Combined model electric quantity prediction method based on CNN and VAR

Abstract

The combined model electric quantity prediction method based on the CNN and the VAR comprises the steps of data collection and processing, electric quantity prediction of the current month and electric quantity prediction of the next month. In the data collection stage, basic data such as electric quantity, weather, economy and population are covered. The data processing relates to bad data correction, weather and week holiday data quantification and targeted data selection. The current month electric quantity prediction utilizes a CNN model, combines historical day electric quantity, weather and week holiday data to predict daily electric quantity from the middle month to the end of the month, and accumulates to obtain a monthly prediction value. And the next-month electric quantity prediction adopts a VAR model, historical monthly electric quantity and large user electric quantity data are input, and multi-dimensional single-step prediction is carried out. According to the method, the problems of long time span and inconsistent time scale in electric quantity prediction are effectively solved, the prediction precision and compatibility are remarkably improved, theoretical support and practical guidance are provided for stable operation of a power grid, and the method has important application value.

Description

Technical Field

[0001] The present invention belongs to the field of power grid planning and operation and power forecasting, and specifically is a power forecasting method based on a combined model of CNN and VAR. Background Art

[0002] In the power industry, accurate prediction of power demand is crucial for power production, dispatching and energy management. With the continuous development of society and the advancement of science and technology, the prediction of power demand has become more and more complicated, and traditional prediction methods can no longer meet actual needs. Therefore, researching and developing new power forecasting methods to improve the accuracy and reliability of forecasts has become an important topic in the power industry.

[0003] At present, the power industry mainly uses statistical models, regression models, artificial neural networks and other methods to forecast electricity. These methods can reflect the changing trend of electricity demand to a certain extent, but they often have certain limitations when dealing with complex electricity demand data. For example, the traditional ARIMA model is usually a unit multi-step forecast, which has certain requirements for the stability of electricity data and cannot cope with abnormal fluctuations in electricity caused by external factors. Therefore, the results of long-term forecasting of electricity using the ARIMA model are usually uncertain.

[0004] Moreover, the electricity forecast in many regions needs to provide electricity forecast data for the current month and the next month in the middle of each month, which requires both the prediction of the remaining daily electricity of the current month and the prediction of the monthly electricity of the next month. Obviously, there are problems such as a large prediction time span and inconsistent prediction time scale. The current electricity forecasting scheme does not provide a corresponding solution to this problem.

[0005] In order to solve the above problems, people have been seeking an ideal technical solution. Summary of the invention

[0006] To solve the above technical problems, the present invention proposes a combined model electricity prediction method based on CNN and VAR, which cleverly combines the advantages of convolutional neural network (CNN) and vector autoregression model (VAR), aiming to achieve accurate short-term and long-term prediction of electricity at different prediction time scales.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] The power forecasting method based on the combined model of CNN and VAR is characterized by comprising the following steps:

[0009] 1) Data collection and processing:

[0010] 2) Data selection:

[0011] Select relevant data for processing and analysis based on the forecast target. For daily electricity forecast, focus on analyzing the impact of weather data and weekend data. For monthly electricity forecast, perform correlation analysis and select data with higher correlation with monthly electricity data as input.

[0012] 3) Electricity forecast for the current month:

[0013] Use the CNN model to learn historical daily electricity data and identify patterns and trends in the data; at the same time, combine weather data and weekend data to provide more comprehensive information support; input the processed data into the CNN model to predict daily electricity data from the middle of the month to the end of the month; add the predicted daily electricity data to the actual electricity data to obtain the monthly electricity forecast data for the entire month;

[0014] 4) Electricity forecast for next month:

[0015] Use VAR model for multi-dimensional single-step forecasting; perform time series test on processed data to determine its correlation with time series and dynamically select parameters of VAR model; use historical monthly electricity data and large user electricity data as input to predict monthly electricity data for next month through VAR model; extract data with correlation exceeding set threshold in the current quarter as input data through correlation analysis to improve the accuracy of monthly electricity forecasting;

[0016] 5) Result analysis steps;

[0017] The absolute value error is used as the error calculation method, and the model parameters are adjusted by analyzing the prediction result error to further improve the model accuracy.

