A method for regional short-term power load forecasting that considers user electricity consumption and seasonal characteristics

Abstract

This invention discloses a method for regional short-term power load forecasting that considers user electricity consumption and seasonal characteristics. The method includes: constructing a user classification and feature extraction model under big data conditions; introducing large-scale electricity consumption samples to classify a massive number of users; and extracting typical load curves for different industries based on their electricity load characteristics. The extracted typical daily load curves are then analyzed in conjunction with user electricity consumption characteristics and seasonal characteristics to deeply mine user electricity consumption features. A regional short-term power load forecasting model based on these typical daily load curves is established, and quadratic mode decomposition technology is introduced to predict the typical daily load curves for different user types. Finally, the prediction results of each typical curve are multiplied by the total number of users of that type and summed to obtain the load forecast value for the target area. This invention deeply extracts user electricity consumption behavior and characteristics, reduces the complexity and required data volume of regional power load forecasting, and improves the accuracy of regional power load forecasting.

Description

Technical Field

[0001] This invention relates to the field of integrated energy system technology, specifically a method for predicting regional short-term power load that takes into account user electricity consumption and seasonal characteristics. Background Technology

[0002] In recent years, the large-scale integration of distributed energy sources into distribution networks has brought significant challenges due to their dispersed and random nature. On the power supply side, a large number of new energy sources such as photovoltaics have been integrated into medium- and low-voltage distribution networks; on the load side, the number of electric vehicles has increased dramatically in recent years, leading to the large-scale integration of electric vehicle charging stations. The randomness and volatility of these new energy sources and loads have made the operation and control of distribution networks increasingly complex.

[0003] With the continuous improvement of the information level of power grid users, the power sector has accumulated a large amount of data, containing a wealth of hidden information worth exploring. Finding suitable data processing technologies to analyze users' electricity consumption behavior and characteristics is beneficial for power grid companies to understand users' electricity preferences and provide targeted assistance for load forecasting, thereby improving energy efficiency. How to use appropriate data mining and processing techniques to maximize the use of raw power load data, uncover users' electricity consumption behavior characteristics, and use this as a guiding factor for load forecasting is an urgent problem to be solved. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention proposes a regional short-term power load forecasting method that considers user electricity consumption and seasonal characteristics. The aim is to deeply explore users' electricity consumption behavior and characteristics, thereby providing targeted assistance for load forecasting and improving the accuracy of load forecasting.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A method for forecasting regional short-term power load considering user electricity consumption and seasonal characteristics, which uses only extracted typical daily load curves to forecast the overall short-term power load of the region, includes the following steps:

[0007] S1: Select typical days from the four seasons of spring, summer, autumn and winter, and perform K-means clustering on the load curves of the massive number of users on the typical days to divide the users into different groups.

[0008] S2: Label different groups based on large user electricity consumption samples and extract typical daily load curves for each type of user;

[0009] S3: The second-order mode decomposition technique is used to process various typical daily load curves. That is, the pre-processed raw electrical load data is decomposed into subsequences using the complete set empirical mode decomposition of adaptive noise. Then, the strong non-stationary components in the subsequences are subjected to second-order variational mode decomposition to obtain stationary subsequences.

[0010] S4: Feed each component into the LSTM network for prediction and obtain the prediction result for each component.

[0011] S5: Multiply the daily load curves of each type of user by the number of users of that type and then overlay them to obtain the prediction result of the overall load of the area where each type of user is located.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0013] 1. This invention utilizes large user electricity consumption samples for analysis to achieve accurate classification of large electricity users. By combining industry characteristics and seasonal characteristics, it can extract user electricity consumption behavior and characteristics at a deeper level.

[0014] 2. This invention only requires the extraction of typical daily load curve data of users to have regional load forecasting capabilities with good results. It has certain guiding significance for production and life, and makes up for the deficiency that it is impossible to make overall regional load forecasting when the electricity consumption information of some users is missing.

