Short-term load prediction method based on multivariate user portraits

Abstract

The invention discloses a short-term load prediction method based on multivariate user portraits, which overcomes the problems of lack of objectivity and low prediction accuracy of short-term load prediction of a power system in the prior art, and comprises the following steps: acquiring historical power consumption data of users, extracting features, and drawing typical power consumption curves of the users; clustering the typical power consumption curves of the users by adopting a stepping clustering algorithm to obtain user groups with different power consumption modes, and generating user portraits; and for different user groups obtained by clustering, respectively constructing a deep learning time sequence short-term load prediction model, and carrying out classified short-term load prediction on users in different power utilization modes. Users are clustered by adopting a stepping clustering algorithm, so that the space-time complexity of the algorithm is reduced, and the clustering result is more accurate; corresponding deep learning time sequence short-term load prediction models can be established for multiple users in different power utilization modes, so that the short-term load of a power system can be predicted quickly and accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of electricity load forecasting, and in particular to a short-term load forecasting method based on multi-user portraits. Background Art

[0002] Fast and accurate power load forecasting is conducive to improving the economy and reliability of the power system, and is the prerequisite for realizing automatic power generation control and economic dispatch control. With the gradual reform of the modern power industry, accurate power load forecasting has largely determined the trend and direction of the reform, but nonlinearity and volatility have increased the difficulty of load forecasting. The influence of economic, weather and other factors has led to changes in load every day, week and year. The accuracy of short-term load forecasting affects the utilization rate of electric energy and the safety and reliability of power system operation.

[0003] At present, a large number of methods have been proposed for short-term power system load forecasting, including traditional time series, neural networks, wavelet analysis, etc. With the increasing complexity of my country's power system and the development of smart grids, the factors affecting the load are also becoming more complex. Therefore, when establishing a forecasting model, determining the relevant factors affecting the load is the first problem to be solved.

[0004] Among the many prediction methods, most models only rely on experience to select influencing factors, which not only lacks objectivity but also affects the prediction accuracy. Moreover, for power systems, only a single prediction method is often used to predict the user's power load. For the current smart grid system under comprehensive management, users are diversified and power consumption is gradually diversified. Therefore, it is urgent to adopt corresponding power load prediction models for the types of multiple users under smart grids, quickly and accurately make short-term power load forecasts, and improve the safety and reliability of power system operation. Summary of the invention

[0005] The purpose of the present invention is to overcome the problems in the prior art of lack of objectivity and low prediction accuracy in short-term load forecasting of power systems, and to provide a short-term load forecasting method based on multi-user portraits. A step-by-step clustering algorithm is used to cluster users, which reduces the algorithm's time and space complexity and makes the clustering results more accurate. A corresponding deep learning time series short-term load forecasting model can be established for multiple users with different electricity consumption patterns, thereby achieving rapid and accurate prediction of the short-term load of the power system.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: A short-term load forecasting method based on multi-user portraits, characterized by comprising the following steps: S1: Obtain the user's historical electricity consumption data, extract features, and draw the user's typical electricity consumption curve; S2: Use the step-wise clustering algorithm to cluster typical power consumption curves of users, obtain user groups with different power consumption patterns, and generate user portraits; S3: For different user groups obtained through clustering, deep learning time series short-term load forecasting models are constructed respectively to perform classified short-term load forecasting for users with different electricity consumption patterns.

[0007] The step-by-step clustering algorithm is used to cluster users and obtain user groups with different power consumption modes (load types), which reduces the algorithm's temporal and spatial complexity, reduces the key information loss rate, and makes the clustering results more accurate; it can establish a corresponding deep learning time series short-term load forecasting model for multiple users with different power consumption modes, so as to quickly and accurately predict the short-term load of the power system. According to the clustering and prediction results, the power grid security management and control strategy can be accurately formulated to ensure the safe and stable operation of the power grid system.

[0008] Preferably, drawing a typical electricity consumption curve of a user comprises: drawing a plurality of daily electricity consumption curves within a period of time according to the user's historical electricity consumption data; clustering the daily electricity consumption curves to obtain a group of daily electricity consumption curves with different change patterns, and taking the cluster center as the typical electricity consumption curve of the group of daily electricity consumption curves.

