Short-Term Electric Load Forecasting Method and System

Abstract

The present invention relates to the technical field of electric load forecasting, specifically to a short-term electric load forecasting method and system. Electric load data, meteorological data, and social activity data affecting the electric load are obtained and preprocessed to form a data set; an artificial intelligence model is used to obtain the daily average load value corresponding to the day to be predicted; an artificial experience method is used to determine the electric load reference curve for the same day to be predicted; the daily average load value in the electric load reference curve is calculated, and the ratio to the [specific value not provided in the original text] is used as the curve adjustment ratio to correct the load value corresponding to each time point in the electric load reference curve, and the corrected load value is used as the prediction result. By combining the AI algorithm and artificial experience, the load curve predicted by artificial experience is corrected, and the changing trend of the electric load can be better reflected.

Description

Technical Field

[0001] The present invention relates to the technical field of power load forecasting, and specifically to a short-term power load forecasting method and system. Background Technique

[0002] The statements in this part merely provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] Short-term power load forecasting refers to a method for forecasting the load changes of a power system in the next few hours or within the next day. Traditional manual forecasting methods rely on rich experience and can comprehensively consider various factors affecting load changes. With the increasing volatility and uncertainty of power loads during short-term forecasting, existing technologies have begun to use AI algorithms for short-term power load forecasting. AI algorithms demonstrate powerful capabilities in data processing and pattern recognition. AI algorithms are more process-oriented and consider more factors than manual forecasting, and the obtained predicted mean results are more credible. However, they are slightly insufficient in the shape of the load curve and still difficult to completely replace the unique experience accumulated by manual experience in complex system analysis and decision-making. Summary of the Invention

[0004] To solve the technical problems existing in the above background technique, the present invention provides a short-term power load forecasting method and system. Through the powerful data analysis ability of AI algorithms, the potential laws between power loads and various influencing factors (such as weather, holidays, economic activities, etc.) are deeply explored. At the same time, combined with manual practical experience and professional knowledge, the prediction results of AI algorithms are verified and optimized. By combining the wisdom of the human brain with the computing power of the computer, the "computerization" of human brain wisdom and more accurate and reliable short-term load forecasting are realized.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] The first aspect of the present invention provides a short-term power load forecasting method, including the following steps:

[0007] Obtain power load data, meteorological data, and social activity data affecting power loads, and form a data set after preprocessing;

[0008] The obtained data set uses a trained artificial intelligence model to obtain the daily average load value corresponding to the day to be predicted ;

[0009] The obtained data set uses the manual experience method to determine the power load reference curve for the same day to be predicted;

[0010] According to the power load values corresponding to all time points in the power load reference curve, calculate the daily average load value corresponding to the reference curve , so as to The ratio with is used as the curve adjustment ratio , and the obtained curve adjustment ratio is used to correct the load value corresponding to each time point in the power load reference curve, and the corrected load value is used as the prediction result.

[0011] Furthermore, the social activity data affecting the power load at least includes holiday information, production start-stop information, and electricity price adjustment information. The date type of the day to be predicted is determined to be at least one of a working day, a rest day, a make-up holiday, a major holiday, and a day affected by economic activities by using the social activity data affecting the power load.

[0012] Furthermore, the preprocessing includes at least one of missing data filling, abnormal data detection and processing, multi-source data fusion and cleaning, and filtering and standardization processing.

[0013] Furthermore, using the manual experience method to determine the power load reference curve of the day to be predicted includes the following steps:

[0014] Obtain power load data, meteorological data, and social activity data affecting the power load;

[0015] Determine the daily cumulative precipitation of the day to be predicted according to the meteorological data;

[0016] Determine the date type of the day to be predicted according to the social activity data;

[0017] When the daily cumulative precipitation of the day to be predicted is not greater than the set value X, select the day with the same date type as the day to be predicted and the maximum similarity of the perceived temperature from all the dates in the same month of the previous year of the day to be predicted and the first 30 days of the current year of the day to be predicted, and obtain the power load curve of the selected day as the first manual experience output curve;

[0018] When the daily cumulative precipitation of the day to be predicted is greater than the set value X, select the day with the same date type as the day to be predicted and the maximum similarity of the precipitation from the dates with the daily cumulative precipitation exceeding the set value X in all the months of the current year and the previous year of the day to be predicted, and obtain the power load curve of the selected day as the second manual experience output curve;

[0019] The obtained first manual experience output curve or second manual experience output curve is the power load reference curve of the day to be predicted.