[0015] 3. This invention introduces mode decomposition technology, which makes the time characteristics of the original load sequence more prominent, making it easier for the neural network to capture information, thereby enabling more effective prediction of power load change trends and improving the accuracy of load forecasting. Attached Figure Description

[0016] Figure 1 This is a general framework diagram of a regional short-term power load forecasting method that considers user electricity consumption and seasonal characteristics according to the present invention;

[0017] Figure 2 This is a basic structural diagram of an LSTM neuron;

[0018] Figure 3 This is the clustering result of the load curves on typical days in spring;

[0019] Figure 4 This is the clustering result of the load curves on a typical summer day;

[0020] Figure 5 This is the clustering result of the load curves for a typical autumn day;

[0021] Figure 6 This is the clustering result of the load curves on a typical winter day;

[0022] Figure 7 This is the predicted result of the typical load curve for the first type of user;

[0023] Figure 8 Typical load curve prediction results for the second type of users;

[0024] Figure 9 This is the predicted result of the typical load curve for the third type of user;

[0025] Figure 10 This is the predicted result of the typical load curve for the fourth type of user;

[0026] Figure 11 This is the predicted result of the typical load curve for the fifth category of users;

[0027] Figure 12 This is the regional overall load prediction result after introducing mode decomposition. Detailed Implementation

[0028] The invention will now be further described with reference to the accompanying drawings.

[0029] like Figure 1 As shown, a method for regional short-term power load forecasting that considers user electricity consumption and seasonal characteristics according to the present invention includes the following steps:

[0030] S1: Select typical days from the four seasons of spring, summer, autumn and winter, and perform K-means clustering on the load curves of the massive number of users collected on typical days to divide users into different groups, including;

[0031] S1-1: The specific process of the K-MEANS algorithm is as follows:

[0032] (1) Determine the initial number of clusters k;

[0033] (2) Initialize k cluster centers;

[0034] (3) Calculate the distance between the sample (the sample is the load curve of 370 users collected below) and the center, and assign each sample to the nearest center;

[0035] (4) Update the cluster centers based on the samples from each cluster;

[0036] (5) Iterate (3) to (4) until the similarity metric function begins to converge.

[0037] The similarity measure function for K-means clustering is expressed as follows:

[0038]

[0039] Where E is the sum of squared errors of all samples, k is the number of clusters, selected using the elbow rule; C iLet p be the i-th cluster; p is the sample point in the cluster space, μ i It is the cluster center of the i-th cluster.

[0040] Here, E is the sum of squared errors of all samples, and k is the sum of squared distances between cluster centers. As the objective function, for a cluster, a smaller sum of squared distances indicates a more compact cluster structure, while a larger sum of squared distances indicates a looser cluster structure. The sum of squared distances decreases with the number of categories, but for data with a certain degree of discriminative power, the sum of squared distances decreases rapidly at a certain critical point, and then decreases slowly. This critical point can be considered as the k value for good clustering performance. Based on this metric, multiple K-means models can be repeatedly trained, selecting different k values ​​to obtain a relatively suitable number of clusters.

[0041] S2: Label different groups based on the electricity consumption samples of large users (as shown in Table 1) and extract the typical daily load curves of various types of users;

[0042] The typical daily load curves of different groups of users obtained in S1 are grouped and averaged to obtain the typical daily load curve of that type of user. Then, the typical daily load curves of various types of users are classified according to the large user electricity consumption examples, and industry labels are affixed at the same time.

[0043] S3: A quadratic mode decomposition technique is used to process various typical daily load curves. This involves decomposing the preprocessed raw electrical load data using Complete Ensemble Empirical Mode Decomposition (CEEMDAN) with adaptive noise to obtain subsequences, and then performing quadratic variational mode decomposition (VMD) on the strongly non-stationary components of the subsequences to obtain stationary subsequences, including:

[0044] S3-1: The specific steps of complete set empirical mode decomposition are as follows:

[0045] 1) Add I pairs of white noise sequences with opposite signs to the original sequence x(t) to obtain the I-th decomposition sequence.