[0009] Preferably, the clustering of typical user electricity consumption curves includes: decomposing the typical user electricity consumption curves into low-frequency components and high-frequency components using a discrete wavelet transform method until b discrete wavelet transforms are completed; and clustering the discrete wavelet transforms from the result of the last discrete wavelet transform to the previous layer in sequence.

[0010] As a preferred method, all typical power consumption curves of all users in the area are obtained, and the user and user group to which each typical power consumption curve belongs are recorded to obtain the power consumption curve-user group and the power consumption curve-user group group; the power consumption curve-user group and the power consumption curve-user group group are traversed to record the user group to which the typical power consumption curve of each user belongs in the new group, and the user-power consumption curve-user group group is output to obtain a clear user classification result.

[0011] Preferably, the method includes: S4: for a single user, establishing an identification model to identify the user group to which the user belongs, and according to the user group identification result, selecting a corresponding deep learning time series short-term load forecasting model to perform short-term load forecasting.

[0012] Preferably, among all daily electricity curve groups obtained by clustering, if there is a daily electricity curve group whose number is less than or equal to a% of the number of daily electricity curves in another daily electricity curve group, the daily electricity curve group is removed and the daily electricity curves are clustered again.

[0013] Preferably, step S4 includes: fitting the distribution of each feature of a single user using a Gaussian mixture model; using all user power consumption features as a sample data set to train a recognition model; obtaining power consumption information of a single user, and using the recognition model to obtain the user group to which the user belongs.

[0014] Preferably, the input of the deep learning time series short-term load forecasting model includes the current moment, the user's load at the previous moment, the user's load at the current moment of the previous day, the load of similar users at the previous moment, and the load of similar users at the current moment of the previous day, and the output is the user's load at the current moment.

[0015] Preferably, when performing feature extraction, the Pearson correlation coefficient between the features is calculated, a correlation analysis is performed on the features, and features with high correlation are extracted or merged.

[0016] Preferably, the root mean square error is used to evaluate the prediction accuracy of the deep learning time series short-term load forecasting model, where the standard value is the fitting value of the change law of the electricity consumption curve.

[0017] Therefore, the present invention has the following beneficial effects: 1. The present invention uses a step-by-step clustering algorithm to cluster users and obtain user groups with different electricity consumption patterns, which reduces the algorithm's time and space complexity, reduces the key information loss rate, and makes the clustering results more accurate.

[0018] 2. The present invention can establish a corresponding deep learning time series short-term load forecasting model for multiple users with different electricity consumption modes, so as to achieve rapid and accurate prediction of the short-term load of the power system.

[0019] 3. The present invention can accurately formulate power grid security management and control strategies based on clustering and prediction results to ensure the safe and stable operation of the power grid system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The figure is a flow chart of the overall steps of the method of the present invention.

[0021] Figure 2 The present invention is a flowchart of the steps of clustering typical power consumption curves of users using a step-by-step clustering algorithm. DETAILED DESCRIPTION

[0022] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments: Embodiment 1: like Figure 1In the illustrated embodiment, a short-term load forecasting method based on multi-user portraits can be seen, and its operation process is as follows: Step 1, obtain the user's historical electricity consumption data, perform feature extraction, and draw the user's typical electricity consumption curve; Step 2, use a step clustering algorithm to cluster the user's typical electricity consumption curve, obtain user groups with different electricity consumption patterns, and generate user portraits; Step 3, for the different user groups obtained by clustering, respectively construct deep learning time series short-term load forecasting models, and classify short-term load forecasts for users with different electricity consumption patterns.

[0023] The short-term load forecasting method based on multi-user portraits provided in this embodiment is clustered through user portraits with the support of a large amount of user electricity consumption data, and the clustering results are used as load forecasting objects. Based on the deep learning time series load forecasting model structure, a short-term forecasting model is proposed, which can accurately predict the load output characteristics during the future optimization period, which is of great significance to the real-time optimization control of the three-phase load balancing of the distribution network.

[0024] Specifically, a step-by-step clustering algorithm is used to cluster users, and user groups with different power consumption patterns (load types) are obtained, which reduces the algorithm's temporal and spatial complexity, reduces the key information loss rate, and makes the clustering results more accurate. Based on the clustering results, user portraits are generated, and corresponding deep learning time series short-term load forecasting models can be established for multiple users with different power consumption patterns, so as to quickly and accurately predict the short-term load of the power system. According to the clustering and prediction results, the power grid security management and control strategy can be accurately formulated to ensure the safe and stable operation of the power grid system.

