Electric quantity prediction interface visual training method and system based on deep learning

Abstract

The invention is suitable for the technical field of computers, and provides an electric quantity prediction interface visual training method based on deep learning, which comprises the following specific steps: S1, reading original historical electric quantity and load data from a database through pymysql, and sending the original historical electric quantity and load data into a model; s3, the model encodes the cleaned data model, and performs multi-downsampling scale decomposition on the encoded data, and S4, the training part calculates the predicted data and real data to obtain the accuracy; according to the method, the mutual influence relation between the scales cannot be found to the maximum extent by adopting a direct addition method for the information between the scales, an attention mechanism is introduced, information fusion is performed after the correlation between the scales is found, and the subsequent prediction accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and more specifically, to a method and system for visualizing and training a power prediction interface based on deep learning. Background Art

[0002] Scientific prediction of regional power grid electricity sales is an important basis for power grid construction planning. For regional power grids, electricity sales forecasting is crucial to voltage and load stability, grid security scheduling, and stable operation. Currently, a time series prediction model based on deep learning is generally used. The past decomposable mixture (PDM) module is used to extract key information from the past, and the future multi-predictor mixture (FMM) module is used to predict future series.

[0003] However, most existing time series prediction models use a direct addition method to fuse information between multiple scales. However, this method is not comprehensive enough and will affect the accuracy of the prediction. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a deep learning-based power prediction interface visualization training method and system.

[0005] To achieve the above object, the present invention provides the following technical solutions: The visualization training method of the power forecasting interface based on deep learning includes the following specific steps: in the S1 step, a long connection is established, the parameters set by the front end are passed to the back-end model, and then the original historical power and load data are read from the database through pymysql and fed into the model; Step S2: Clean the incoming raw data, fill in missing data, and remove outliers to avoid data loss and excessive data deviation. In step S3, the model encodes the cleaned data model and decomposes the encoded data into four scales by multi-downsampling. The seasonal part and the trend cycle part are gradually separated from the information of each scale. S4, step, the training part calculates the accuracy of the predicted data and the actual data, displays a line chart comparing the predicted value and the true value of the power, and has a power correction function. It displays the load value and the maximum and minimum predicted load occurrence time table. The prediction part only retains the predicted value, and the web page realizes the visualization of the results.

[0006] The present invention is further configured as follows: in the S3 step, the seasonal part is downsampled three times, and the trend part is upsampled three times. The seasonal mixing adopts a bottom-up approach to incorporate information of low-level fine-scale time series, which can supplement detailed information for coarse-scale seasonal modeling. The trend mixing is opposite to the seasonal part in that the detailed changes in the trend item may introduce noise and affect the capture of macro trends.

[0007] The present invention is further configured as follows: when performing information fusion between multiple scales of trend and seasonal parts in the S3 step, a method of respectively seeking attention between the two scale information is adopted, and information fusion is performed after correlation is obtained, so as to better extract the relevant influences between multiple scales.

[0008] The present invention is further configured as follows: in the step S3, multi-scale past information is obtained after L PDM blocks. Since sequences of different scales exhibit different dominant changes, their predictions also exhibit different capabilities.

[0009] The present invention is further configured as follows: in the step S3, in order to make full use of the multi-scale information, a method of aggregating multi-scale sequence prediction and presenting a future multi-predictor hybrid block is used to regress each scale, extract from the past information and map it to the output variable.

[0010] A power prediction interface visualization training system based on deep learning includes a training model, which includes a cleaning module, an encoding module, a scale decomposition module and a training module. The output end of the cleaning module is data-connected to the encoding module, the output end of the encoding module is data-connected to the scale decomposition module, the output end of the scale decomposition module is data-connected to the training module, and the output end of the training module is data-connected to the visualization module.

[0011] The present invention is further configured as follows: the cleaning module is used to clean the incoming raw data, fill in missing data and remove outliers; the encoding module can encode the cleaned data model; the scale decomposition module can perform multi-downsampling scale decomposition on the encoded data, and gradually separate the seasonal part and the trend cycle part in the information of each scale; the training module can calculate the data predicted by the scale decomposition module and the real data and obtain the accuracy; the visualization module can visualize the results based on the data calculated by the training module, and display a line chart comparing the predicted value and the true value of the power, a power correction function, and a load value and a table of the time when the maximum and minimum predicted loads occur.