[0046] x i (t)=x(t)+εδ i (t), i∈{1,…,I} (2)

[0047] Where x i (t) represents the new sequence after the i-th addition of positive and negative white noise, ε is the amplitude coefficient of the white noise, and δ i (t) represents the standard white noise sequence added for the i-th time;

[0048] 2) Perform empirical mode decomposition (EMD) on the sequence after adding white noise, and take the first-order intrinsic mode component IMF1(t) and residual component r1(t) obtained after decomposing the original sequence after each addition of white noise;

[0049]

[0050] r1(t)=x(t)-IMF1(t) (4)

[0051] In the formula, IMF1(t) is the first component generated by CEEMDAN decomposition, i.e., the first-order intrinsic mode component; r1(t) is the residual component; IMF1 i (t) is the i-th modal component generated for the first time. The obtained residual component is then added to a white noise sequence for EMD decomposition. The mean of the first component obtained from the decomposition is taken to obtain IMF2(t). Then, IMF2(t) is obtained by subtracting IMF2(t) from r1(t).

[0052]

[0053] Where E1 represents the first-order intrinsic mode component operator; repeating the above steps, we obtain the k+1 order component and residual component:

[0054] r k (t)=r k-1 (t)-IMF k (t), k=2,3,…,K (6)

[0055] Where K represents the number of components generated by CEEMDAN, r k (t) represents the generated k-th residual component, IMF k (t) represents the k-th component generated after CEEMDAN decomposition.

[0056] 3) If the EMD stopping condition is met, i.e. the residual signal of the nth decomposition is monotonic, the iterative process terminates, and the decomposition process of the complete set empirical mode decomposition is completed.

[0057] S3-2: The second-order variational mode decomposition includes the following steps:

[0058] 1) Construct a variational constraint problem; the variational constraint expression is:

[0059]

[0060] In the formula, {u1,u2,…,u K} represents the set of modal components, {ω1,ω2,…,ω k} represents the set of center frequencies, δ(t) represents the Dirac distribution; ω k The center frequency of each intrinsic mode fraction (IMF) is represented by u. k Let f(t) represent the modal components with finite bandwidth, and f(t) be a time-series signal.

[0061] By introducing the Lagrange operator and the penalty factor, the variational constraint problem is transformed into an unconstrained problem.

[0062]

[0063] In the formula, λ(t) represents the Lagrange multiplication operator, α represents the penalty factor, and <·> denotes the inner product; L() represents the Lagrange expression. Let j represent the square of the L2 norm. 2 =-1, f(t) is the original time series signal.

[0064] 2) Optimize by using the alternating multiplication operator, continuously update the modal components and center frequency set, and finally obtain the components with different center frequencies.

[0065] After decomposing the load series using CEEMDAN, stationary and strongly non-stationary components are generated. The strongly non-stationary components contain fluctuations from the original load information, making them difficult to predict. Therefore, we propose a VMD secondary decomposition of the strongly non-stationary components, using the stationary components for subsequent predictions.

[0066] S4: Feed each component into the LSTM network for prediction and obtain the prediction result for each component.

[0067] The basic neuron structure diagram of LSTM is as follows: Figure 2 As shown, the LSTM has three inputs, namely cell state C. t-1 Hidden unit state h t-1 Input quantity x at time t t The output has two parts, one being the cell state C. t and hidden unit state h t t represents time.

[0068] f t =σ(W f ·[h t-1 ,x t ]+b f (9)

[0069] Among them, f t This is the output of the forget gate at time t, where σ is the activation function of the sigmoid function, and W... f Let x be the weight matrix corresponding to the forget gate. t Input variable b at time t f This is the bias value for the forget gate.

[0070] The formula for the input gate is:

[0071] i t =σ(W i ·[h t-1,x t ]+b i (10)

[0072]

[0073]

[0074] Among them, i t This represents the output signal of the input gate. C represents an intermediate state of the cell. t-1 and C t h represents the cell state at the previous and current time points, respectively. t-1 The hidden unit state at the previous time step, where tanh is the hyperbolic tangent activation function. W C and W i These are the weight matrices for the input nodes and the input gate, respectively. C and b i is the bias of the input gate, and * is the product of the elements of the two vectors.

[0075] The formula for the output gate is:

[0076] o t =σ(W o [h t-1 ,x t ]+b o (13)

[0077] h t =o t *tanh(C t (14)

[0078] Among them, o t It refers to the output signal of the output gate at time t, h t W represents the current hidden unit state. o Let b be the weight matrix of the output gate. o This is the bias value of the output gate.

[0079] LSTM continuously updates the cell state through forgetting, remembering, and output phases, retaining previously useful information while discarding useless information. Its basic steps are as follows:

[0080] (1) The input variable x at time t t and the hidden unit state h at time t-1 t-1 The characteristic variables of the three gates and the cell state C at the current moment were calculated. t .