[0025] The power user portrait analysis is based on massive electricity consumption data. It takes power users as the main body, mines information on user personal characteristics and user electricity consumption behavior data, assigns user portrait labels based on user characteristics, analyzes the differences in electricity consumption behaviors of different users, and then explores the differences in user behaviors, providing a theoretical basis for the intelligent management of power companies.

[0026] The technical scheme and technical effects of the present invention will be further described below through specific examples. The following examples are explanations of the present invention and the present invention is not limited to the following examples.

[0027] Step 1: Obtain the user's historical electricity consumption data, extract features, and draw the user's typical electricity consumption curve.

[0028] Among them, drawing the typical electricity consumption curve of the user includes: drawing multiple daily electricity consumption curves within a period of time according to the user's historical electricity consumption data; clustering the daily electricity consumption curves to obtain daily electricity consumption curve groups with different change patterns, and taking the cluster center as the typical electricity consumption curve of the daily electricity consumption curve group.

[0029] According to the above, the same user may get multiple typical electricity consumption curves. Therefore, all typical electricity consumption curves of all users in the area are obtained, and the user to which each typical electricity consumption curve belongs is recorded to obtain the electricity consumption curve-user group, and the user group to which each typical electricity consumption curve belongs is recorded to obtain the electricity consumption curve-user group group; traverse the electricity consumption curve-user group and the electricity consumption curve-user group group, record the user group to which the typical electricity consumption curve of each user belongs in the new group, and output the user-electricity curve-user group correspondence group.

[0030] Step 2: Use the step-by-step clustering algorithm to cluster users’ typical electricity consumption curves, obtain user groups with different electricity consumption patterns, and generate user portraits.

[0031] like Figure 2 As shown, the specific steps are: For a certain user, the user's typical electricity consumption curve is decomposed into low-frequency components and high-frequency components using the discrete wavelet transform method, and then the low-frequency components are decomposed into lower-level low-frequency components and high-frequency components using the discrete wavelet transform method, and then the lower-level low-frequency components are decomposed into lower-level low-frequency components and high-frequency components using the discrete wavelet transform method, until b discrete wavelet transforms are completed; the same dimensionality reduction method is used for all users.

[0032] The low-frequency components and high-frequency components obtained from the last decomposition are clustered, and the clustering results are clustered (merged and reconstructed) with the high-frequency components of the previous layer, and so on, until they are clustered (merged and reconstructed) with the high-frequency components obtained from the first decomposition to obtain the clustering results.

[0033] When clustering, set the clustering data, initialize the cluster center, calculate the distance from the power consumption curve to the cluster center, and assign the power consumption curve to the corresponding cluster cluster.

[0034] Among all the daily electricity consumption curve groups obtained by clustering, if the number of daily electricity consumption curves in a certain daily electricity consumption curve group is less than or equal to a% of the number of daily electricity consumption curves in another daily electricity consumption curve group, then the certain daily electricity consumption curve group is eliminated and the daily electricity consumption curves are clustered again.

[0035] The mean square error is used as the clustering performance evaluation index before and after eliminating a certain day's electricity consumption curve group.

[0036] Step 3: For different user groups obtained through clustering, deep learning time series short-term load forecasting models are constructed respectively to perform classified short-term load forecasting for users with different electricity consumption patterns.

[0037] Classification forecasting is carried out based on the electricity consumption characteristics of user groups.

[0038] Taking a certain user group as an example, the process of establishing a deep learning time series short-term load forecasting model includes: Obtain the electricity usage time of all users in the user group and the corresponding electricity load at that time, and take the current electricity usage time, the load at the current electricity usage time, the load of the user at the previous time, the load of the user at the current time of the previous day, the load of similar users at the previous time, and the load of similar users at the current time of the previous day as a group of data to establish a data set.