[0020] Furthermore, the power load reference curve of the day to be predicted reflects the power load data every 15 minutes and has the power load values corresponding to 96 time points.

[0021] Further, the maximum similarity is specifically the minimum root mean square error.

[0022] Further, according to the power load values corresponding to all time points in the power load reference curve, calculate the daily average load value corresponding to the reference curve , as shown in the following formula:

[0023] ;

[0024] In the formula, N is the number of load points, which is 96 points; is the load at the time corresponding to the i-th point.

[0025] Further, take and The ratio of is used as the curve adjustment ratio , using the obtained curve adjustment ratio , correct the load values corresponding to each time point in the power load reference curve, and use the corrected load values as the prediction results, as shown in the following formula:

[0026] , ;

[0027] In the formula, is the daily average load value corresponding to the reference curve, is the daily average load value corresponding to the day to be predicted, is the load at the time corresponding to the i-th point in the reference curve.

[0028] Further, the artificial intelligence model uses historical power load data, historical meteorological data, and social activity data affecting power load as the training set, and through optimization and training, obtains the daily average load value corresponding to the day to be predicted .

[0029] The second aspect of the present invention provides a short-term power load prediction system, including:

[0030] A data acquisition and preprocessing module, configured to: obtain power load data, meteorological data, and social activity data affecting power load, and form a data set through preprocessing;

[0031] An AI prediction module, configured to: use the obtained data set and the trained artificial intelligence model to obtain the daily average load value corresponding to the day to be predicted ;

[0032] An artificial experience prediction module, configured to: use the obtained data set and the artificial experience method to determine the power load reference curve for the same day to be predicted;

[0033] The fusion processing module is configured to calculate the daily average load value corresponding to the reference curve based on the power load values corresponding to all time points in the power load reference curve. , taking the ratio with as the curve adjustment ratio , and using the obtained curve adjustment ratio , to correct the load value corresponding to each time point in the power load reference curve, and taking the corrected load value as the prediction result.

[0034] Compared with the prior art, the above one or more technical solutions have the following beneficial effects:

[0035] Considering that the mean result predicted by AI (artificial intelligence algorithm) is more credible, but the result is slightly insufficient in terms of curve shape, and the curve shape can reflect the change trend of power load, making the prediction result obtained by the AI algorithm have the characteristic of uneven distribution. This characteristic is mainly manifested as: the predicted load in different time periods deviates greatly from the actual load locally, and the local deviation is small. In this case, the overall deviation degree of the predicted load is large and the accuracy rate is relatively low. On the one hand, the manual experience method finds a reference curve that is closer to the day to be predicted, such as the same rest day, the same working day, or the date that is similarly affected by economic activities, or the similarity in precipitation and perceived temperature. On the other hand, it finds the influence of the load characteristics in different time periods on the accuracy rate of the whole network load to achieve load prediction, and has more advantages in the overall load change trend. Therefore, according to the ratio of the power load mean predicted by AI to the load mean of the manual experience reference curve, the amplitude fluctuation of the reference curve is adjusted, and the shape of the day to be predicted and the reference curve are regarded as the same. After adjusting the curve shape, the predicted load can become relatively more uniform (the 96-point predicted load basically coincides with or is parallel to the actual load). When the accuracy rate of the predicted average load value approaches the actual load mean, the adjusted predicted load curve will be almost coincident with the actual load (if there are no extreme weather or social events and other special circumstances affecting the day to be predicted), and the accuracy rate will be greatly improved; even if there are some deviations in the mean value, the two curves are still in a relatively parallel state, and there will be no extreme deviation in any time period of the day, and the accuracy rate will still be improved, which makes the load curve obtained by manual experience better reflect the change trend of power load after correction. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The specification drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0037] Figure 1 is a schematic diagram of the short-term power load prediction process provided by one or more embodiments of the present invention. Detailed implementation manners

[0038] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0039] It should be noted that the following detailed descriptions are all exemplary and are intended to provide further descriptions of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0040] The following embodiments provide a short-term power load forecasting method. Through the powerful data analysis ability of the AI algorithm, the potential laws between the power load and various influencing factors (such as weather, holidays, economic activities, etc.) are deeply explored. At the same time, combined with the practical experience and professional knowledge of humans, the forecasting results of human experience are verified and optimized. By combining the wisdom of the human brain with the computing power of the computer, the "computerization" of the human brain wisdom and more accurate and reliable short-term load forecasting are realized.