[0081] (2) Update the cell state through the forget gate and the input gate.

[0082] (3) The hidden layer information is passed to the current hidden unit state h through the output gate. t .

[0083] S5: Multiply the daily load curves of each type of user by the number of users of that type and then overlay them to obtain the predicted result of the overall load of the region.

[0084] Three evaluation metrics were selected: mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R-squared, R²) to evaluate the accuracy of the prediction results.

[0085]

[0086]

[0087]

[0088] In the formula, y i and represents the true value and the sample value at time t, respectively; i represents the i-th sample; and n represents the number of samples involved in the calculation. The closer MAPE is to 0, the closer RMSE is to 0, and the closer R² is to 1, the better the model's prediction accuracy.

[0089] To verify its effectiveness, this invention uses electricity load data from 370 users in Portugal in 2014, provided by the University of California, Irvine, with a granularity of 15 minutes. The K-means clustering algorithm was used, Python was employed, and TensorFlow was selected as the deep learning framework. The LSTM neural network was set to one layer with 50 neurons, and the Adam optimization algorithm was used. Dropout layers were added to prevent overfitting.

[0090] To fully demonstrate the effectiveness of the proposed method, typical days from each of the four seasons were selected from the annual electricity load data of 370 users, and K-means cluster analysis was performed, with a cluster size of 5. The clustering results are as follows. Figures 3-6 As shown in Table 1, the user classification results are as follows.

[0091] Table 1 User Classification by Season (Spring, Summer, Autumn, Winter)

[0092]

[0093] The user classification results remained consistent across all four seasons, dividing the 370 users into five categories. Category 1 users had the highest electricity load, with peak loads around 70,000 kW, clearly indicating industrial load. Peak electricity consumption was concentrated between 10:00 and 20:00, showing a distinct and relatively stable peak period. This suggests that this type of load likely occurs outside of normal rest hours and belongs to the general manufacturing sector. Category 2 users had relatively stable loads throughout the day, with peak and off-peak loads both meeting the criteria in the table above, around 5,000 kWh, belonging to the chemical raw materials and chemical products manufacturing sector. Category 3 users had a more stable load trend, with peak, average, and off-peak electricity all around 1,500 kWh, belonging to the light industry sector. Category 4 users were large users, with peak and off-peak electricity around 650 and 450 kWh respectively, basically meeting the standards for the non-metallic mineral manufacturing sector. Category 5 users were basic residential users.

[0094] We will make predictions for any day in 2014, here choosing April 16th. We will then use the typical daily load curves to make predictions for various types of typical load curves. Finally, we will multiply these predictions by the number of users in each type and sum them to obtain the total load prediction for that type of area.

[0095] Among them, the first and fifth categories of users were predicted using LSTM after mode decomposition, while other users were predicted directly using LSTM. The prediction results for the five categories of users are as follows: Figures 7-11 As shown, the final load forecast results are as follows: Figure 12 As shown.

[0096] Before introducing mode decomposition, the RMSE was 7841.430, MAPE was 0.026, and r2_score was 0.984734. For the first type of user, combined with mode decomposition, the RMSE was 4936.553, MAPE was 0.019, and r2_score was 0.993950. It is evident that introducing mode decomposition significantly improves prediction accuracy, and the resulting load predictions effectively reflect load change trends, providing guidance for regional production, daily life, and load regulation.

Claims

1. A method for regional short-term power load forecasting that considers user electricity consumption and seasonal characteristics, characterized in that, The method for forecasting the overall short-term power load of a region using only the extracted typical daily load curves includes the following steps: S1: Select typical days from the four seasons of spring, summer, autumn and winter, and perform K-means clustering on the load curves of the massive number of users on the typical days to divide the users into different groups. S2: Label different groups based on large user electricity consumption samples and extract typical daily load curves for each type of user; S3: The second-order mode decomposition technique is used to process various typical daily load curves. The preprocessed raw electrical load data is first decomposed into subsequences using the complete set empirical mode decomposition of adaptive noise. Then, the strong non-stationary components in the subsequences are subjected to second-order variational mode decomposition to obtain stationary subsequences. S4: Feed each component into the LSTM network for prediction and obtain the prediction result for each component. S5: Multiply the daily load curves of each type of user by the number of users of that type and then overlay them to obtain the prediction result of the overall load of the area where each type of user is located.