[0039] The data set is divided into training data set, test data set and validation data set in the ratio of 7:1:2. The training data set is used to train the deep learning time series short-term load forecasting model, and the load forecasting model parameters are continuously adjusted during the training process; the validation data set is used to verify the accuracy of the load forecasting model and adjust the hyperparameters of the load forecasting model. The relationship between the loss values ​​of the training set and the validation set as the epoch changes is used to determine whether the load forecasting model is overfitting. If so, the training is stopped, and then the load forecasting model structure and hyperparameters are adjusted according to the situation, which greatly saves time; the test data set is used to verify the generalization ability of the load forecasting model.

[0040] After the load forecasting model is established, the root mean square error is used to evaluate the prediction accuracy of the deep learning time series short-term load forecasting model, where the standard value is the fitting value of the power consumption curve change law.

[0041] Based on the above, the input of the deep learning time series short-term load forecasting model includes the current moment, the user's load at the previous moment, the user's load at the current moment of the previous day, the load of similar users at the previous moment, and the load of similar users at the current moment of the previous day, and the output is the user's load at the current moment.

[0042] By adopting the method of this embodiment and using the constructed deep learning time series short-term load forecasting model, clustered short-term load forecasting can be performed for users in a certain area. For example, the electricity consumption data of all users in a certain area is collected, feature extraction is performed, typical electricity consumption curves are drawn, and the typical electricity consumption curves are clustered, so that users in the area are divided into user groups with different electricity consumption patterns according to their electricity consumption behaviors. For users in a certain group, the load forecasting model corresponding to the group is selected, and short-term load forecasting is performed for users in the group based on the collected electricity consumption data, and the forecasting results are visualized.

[0043] Embodiment 2: This embodiment provides a short-term load forecasting method based on multi-user portraits. On the basis of the first embodiment, the short-term load forecasting method is further optimized.

[0044] A short-term load forecasting method based on multi-user portraits, whose operation process is as follows: step one, obtain user historical electricity consumption data, perform feature extraction, and draw user typical electricity consumption curves; step two, use a step-by-step clustering algorithm to cluster user typical electricity consumption curves, obtain user groups with different electricity consumption patterns, and generate user portraits; step three, for different user groups obtained by clustering, respectively construct deep learning time series short-term load forecasting models, and classify users with different electricity consumption patterns for short-term load forecasting; step four: for a single user, establish an identification model to identify the user group to which the user belongs, and according to the user group identification result, select the corresponding deep learning time series short-term load forecasting model for short-term load forecasting.

[0045] The specific manifestations are: Step 1: Obtain the user's historical electricity consumption data, extract features, and draw the user's typical electricity consumption curve.

[0046] The user's historical electricity consumption data can be collected through the user's smart meter, and the extracted features include user number, data time, maximum load value, minimum load value, load peak-to-valley difference, daily load, daily frozen amount, daily maximum load rate, etc.

[0047] Preprocess the collected historical electricity consumption data, including: For outliers, they are removed and treated as missing values.

[0048] Determine whether the data missing rate of the electricity consumption curve is less than the threshold. If not, discard the electricity consumption curve. If so, use the linear difference method to fill the missing values ​​according to the changing trend of the electricity consumption curve.

[0049] For repeated values, the average of the first and last data returned at the same time is used as the user's data at that time to maintain the basic structure of the data set.

[0050] For users whose valid data days do not exceed the threshold, the user data is discarded and not used.

[0051] The collected historical electricity consumption data of users are plotted into user electricity consumption curves, and based on the user electricity consumption curves, the user's typical electricity consumption curves are plotted.

[0052] A typical electricity consumption curve is drawn for a single user, and a step-wise clustering algorithm can also be used to cluster all daily electricity consumption curves of the user, and daily electricity consumption curves with similar change patterns are regarded as one category to obtain initial grouping categories.

[0053] Step 2: Use the step-by-step clustering algorithm to cluster users’ typical electricity consumption curves, obtain user groups with different electricity consumption patterns, and generate user portraits.

[0054] By clustering the power consumption curves, the power consumption curves with similar changing patterns are classified into one group, thereby dividing the users into different user groups.

[0055] When extracting features, the Pearson correlation coefficient between each feature is calculated, and correlation analysis is performed on each feature, and features with high correlation are extracted or merged.

[0056] Because any user may have multiple typical power consumption curves, the typical power consumption curves of any user may belong to user groups with different power consumption patterns. Therefore, the clustering results should be marked to clarify the user group to which a typical power consumption curve of the user belongs.