[0041] Embodiment 1:

[0042] A short-term power load forecasting method includes the following steps:

[0043] Obtain power load data, meteorological data, and social activity data affecting the power load, and form a data set through preprocessing;

[0044] The obtained data set uses the trained artificial intelligence model to obtain the daily average load value corresponding to the day to be predicted ;

[0045] The obtained data set uses the artificial experience method to determine the power load reference curve for the same day to be predicted;

[0046] According to the power load values corresponding to all time points in the power load reference curve, calculate the daily average load value corresponding to the reference curve , and use The ratio of as the curve adjustment ratio , and use the obtained curve adjustment ratio to correct the load value corresponding to each time point in the power load reference curve, and use the corrected load value as the prediction result.

[0047] As Figure 1 shown, the short-term power load forecasting method in this embodiment includes a part of obtaining the predicted load value through the AI algorithm and a part of obtaining the shape trend of the load curve through artificial experience. By combining the two parts, the predicted value obtained by the AI algorithm is corrected using artificial experience, and the corrected predicted value is used as the final output prediction result.

[0048] The process of an AI algorithm (artificial intelligence algorithm) obtaining the predicted load value includes the following steps:

[0049] Data preparation: Obtain data to form a data set. The data includes, but is not limited to, meteorological data, load data, production start / stop information, and economic activity information, etc., and is divided into a load data set and a meteorological data set according to the data type;

[0050] Preprocess the data, including filling missing data, detecting and processing abnormal data, fusing and cleaning multi-source data, and filtering and normalizing the data;

[0051] Use the preprocessed load data, meteorological data, and production start / stop and economic activity data as the input data for the AI algorithm;

[0052] When the AI algorithm obtains the predicted load value, the input data is segmented and processed to construct a training set and a test set. By constructing a load prediction model and optimizing and adjusting the model, the predicted output of the load sequence is obtained.

[0053] During the period when the load prediction model obtains the predicted load value, by performing a correlation analysis on the load data set and the meteorological data set, analyze the factors affecting load prediction, extract feature factors, perform numerical processing on the extracted feature factors, and select the required influencing factors according to the factor sizes of the load and different influencing factors. The same analysis can also be done on time factors such as seasons, date types, major holidays, etc. and policy factors such as time-of-use electricity prices and production start / stop to obtain the corresponding influencing factors.

[0054] In terms of data set segmentation, use the random segmentation method. Divide the data set into a training set, a test set, and a validation set according to a certain proportion for model training, and use the cross-validation technique to evaluate the model performance to obtain the preliminary prediction result of load prediction.

[0055] Implement hyperparameter tuning of the model. Continuously optimize the model and the feature set based on the feedback information. By increasing the number of training samples and making them closer to the characteristics of the date to be predicted and the self-learning mechanism of the algorithm, continuously improve the prediction ability of the model to achieve the optimization of the preliminary prediction of electric power load by artificial intelligence methods.

[0056] As Figure 1 shown, the part of manually predicting the shape trend of the load curve includes the following steps:

[0057] Obtain basic data. The resolution of the electric power load data and meteorological data in the basic data is 15 minutes. The social activity data in the basic data includes information such as major holidays, production start / stop, electricity price policy adjustments, and economic activities;

[0058] Determine the daily cumulative precipitation of the day to be predicted according to the meteorological data;

[0059] When the daily cumulative precipitation of the day to be predicted is not greater than the set value, determine the date type of the day to be predicted according to the social activity data. Among all the dates in the same month of the previous year as the day to be predicted and the first 30 days of the current year of the day to be predicted, select the day with the same date type as the day to be predicted and the maximum similarity of the perceived temperature. The obtained power load curve of the selected day is the first artificial experience output curve;

[0060] When the daily cumulative precipitation of the day to be predicted is greater than the set value, determine the date type of the day to be predicted according to the social activity data. Among the dates with the daily cumulative precipitation exceeding the same set value in all months of the current year and the previous year of the day to be predicted, select the day with the same date type as the day to be predicted and the maximum similarity of the precipitation. The obtained power load curve of the selected day is the second artificial experience output curve;

[0061] The obtained first artificial experience output curve or second artificial experience output curve is recorded as the reference curve. The reference curve reflects the power load data every 15 minutes, forming a 96-point power load reference curve.