2. The method for regional short-term power load forecasting considering user electricity consumption and seasonal characteristics as described in claim 1, characterized in that, In step S1, the calculation process of K-means clustering is as follows: (1) Determine the initial number of clusters k; (2) Initialize k cluster centers; (3) Calculate the distance between the sample and the center, and assign each sample to the nearest center; (4) Update the cluster centers based on the samples from each cluster; (5) Iterate through steps (3) to (4) until the similarity metric function begins to converge; The similarity measure function for K-means clustering is expressed as follows: Where E is the sum of squared errors of all samples, k is the number of clusters, selected using the elbow rule; C i Let p be the i-th cluster; p is the sample point in the cluster space, μ i It is the cluster center of the i-th cluster.

3. The method for regional short-term power load forecasting considering user electricity consumption and seasonal characteristics as described in claim 2, characterized in that, In step S2, big data samples are introduced, users are divided into different categories and labeled, and user electricity consumption characteristics are extracted in depth by combining industry characteristics and seasonal characteristics, and typical daily load curves are extracted. The typical daily load curves of different groups of users obtained in S1 are grouped and averaged to obtain the typical daily load curve of that type of user. Then, the typical daily load curves of various types of users are classified according to the large user electricity consumption examples, and industry labels are affixed.

4. The method for regional short-term power load forecasting considering user electricity consumption and seasonal characteristics as described in claim 3, characterized in that, Step S3 includes: S3-1: The specific steps of complete set empirical mode decomposition are as follows: 1) Add I pairs of white noise sequences with opposite signs to the original sequence x(t) to obtain the I-th decomposition sequence; x i (t)=x(t)+εδ i (t),i∈{1,…,I} (2) In the formula, x i (t) represents the new sequence after the i-th addition of positive and negative white noise, ε is the amplitude coefficient of the white noise, and δ i (t) represents the standard white noise sequence added for the i-th time; 2) Perform empirical mode decomposition on the sequence after adding white noise, and take the first-order intrinsic mode component IMF1(t) and residual component r1(t) obtained after decomposing the original sequence after each addition of white noise; r1(t)=x(t)-IMF1(t) (4) In the formula, IMF1(t) is the first component generated by CEEMDAN decomposition, i.e., the first-order intrinsic mode component; r1(t) is the residual component. It is the first generated i-th mode component; add white noise sequence to the obtained residual component r1(t) and perform empirical mode decomposition, take the first component obtained by decomposition and take the mean to obtain IMF2(t), and then subtract IMF2(t) from r1(t) to obtain r2(t); Where E1 represents the first-order intrinsic mode component operator; repeating the above steps, we obtain the k+1 order component and residual component: r k (t)=r k-1 (t)-IMF k (t),k=2,3,…,K (6) Where K represents the number of components generated by the complete set of empirical modes, r k (t) represents the generated k-th residual component, IMF k (t) represents the kth component generated after complete set empirical mode decomposition; 3) If the EMD stopping condition is met, i.e. the residual signal of the nth decomposition is monotonic, the iterative process terminates, and the decomposition process of the complete set empirical mode decomposition is completed. S3-2: The second-order variational mode decomposition includes the following steps: 1) Construct a variational constraint problem; the variational constraint expression is: In the formula, {u1,u2,…,u K } represents the set of modal components, {ω1,ω2,…,ω k } represents the set of center frequencies, δ(t) represents the Dirac distribution; ω k u represents the center frequency of each intrinsic modal component. K Let f(t) represent the modal components with finite bandwidth, and f(t) be a time-series signal. Introducing the Lagrange operator and penalty factor transforms the variational constrained problem into an unconstrained problem: In the formula, λ(t) represents the Lagrange multiplication operator, α represents the penalty factor, and <·> denotes the inner product; L() represents the Lagrange expression. Let f(t) represent the square of the L2 norm, j be the imaginary number, and f(t) be the original time series signal. 2) Optimize by using the alternating multiplication operator, continuously update the modal components and center frequency set, and finally obtain the components with different center frequencies.

Images