[0018] As a further improvement of the present invention, the data collection and processing in step 1) are specifically as follows:

[0019] a) Collect basic data, which mainly includes electricity data, weather data, economic and population data and other relevant data; economic and population data are used to analyze the long-term trend of electricity demand, and other data may include electricity price policy and power network structure information;

[0020] b) Data processing: Bad data correction: Check the collected power data, identify and correct missing power data, zero power data, continuous constant power data and abnormal step power data to ensure the integrity and accuracy of the power sequence; Quantify weather data: Simplify weather conditions into four conditions: sunny, cloudy, rainy and snowy, and quantify them as 1, 2, 3 and 4 respectively; Quantify weekday and holiday data: Quantify week types into 1-7, and assign special quantized values ​​to holidays and specific days before and after holidays to distinguish the different effects of weekdays and weekends, holidays and the first day after holidays on power consumption;

[0021] As a further improvement of the present invention, the CNN model adopts a multi-dimensional non-rolling multi-step prediction strategy when predicting daily electricity data.

[0022] As a further improvement of the present invention, when predicting monthly electricity data, the VAR model can dynamically select model parameters, such as lag order, etc., to improve the robustness and adaptability of the model.

[0023] As a further improvement of the present invention, the corresponding system of the combined model electricity prediction method based on CNN and VAR includes a data collection module, a data processing module, a current month electricity prediction module, a next month electricity prediction module and a result analysis module.

[0024] Beneficial effects: Compared with the prior art, the present invention adopts the above technical solution and has the following advantages:

[0025] (1) The present invention can more accurately predict electricity consumption by combining the advantages of the CNN model and the VAR model. The CNN model is good at processing data with spatial structure and can identify patterns and trends in historical daily electricity consumption data; while the VAR model is a time series analysis tool that can model the dynamic relationship between multiple related variables and is suitable for multi-dimensional single-step prediction of monthly electricity consumption. This combined model effectively solves the problems of long prediction time span and inconsistent prediction time scale in electricity consumption forecasting in some regions, and significantly improves the accuracy of electricity consumption forecasting.

[0026] (2) The method of the present invention not only considers the electricity data itself, but also integrates various related information such as weather data, weekend data, economic and population data, thereby improving the compatibility of the model. Through the comprehensive analysis and utilization of these data, the model can better adapt to the characteristics of electricity demand in different regions, providing strong theoretical support and practical guidance for the stable operation of the power grid.

[0027] (3) The present invention improves the efficiency of electricity forecasting by optimizing steps such as data processing and model parameter selection. For example, in the data processing stage, bad data is corrected, weather data and weekend data are quantified, and relevant data is selected according to the forecast target, providing high-quality and targeted input for the model; in terms of model parameter selection, the VAR model can dynamically select model parameters, such as the lag order, to improve the robustness and adaptability of the model.

[0028] In summary, the present invention has the advantages of improving prediction accuracy, enhancing compatibility and improving prediction efficiency, and provides a new solution for electricity quantity prediction in the power industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1is a flow chart of the method of the present invention;

[0030] Figure 2 It is a basic data type graph;

[0031] Figure 3 This is the correlation analysis result graph of monthly data;

[0032] Figure 4 It is the time series test result chart of monthly data;

[0033] Figure 5 This is the result of the correlation analysis of quarterly data. DETAILED DESCRIPTION

[0034] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0035] A method for planning and configuring power grid energy storage taking into account a new power system security and stability domain, the method comprising the following steps:

[0036] 1) Data collection and processing:

[0037] a) Collect basic data, which mainly includes electricity data, weather data, economic and population data and other related data; economic and population data are used to analyze the long-term trend of electricity demand, and other data may include information such as electricity price policy and power network structure. Specific data types include Figure 2 shown.

[0038] b) Data processing: Bad data correction: Check the collected power data, identify and correct missing power data, zero power data, continuous constant power data and abnormal step power data to ensure the integrity and accuracy of the power sequence; Quantify weather data: Simplify weather conditions into four conditions: sunny, cloudy, rainy and snowy, and quantify them as 1, 2, 3 and 4 respectively; Quantify weekday and holiday data: Quantify week types into 1-7, and assign special quantized values ​​to holidays and specific days before and after holidays to distinguish the different effects of weekdays and weekends, holidays and the first day after holidays on power consumption;

[0039] 2) Data selection: Select relevant data for processing and analysis based on the prediction target. For daily power forecast, focus on analyzing the impact of weather data and weekend data. For monthly power forecast, perform correlation analysis and select data with a higher correlation with monthly power data as input.