[0057] Step 3: For different user groups obtained through clustering, deep learning time series short-term load forecasting models are constructed respectively to perform classified short-term load forecasting for users with different electricity consumption patterns.

[0058] Step 4: For a single user, establish an identification model to identify the user group to which the user belongs. Based on the user group identification results, select the corresponding deep learning time series short-term load forecasting model for short-term load forecasting.

[0059] A Gaussian mixture model is used to fit the distribution of each feature of a single user; all user power consumption features are used as sample data sets to train a recognition model; power consumption information of a single user is obtained, and the user group to which the user belongs is obtained using the recognition model.

[0060] In other embodiments, for load prediction of a single user, the user's historical electricity consumption data can also be divided into a training set and a test set, and a prediction model can be constructed for prediction. The prediction model input is the user's load at the previous moment, the user's load at the same moment of the previous day, and the current moment, and the output is the user's load at the current moment.

[0061] Embodiment three: This embodiment provides a short-term load forecasting method based on multi-user portraits, and implements the short-term load forecasting method based on multi-user portraits in a specific application scenario.

[0062] With the continuous development of the economy, the use of relevant energy is also increasing. The development of non-renewable energy will lead to energy shortages and affect the ecological environment. The environmental problems we face are becoming increasingly obvious. Renewable energy has emerged and its utilization rate has gradually increased. Renewable energy is concentrated on the development of wind and solar energy. Renewable energy not only has strong renewability, but also has convenient energy conversion. As a novel energy generation method, solar photovoltaic power generation has injected a steady stream of vitality into the development of my country's economy and has been vigorously promoted in China. When photovoltaic power sources are connected to the distribution network, it will affect the stability of the distribution network. If no measures are taken, some power quality problems may occur, including three-phase load imbalance, voltage problems, harmonic pollution, etc.

[0063] Distributed photovoltaic power will generate certain pollution during the access process, which is collectively called harmonic pollution. Since the type of current changes from the original direct current to alternating current, it is easy to generate harmonics during the change. When a large number of photovoltaic power sources are connected, the number of harmonics will continue to increase, causing the frequency of harmonics to increase in amplitude. The generated harmonic pollution will not only affect the stability of the distribution network, but also generate certain radiation, affecting people's physical safety.

[0064] For the distribution network, the quality of electric energy plays a very important role. During the process of distributed photovoltaic intervention, it will affect the voltage distribution of the distribution network. The advancement of the distribution network is also constantly upgrading with the changes of the times. The operation mode of the traditional distribution network is different. It often presents a radial shape, and the current will decrease layer by layer according to the direction of presentation. Distributed photovoltaic access can change the emergence of this problem. It will change the original flow direction, which will cause large fluctuations in the strategy within the distribution network, and will cause the local voltage to increase or decrease. For photovoltaic power stations, strict control must be carried out to reasonably control the voltage situation.

[0065] The short-term load forecasting method based on multi-user portraits provided in this embodiment is adopted. With the support of a large amount of user electricity consumption data, user portraits are clustered, and the clustering results are used as load forecasting objects. Based on the deep learning time series load forecasting model structure, a short-term forecasting model is proposed, which can accurately predict the load output characteristics during the future optimization period, so that the distribution network can be accurately managed and controlled safely, which is of great significance to the real-time optimization control of the three-phase load balancing of the distribution network.

[0066] The short-term load forecasting method based on multi-user portraits provided in this embodiment includes: Step 1: Obtain the user's historical electricity consumption data, extract features, and draw the user's typical electricity consumption curve.

[0067] Step 2: Use the step-by-step clustering algorithm to cluster users’ typical electricity consumption curves, obtain user groups with different electricity consumption patterns, and generate user portraits.

[0068] Users can be initially grouped into several user groups, including those with good power consumption stability, those with poor power consumption stability, those with medium power consumption stability, those with low temperature sensitivity, and those with high temperature sensitivity. Then, a step-by-step clustering algorithm is used to further subdivide and cluster the users in each user group, because even users with poor power consumption stability have different power consumption patterns.

[0069] Step 3: For different user groups obtained through clustering, deep learning time series short-term load forecasting models are constructed respectively to perform classified short-term load forecasting for users with different electricity consumption patterns.