[0012] The advantages of the present invention are: First, the direct addition method of information between scales in the present invention cannot maximize the mutual influence relationship between scales. Introducing the attention mechanism to find the correlation between scales and then perform information fusion is conducive to improving the subsequent prediction accuracy.

[0013] Secondly, the present invention has a visual interface system, provides data analysis functions, integrates multiple models for selection, and can be directly applied to engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a flow chart of the visualization training method for power prediction interface based on deep learning of the present invention; Figure 2 This is a diagram showing the power prediction training of the system interface of the present invention; Figure 3 This is a diagram showing the daily load forecast interface of the system interface of the present invention; Figure 4 This is a diagram showing the multi-algorithm selection interface of the present invention; Figure 5 This is a block diagram of the power prediction interface visualization training system based on deep learning in the present invention. DETAILED DESCRIPTION

[0015] See also Figure 1-4 , the present invention provides the following technical solutions: Specifically, it refers to a visualization training method for the power prediction interface based on deep learning. The specific steps are as follows: S1. Establish a long connection, pass the parameters set by the front-end into the back-end model, and then read the original historical power and load data from the database through pymysql and feed it into the model; S2. Clean the incoming raw data, fill in missing data, and remove outliers to avoid missing data and excessive data deviation. S3. The model encodes the cleaned data model and decomposes the encoded data into four scales by multi-sampling. The seasonal part and the trend cycle part are gradually separated from the information of each scale. S3-1. The seasonal component is downsampled three times, and the trend component is upsampled three times. Seasonal mixing adopts a bottom-up approach, incorporating information from low-level fine-scale time series, which can supplement detailed information for coarse-scale seasonal modeling. Trend mixing targets the trend component in the opposite direction to the seasonal component. Detailed changes in the trend term may introduce noise, affecting the capture of macro trends. When fusing information at multiple scales for trend and seasonality, we use a method to obtain attention between information at two scales, find the correlation, and then fuse the information to better extract the relevant influences between multiple scales. S3-3, after L PDM blocks, multi-scale past information is obtained. Since sequences at different scales exhibit different dominant changes, their predictions also show different capabilities; S3-4. In order to make full use of this multi-scale information, a method of aggregating multi-scale sequence predictions and presenting future multi-predictor hybrid blocks is used to regress each scale, extract past information and map it to the output variable; S4. The training part calculates the accuracy of the predicted data and the real data, displays the power prediction value and the real value comparison line chart, power correction function, displays the load value and the maximum and minimum predicted load occurrence time table, the prediction part only retains the predicted value, and the web page realizes the visualization of the results (see the appendix of the manual). Figure 2-4 ).

[0016] You can choose a variety of models for training according to different regional and industrial scenarios, and select the most suitable prediction model for the scenario. The prediction results are shown in the following five major industry prediction accuracy tables: Power forecast accuracy the whole society All industries resident Agriculture and Forestry industry TimeMixer 91.36% 91.02% 85.46% 94.09% 90.52% TimeMixer+Scale Fusion 92.20% 91.94% 86.62% 94.31% 91.42% According to the above table, the prediction accuracy has been steadily improved in the tests conducted on the electricity data sets of various cities. At the same time, a visual interface is also provided for direct operation. It is only necessary to activate the environment variables on the server, log in to the URL, and enter the database login information to conveniently implement training and prediction related functions, and automatically return the processed data. The training part draws a comparison between the predicted data and the true value to display the prediction accuracy, load pie charts of various industries and the maximum load time, and a table of predicted values ​​for various industries in various cities. It provides a variety of algorithm options. You can choose the most appropriate algorithm for future prediction work based on the accuracy during training and the test chart.