[0040] 3) Electricity forecast for the current month:

[0041] The CNN (convolutional neural network) model is used to learn historical daily electricity data and identify patterns and trends in the data. At the same time, it is combined with weather data and weekend data to provide more comprehensive information support. The processed data is input into the CNN model to predict daily electricity data from the middle of the month to the end of the month. The predicted daily electricity data is added to the actual electricity data to obtain the monthly electricity forecast data for the entire month.

[0042] 4) Electricity forecast for next month:

[0043] The VAR (vector autoregression) model is used for multi-dimensional single-step forecasting; the processed data is subjected to a time series test to determine its correlation with the time series, and the parameters of the VAR model are dynamically selected; the historical monthly electricity data and large user electricity data are used as input, and the monthly electricity data for the next month is predicted through the VAR model; through correlation analysis, data with correlation exceeding the set threshold (such as 0.9) in the current quarter are extracted as input data to improve the accuracy of monthly electricity forecasting.

[0044] As a further improvement of the present invention, the step 1) fully collects data on power consumption. The basic data for power consumption forecast mainly includes four categories, namely power consumption data, weather data, economic and population data, and other data. These four categories of data include data from multiple sources.

[0045] There are three main steps to process the collected data: 1. Correction of bad data: Bad power data mainly includes vacant power data, zero power data, continuous constant power data, and abnormal step power data. If these bad data cannot be effectively processed, they will not be able to provide real and reliable information for modeling and prediction. Therefore, before analyzing the load characteristics and establishing the prediction model, these bad data need to be processed manually to ensure that the power sequence is complete and correct. When processing bad data, one principle needs to be followed: In order to ensure that the power data samples are not distorted, the power data should be modified less and the original power data should be used as much as possible to avoid unnecessary errors. 2. Quantification of weather data: Weather data includes weather conditions, maximum temperature, minimum temperature, and wind speed. First, simplify the weather conditions into four conditions: sunny, cloudy, rainy, and snowy, and quantize them into 1, 2, 3, and 4 respectively. Such simplification helps to reduce the complexity of the data while retaining the impact of weather conditions on power forecasting. 3. Quantification of weekday and holiday data: Week types and holidays have a great impact on electricity consumption. The load curve presents different characteristics on weekdays and weekends, holidays and the first day after holidays. Therefore, when establishing an electricity forecasting model, it is necessary to fully consider the week types and holidays that affect electricity consumption. Weekly data is quantified into 1-7, and statutory holidays are assigned special quantitative values. For example, we define holidays as 6, the day before the holiday as 5, and the day after the holiday as 1. This quantification method helps to distinguish the impact of different types of holidays on electricity demand.

[0046] As a further improvement of the present invention, in step 2), when establishing a CNN model for daily electricity forecasting, due to the short forecasting time span, it is not necessary to consider the impact of factors such as economic development level, urban and rural population, and electricity price policy. The focus is on analyzing the impact of weather factors such as temperature, humidity, and precipitation on the load and the changing rules of electricity with different daily attributes. At the same time, according to experience accumulation and statistics, the rotation of holidays and weekdays and weekends will also have a certain impact on the change of electricity. Therefore, the basic data required to establish a short-term electricity forecasting model only includes electricity data, weekly holiday data, and weather data.

[0047] At the same time, we conduct a correlation analysis between the predicted monthly electricity data and other monthly data available for input, so as to find out the monthly data with higher correlation with the monthly electricity data as input to improve the accuracy of the model. The correlation analysis heat map is shown in the figure below. Figure 3 shown.

[0048] As a further improvement of the present invention, the data processed in step 3) is input into the constructed CNN model for prediction. CNN (Convolutional Neural Network) is a deep learning algorithm that is particularly good at processing data with spatial structure. In the present invention, we use the CNN model to learn historical daily electricity data to identify patterns and trends in the data. At the same time, we also combine weather data and weekend holiday data to provide more comprehensive information support. In this way, we can predict daily electricity data from the middle of the month to the end of the month.