[0070] For user groups with different electricity consumption patterns, different types of deep learning time series short-term load forecasting models can be established for prediction.

[0071] At the same time, when predicting the short-term load of users, the reactive power of the distribution network can also be controlled. The specific performance is: 1. Collect distribution network information and analyze the distribution network status.

[0072] For example, a substation in place A has 10 20kv lines with a cable rate of 100%; a substation in place B has 2 20kv lines with a cable rate of 100%; a substation in place A has 18 10kv lines with a cable rate of 11 lines.

[0073] 2. Obtain the reactive power reverse transmission data of dedicated transformers and public transformers.

[0074] For a feeder, a reactive data point is collected every t minutes, and a total of (3600 / t) reactive data points are collected within 24 hours. Then the reactive reverse transmission of the feeder is: Daily feeder reactive reverse transmission amount = Sum((3600 / t) points of reactive power data) / -24.

[0075] The reactive power value at a point represents the reactive power value within t minutes.

[0076] Calculate the reactive power of the public transformer: Public transformer reactive power = public transformer four-quadrant reactive power - public transformer one-quadrant reactive power; Special transformer reactive power = special transformer four-quadrant reactive power - special transformer one-quadrant reactive power.

[0077] Distribution line reactive power = feeder reactive power - public transformer net reactive power - private transformer net reactive power.

[0078] 3: Carry out short-term load forecasting for users on the user side of the distribution network.

[0079] The method provided in this embodiment is used to perform short-term load forecasting on the user side of the distribution network. Based on the load forecasting results and the reactive power reverse transmission data of the distribution network, the distribution network is regulated to solve the reactive power reverse transmission problem of the distribution network, and the user group of large users is controlled at a power factor below 0.98.

[0080] The control plan includes: (1) Install a reactor on the 10kV bus side.

[0081] (2) Based on the short-term load forecast results, SVG is installed in distribution transformers where power reverse is more serious, and dynamic compensation of reactive power is implemented according to the power factor or voltage level of the load.

[0082] Embodiment 4: This embodiment provides a short-term load forecasting system based on multiple user portraits, which is brought into a specific application scenario to execute the short-term load forecasting method based on multiple user portraits in Embodiment 1.

[0083] A short-term load forecasting system based on multi-user portraits is installed inside an electric power system. The short-term load forecasting system includes: a user data acquisition module, a storage module, a data preprocessing module, a feature extraction module, a typical electricity consumption curve drawing module, a clustering module, a deep learning time series short-term load forecasting model modeling module, and a terminal display module. The storage module is respectively connected to the user data acquisition module, the data preprocessing module, the feature extraction module, the typical electricity consumption curve drawing module, the clustering module, and the deep learning time series short-term load forecasting model modeling module. The user data acquisition module is connected to the data preprocessing module, the data preprocessing module is connected to the feature extraction module and the typical electricity consumption curve drawing module, the clustering module is connected to the feature extraction module and the typical electricity consumption curve drawing module, and the deep learning time series short-term load forecasting model modeling module is connected to the clustering module and the terminal display module.

[0084] The user data collection module is used to obtain the historical electricity consumption data of the proxy users and external influence data such as weather and date.

[0085] The storage module is used to store the data generated by each module and the related application programs and configuration files that can implement the method of this embodiment. It can be a random access memory RAM, a read-only memory ROM, a flash memory, a hard disk HDD or a solid state drive SSD, or a combination of several of them.

[0086] The data preprocessing module performs data preprocessing on the collected historical electricity consumption data of users, including correction and normalization of abnormal data, and supplementation of missing values.

[0087] The feature extraction module extracts features from the preprocessed historical electricity consumption data of users.

[0088] The typical electricity consumption curve drawing module draws multiple daily electricity consumption curves within a period of time based on the user's historical electricity consumption data; clusters the daily electricity consumption curves to obtain daily electricity consumption curve groups with different change patterns, and uses the cluster center as the typical electricity consumption curve of the daily electricity consumption curve group, thereby realizing the drawing of the user's typical electricity consumption curve.

[0089] The clustering module uses a step-by-step clustering algorithm to cluster users' typical electricity consumption curves to obtain user groups with different electricity consumption patterns.