[0017] See also Figure 5Based on the above method, the present invention provides another new technical solution: specifically, it refers to a power prediction interface visualization training system based on deep learning, including a training model, the training model includes a cleaning module, an encoding module, a scale decomposition module and a training module, the cleaning module is used to clean the incoming raw data, fill in the missing data and eliminate abnormal values, the output end of the cleaning module is connected to the encoding module data, so that the encoding module can encode the cleaned data model, the output end of the encoding module is connected to the scale decomposition module data, the scale decomposition module can perform multi-downsampling scale decomposition on the encoded data, and gradually separate the seasonal part and the trend cycle part in the information of each layer of scale, the output end of the scale decomposition module is connected to the training module data, so that the training module can calculate the data predicted by the scale decomposition module and the real data and obtain the accuracy, the output end data of the training module is connected to the visualization module, the visualization module can realize result visualization according to the data calculated by the training module, and display the power prediction value and true value comparison line chart, power correction function, load value and maximum and minimum predicted load occurrence time table.

Claims

1. A visualization training method for power prediction interface based on deep learning, characterized by: The specific steps are as follows: in step S1, a long connection is established, the parameters set by the front end are passed into the back-end model, and then the original historical power and load data are read from the database through pymysql and sent into the model; Step S2: clean the incoming raw data, fill in missing data and remove outliers to avoid missing data and excessive data deviation; In step S3, the model encodes the cleaned data model, performs multi-downsampling scale decomposition on the encoded data, and divides it into four scales, and gradually separates the seasonal part and the trend cycle part from the information of each scale; In step S4, the training part calculates the accuracy of the predicted data and the actual data, displays a line chart comparing the predicted value and the true value of the power, and has a power correction function. It displays the load value and the maximum and minimum predicted load occurrence time table. The prediction part only retains the predicted value, and the web page realizes the visualization of the results.

2. The method for visualizing the power prediction interface based on deep learning according to claim 1 is characterized in that: In the S3 step, the seasonal part is downsampled three times and the trend part is upsampled three times. The seasonal mixing adopts a bottom-up approach and incorporates the information of low-level fine-scale time series, which can supplement detailed information for coarse-scale seasonal modeling. The trend mixing is opposite to the seasonal part in terms of the trend part. The detailed changes in the trend item may introduce noise and affect the capture of macro trends.

3. The method for visualizing the power prediction interface based on deep learning according to claim 2 is characterized in that: When performing information fusion between multiple scales of trend and seasonal parts in the step S3, a method of respectively seeking attention between two scale information is adopted, and information fusion is performed after correlation is obtained, so as to better extract the relevant influences between multiple scales.

4. The method for visualizing the power prediction interface based on deep learning according to claim 3 is characterized in that: In the S3 step, multi-scale past information is obtained after L PDM blocks. Since sequences of different scales present different dominant changes, their predictions also show different capabilities.

5. The method for visualizing the power prediction interface based on deep learning according to claim 4 is characterized in that: In order to make full use of the multi-scale information in the step S3, a method of aggregating multi-scale sequence predictions and presenting a future multi-predictor hybrid block is used to regress each scale, extract past information and map it to the output variable.

6. A power prediction interface visualization training system based on deep learning, based on the power prediction interface visualization training method based on deep learning according to claim 5, comprising a training model, the training model comprising a cleaning module, an encoding module, a scale decomposition module and a training module, characterized in that: The output end of the cleaning module is data-connected to the encoding module, the output end of the encoding module is data-connected to the scale decomposition module, the output end of the scale decomposition module is data-connected to the training module, and the output end of the training module is data-connected to the visualization module.

7. The power prediction interface visualization training system based on deep learning according to claim 6 is characterized by: The cleaning module is used to clean the incoming raw data, complete the missing data and remove the outliers. The encoding module can encode the cleaned data model. The scale decomposition module can perform multi-downsampling scale decomposition on the encoded data and gradually separate the seasonal part and the trend cycle part from the information of each scale. The training module can calculate the data predicted by the scale decomposition module and the real data and obtain the accuracy. The visualization module can visualize the results based on the data calculated by the training module, and display a line chart comparing the predicted value and the true value of the power, the power correction function, and the load value and the maximum and minimum predicted load occurrence time table.

Images