[0049] After obtaining the predicted daily electricity data, we add it to the actual data that has been generated to obtain the monthly electricity forecast data for the entire month. This calculation process provides power companies with a basis for understanding the overall electricity demand for the month in advance, which helps power companies make reasonable production and scheduling decisions.

[0050] As a further improvement of the present invention, in step 4), for the prediction of the total monthly electricity consumption of the next month, the present invention adopts a VAR model for multi-dimensional single-step prediction. The VAR (vector autoregression) model is a time series analysis tool for modeling the dynamic relationship between multiple related variables. In the present invention, we use historical monthly electricity consumption data and large user electricity consumption data as input, and use the VAR model to predict the monthly electricity consumption data of the next month.

[0051] We perform a time series test on the processed data to determine its correlation with the time series, so that we can dynamically select the parameters of the VAR model. The time series test of one column in the monthly data is as follows: Figure 4 shown.

[0052] Since different input data will have different time series correlations, in order to improve the robustness and adaptability of the model, when establishing the VAR model, we set it to dynamically select model parameters, such as the lag order, etc. This can help find a balance between overfitting and underfitting, thereby improving the prediction accuracy of the model on various types of data.

[0053] By analyzing the correlation between monthly electricity consumption and other related data in different quarters, we can specifically extract the main influencing factors of monthly electricity consumption in different quarters, and we can also remove data with low correlation to improve the accuracy of monthly electricity consumption forecast. Figure 5 shown.

[0054] By analyzing the results, a related program is designed to extract the data with correlation exceeding 0.9 in the current quarter as the input data for monthly electricity forecasting, while ignoring the data with correlation below 0.9.

[0055] The processed data is input into the constructed VAR model for prediction. The VAR model makes predictions by establishing an autoregressive relationship between variables. In the present invention, we use historical data to estimate model parameters, and then use these parameters to predict the monthly electricity data for the next month.

[0056] This calculation process provides power companies with a basis for understanding the overall electricity demand for the next month in advance, which helps power companies to make long-term power supply plans and resource allocations.

[0057] The method flow chart of the present invention is as follows Figure 1 shown.

[0058] This article predicts regional electricity consumption in September and October based on historical electricity consumption data and weather data.

[0059] According to the results of correlation analysis, the electricity consumption of the secondary industry is significantly correlated with Sunday, and the electricity consumption of the primary industry, tertiary industry and residents is sensitive to the meteorological conditions of the day. Based on this, we constructed a combined model electricity prediction method based on CNN and VAR proposed in this paper. The model is based on historical electricity data and combines variables such as weather conditions of the day, maximum and minimum temperatures, wind force level and weekly data as input. The core goal of this model is to achieve the prediction of electricity sales based on weather forecast values.

[0060] The electricity consumption forecast results for September are shown in the figure below.

[0061] The quantitative values ​​of weather forecast and historical electricity sales are shown in the following table:

[0062] date Electricity Sales Week Weather Weather Maximum temperature Minimum temperature Wind Electricity of the previous day 2024 / 9 / 17 43703460 2 3 2 26 22 3 53187164.43 2024 / 9 / 18 49726636 3 2 0 32 23 2 56137003.24 2024 / 9 / 19 51168988 4 2 0 32 25 2 59477966.83 2024 / 9 / 20 47214284 5 3 0 29 21 4 59610776.92 2024 / 9 / 21 46165104 6 3 0 25 21 2 53087028.87 2024 / 9 / 22 42462464 7 2 1 22 17 3 55418690.61 2024 / 9 / 23 44886276 1 2 0 23 17 3 58607187.82 2024 / 9 / 24 45911996 2 3 0 26 18 2 53731971.1 2024 / 9 / 25 45783490 3 1 3 27 19 2 54469200.85 2024 / 9 / 26 47283404 4 2 0 28 21 2 55202372.07 2024 / 9 / 27 48943880 5 3 0 27 23 2 56541899.12 2024 / 9 / 28 49399450 6 3 2 33 23 2 58068323.24 2024 / 9 / 29 46135320 7 1 3 33 24 2 48175841.39 2024 / 9 / 30 39227430 1 3 0 26 15 3 43546704.77

[0063] According to the calculation results, it is estimated that from September 16 to 30, the total power sales of the power grid will be 648.0122 million kWh. Taking the actual line loss rate (4.64%) in September as a reference, the power supply is 679.5430 million kWh. It is predicted that the power supply of the power grid in September will be 1551.8895 million kWh, and the power sales will be 1478.987 million kWh.

