional feature space through nonlinear mapping, and a regression hyperplane is obtained by calculation, so that the distance sum of all sample points in the set to the hyperplane is the smallest; nonlinear mapping is also called a kernel function; set the input The sample of is M={( _xi ,y _i ),i=1,2,…n}, x _i ∈ R ^d , y _i ∈ R, the hyperplane function expression obtained by SVR is: f(x)=ωΦ(x)+b (4) In the formula: Φ( ) is the kernel function, b is the threshold, and ω is the weight vector; The SVR loss function is:

In the formula: ε is the distance from the tolerance sample point to the hyperplane, that is, when the distance between the sample point and the hyperplane is less than or equal to ε, the loss is 0; Linear regression algorithm mechanism: Linear regression is an analytical method that uses a regression equation to model the relationship between one or more independent variables x and dependent variables y; its model is: f(x)＝w ₀ +w ₁ x ₁ +w ₂ x ₂ +...+w _n x _n (6) y=f(x)+δ (7) In the formula: w is the weight coefficient, δ is the residual error; the main purpose of linear regression is to obtain the coefficient w so that for all sample points in the set M={( _xi ,y _i ),i=1,2,…n } The distance from f(x) is the closest; linear regression uses the least square method to find its equation;

ω ₁ ,...,ω _n ＝(X ^T Y) ^-1 X ^T Y (9) Using linear regression to do load forecasting requires a small amount of calculation, and the relationship between input features and forecast information can be directly obtained. 4. the short-term load forecasting method based on pattern matching as claimed in claim 3, is characterized in that, step S3 is specifically as follows: Input feature parameters: To predict the load P _t at time t, it is necessary to input characteristic parameters to the model that are highly correlated with the load at time t; the input characteristic parameters are: x＝{P _th ,P _th-1 ,P _th-2 ,P _t-h+1 ,P _t-h+2 ,P _{t-2h ,P t-2h -1} ,P _t-2h+1 ,P _t-3h ,P _t-7h ,weekday,hour} In the formula, the first seven are historical load data, h is the number of hours in a day when the load data is collected, weekday and hour represent the week and hour respectively; Electric load pattern matching based on Pearson correlation coefficient: Set a month to be calculated as 28. When predicting the load on day q+1 (q=0, 1, ... 27) of a certain month, it is necessary to select the load data of the 28 days before the forecast date for pattern matching, which includes the q The data data1 of the day, and the load data data2 of 28-q after the previous month; in order to make the pattern matching time consistent, the data of data1 is moved before data2, and spliced into a user's completed monthly load data data3 from the beginning of the month to the end of the month; Calculate the Pearson correlation coefficient between each user data3 and each monthly load pattern, and classify each user into the most similar monthly load pattern with its lowest value; The overall process of load forecasting: (1) Preprocessing the user load data; (2) Taking the monthly load of all users as a sample, perform multiple K-means clustering based on Pearson's coefficient on the sample, calculate the clustering index V, and obtain the best clustering number K _b ; (3) Select the number of clusters as K _b , perform hierarchical K-means clustering based on the Pearson correlation coefficient, and obtain the monthly load of the K _b category, and use the cluster center of each category of monthly load as the monthly load pattern; (4) Build BPNN, SVR and LR models for each monthly load pattern; select the model with the best test effect to match the monthly load pattern; (5) Match the load of the month before each user's forecast with the monthly load that is most similar to it; (6) Sum the monthly loads of users matched to the same pattern, and calculate the matching degree with the monthly load pattern, that is, the Pearson correlation coefficient; (7) Carry out load forecasting with the best forecasting model of the monthly load mode, and add up the forecasting results of each mode to obtain the final load forecasting result; (8) In (7), if the best prediction model of the monthly load pattern is BPNN, and the matching degree between the obtained user load data and the monthly load pattern is less than a certain threshold

Then use the BPNN forecasting model, otherwise use the suboptimal forecasting model of the monthly load pattern; (9) Use MAPE and RMSE to evaluate the user cluster prediction effect;

In the formula: e is the actual value, o is the predicted value, and N is the predicted number.

Images