[0090] The deep learning time series short-term load forecasting model modeling module establishes a deep learning time series short-term load forecasting model for each user group based on the clustering results of the clustering module, and selects the load forecasting model with the highest prediction accuracy as the final load forecasting model for short-term load forecasting.

[0091] The terminal display module provides a visual display of the user's short-term load forecast results.

[0092] The above-described embodiment is only a preferred solution of the present invention and does not limit the present invention in any form. There are other variations and modifications without exceeding the technical solution described in the claims.

Claims

1. A short-term load forecasting method based on multi-user portraits, characterized in that: include: S1: Obtain the user's historical electricity consumption data, extract features, and draw the user's typical electricity consumption curve; S2: Use the step-wise clustering algorithm to cluster typical power consumption curves of users, obtain user groups with different power consumption patterns, and generate user portraits; S3: For different user groups obtained through clustering, deep learning time series short-term load forecasting models are constructed respectively to perform classified short-term load forecasting for users with different electricity consumption patterns.

2. A short-term load forecasting method based on multi-user portraits according to claim 1, characterized in that: The drawing of the typical power consumption curve of the user includes: drawing a plurality of daily power consumption curves within a period of time according to the historical power consumption data of the user; clustering the daily power consumption curves to obtain a group of daily power consumption curves with different change rules, and taking the cluster center as the typical power consumption curve of the group of daily power consumption curves.

3. The short-term load forecasting method based on multi-user portraits according to claim 1 is characterized in that: The clustering of typical user electricity consumption curves includes: using a discrete wavelet transform method to decompose the typical user electricity consumption curve into low-frequency components and high-frequency components until b discrete wavelet transforms are completed; and clustering the discrete wavelet transforms from the result of the last discrete wavelet transform to the previous layer in sequence.

4. A short-term load forecasting method based on multi-user portraits according to claim 1 or 3, characterized in that: The step S2 includes: obtaining all typical electricity consumption curves of all users in the area, and recording the user and user group to which each typical electricity consumption curve belongs, to obtain the electricity consumption curve-user group and the electricity consumption curve-user group group; traversing the electricity consumption curve-user group and the electricity consumption curve-user group group, recording the user group to which the typical electricity consumption curve contained in each user belongs in a new group, and outputting the user-electricity consumption curve-user group group.

5. A short-term load forecasting method based on multi-user portraits according to claim 1, 2 or 3, characterized in that: The method comprises: S4: for a single user, establishing an identification model to identify the user group to which the user belongs, and according to the user group identification result, selecting a corresponding deep learning time series short-term load forecasting model to perform short-term load forecasting.

6. The short-term load forecasting method based on multi-user portraits according to claim 2 is characterized in that: Among all the daily electricity consumption curve groups obtained by clustering, if the number of daily electricity consumption curves in a certain daily electricity consumption curve group is less than or equal to a% of the number of daily electricity consumption curves in another daily electricity consumption curve group, then the certain daily electricity consumption curve group is eliminated and the daily electricity consumption curves are clustered again.

7. The short-term load forecasting method based on multi-user portraits according to claim 5 is characterized in that: The step S4 includes: using a Gaussian mixture model to fit the distribution of each feature of a single user; using all user power consumption features as a sample data set to train a recognition model; obtaining power consumption information of a single user, and using the recognition model to obtain the user group to which the user belongs.

8. A short-term load forecasting method based on multi-user portraits according to claim 1, 2, 3 or 6, characterized in that: The input of the deep learning time series short-term load forecasting model includes the current moment, the user's load at the previous moment, the user's load at the current moment of the previous day, the load of similar users at the previous moment, and the load of similar users at the current moment of the previous day, and the output is the user's load at the current moment.

9. A short-term load forecasting method based on multi-user portraits according to claim 1, 2, 3 or 6, characterized in that: When extracting features, the Pearson correlation coefficient between each feature is calculated, and correlation analysis is performed on each feature, and features with high correlation are extracted or merged.

10. A short-term load forecasting method based on multi-user portraits according to claim 1, 2, 3 or 6, characterized in that: The root mean square error is used to evaluate the prediction accuracy of the deep learning time series short-term load forecasting model, where the standard value is the fitting value of the change law of the electricity consumption curve.

Images