[0064] The electricity forecast results for October are as follows:

[0065] Among them, the VAR model is used with historical electricity data, predicted September electricity data and weather week data as input, and the October forecast value is 1,248,213,200 kWh. Taking the actual line loss rate (4.64%) in October 2023 as a reference, it is predicted that the power supply of the power grid in October will be 1,308,948,400 kWh, and the electricity sales will be 1,248,213,200 kWh.

[0066] In summary, the power supply of the power grid in September is predicted to be 155188950 kWh, and the power sales are 147898700 kWh. The power supply of the power grid in October is predicted to be 130894840 kWh, and the power sales are 1248213200 kWh.

[0067] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. The power forecasting method based on the combined model of CNN and VAR is characterized by: The following steps are involved: 1) Data collection and processing: 2) Data selection: Select relevant data for processing and analysis based on the forecast target. For daily electricity forecast, focus on analyzing the impact of weather data and weekend data. For monthly electricity forecast, perform correlation analysis and select data with higher correlation with monthly electricity data as input. 3) Electricity forecast for the current month: Use the CNN model to learn historical daily electricity data and identify patterns and trends in the data; at the same time, combine weather data and weekend data to provide more comprehensive information support; input the processed data into the CNN model to predict daily electricity data from the middle of the month to the end of the month; add the predicted daily electricity data to the actual electricity data to obtain the monthly electricity forecast data for the entire month; 4) Electricity forecast for next month: Use VAR model for multi-dimensional one-step forecasting; Perform time series inspection on the processed data to determine its correlation with the time series and dynamically select the parameters of the VAR model; use the historical monthly electricity data and large user electricity data as input to predict the monthly electricity data for the next month through the VAR model; extract data with correlation exceeding the set threshold in the current quarter as input data through correlation analysis to improve the accuracy of monthly electricity forecasting; 5) Result analysis steps; The absolute value error is used as the error calculation method, and the model parameters are adjusted by analyzing the prediction result error to further improve the model accuracy.

2. The power consumption prediction method based on the combined model of CNN and VAR according to claim 1 is characterized in that: The data collection and processing of step 1 are as follows: a) Collect basic data, which mainly includes electricity data, weather data, economic and population data and other relevant data; economic and population data are used to analyze the long-term trend of electricity demand, and other data may include electricity price policy and power network structure information; b) Data processing: Correction of bad data: Check the collected electricity data, identify and correct missing electricity data, zero electricity data, continuous constant electricity data and abnormal step electricity data to ensure the integrity and accuracy of the electricity sequence; Quantification of weather data: Simplify the weather conditions into four conditions: sunny, cloudy, rainy and snowy, and quantify them as 1, 2, 3 and 4 respectively; Quantification of week and holiday data: Quantify the week type as 1-7, and assign special quantification values ​​to holidays and specific days before and after holidays to distinguish the different effects of weekdays and weekends, holidays and the first day after holidays on electricity consumption.

3. The power consumption prediction method based on the combined model of CNN and VAR according to claim 1 is characterized in that: The CNN model adopts a multi-dimensional non-rolling multi-step prediction strategy when predicting daily electricity data.

4. The method for predicting power consumption based on the combined model of CNN and VAR according to claim 1 is characterized in that: The VAR model can dynamically select model parameters when predicting monthly electricity data.

5. The method for predicting power consumption based on the combined model of CNN and VAR according to claim 2 is characterized in that: The corresponding system of the combined model electricity forecasting method based on CNN and VAR includes a data collection module, a data processing module, a current month electricity forecasting module, a next month electricity forecasting module and a result analysis module.

Images