[0062] In this embodiment, by obtaining power load data with a resolution of 15 minutes, cleaning and preprocessing the obtained load data, a load data set is constructed.

[0063] In this embodiment, by obtaining meteorological data such as temperature, humidity, wind speed, precipitation, air pressure, and irradiance and weather type that affect the load with a resolution of 15 minutes, cleaning and preprocessing the obtained meteorological data, a meteorological data set is constructed.

[0064] In this embodiment, by obtaining data such as major holidays, production start and stop, electricity price policy adjustment, and economic activities, a social activity data set is constructed.

[0065] According to the corresponding calculation principles, calculate the perceived temperature and daily cumulative precipitation participating in the load prediction respectively;

[0066] The calculation principle of the perceived temperature is shown in the following formula:

[0067] ;

[0068] In the formula, represents the temperature; represents the humidity; represents the perceived temperature.

[0069] The principle for calculating the daily cumulative precipitation is as follows: Obtain the measured meteorological data of a single meteorological station, and calculate the daily cumulative precipitation by dividing the 96-point cumulative measured precipitation by N (where N is the number of measured meteorological stations). The daily cumulative precipitation of the day to be predicted is obtained through prediction based on the already acquired meteorological data. The method for predicting precipitation is a common method in the meteorological field and will not be elaborated in this embodiment.

[0070] When multiple meteorological stations provide measured meteorological data, first determine the average precipitation data at each moment of the multiple meteorological stations, and then calculate the daily cumulative precipitation according to the 96-point cumulative precipitation divided by N.

[0071] Based on the obtained daily cumulative precipitation, determine the selection method of the reference curve for the day to be predicted;

[0072] The principle for determining the consistency of date types is as follows: The date type of the day to be predicted and the date type of the reference curve are the same for working days, rest days, adjusted days, major holidays, and days affected by economic activities;

[0073] When the daily cumulative precipitation is not greater than 10 mm, calculate the similarity of the perceived temperature between the day to be predicted and the dates in the selected interval of the reference curve in a way of similar perception. Adopt the principle of consistent date types, and determine the date with the maximum similarity (minimum root mean square error) and the same date type as the day to be predicted as the perceived temperature similar reference curve for the day to be predicted;

[0074] The selected interval for determining the perceived temperature similar reference curve is: all the dates of the same month in the previous year and the first 30 days before the day to be predicted in the current year.

[0075] In this embodiment, 10 mm is the principle selected by artificial experience.

[0076] When the daily cumulative precipitation is greater than 10 mm (medium rain or above rainfall level), calculate the similarity of the precipitation between the day to be predicted and the dates in the selected interval of the reference curve in a way of similar precipitation. Adopt the principle of consistent date types, and determine the date with the maximum similarity (minimum root mean square error) and the same date type as the day to be predicted as the precipitation similar reference curve for the day to be predicted;

[0077] The selected interval for determining the precipitation similar reference curve is: the dates before the day to be predicted and in the current year and all the months of the same season in the previous year where the daily cumulative precipitation exceeds 10 mm.

[0078] Obtain the 96-point electric load data of the perceived temperature similar reference curve or the precipitation similar reference curve for the day to be predicted, and record it as the reference curve; the obtained reference curve reflects the predicted load every 15 minutes on the prediction day. All the times in a day are divided into 96 data points, that is, the 96-point electric load of the reference curve.

[0079] Combining manual experience and AI technology to determine the short-term prediction result of power load, including the following steps:

[0080] Taking the reference curve obtained from manual experience as the benchmark curve for prediction, denoted as the load shape curve, and calculating the daily average load of the load shape curve , ;

[0081] In the formula, is the daily average load of the calculated shape curve; N is the number of points of the whole network load, which is 96 points; is the whole network load at the corresponding moment of the i-th point.

[0082] Predicting the daily average load value through an artificial intelligence algorithm (such as CNN-LSTM), denoted as the predicted load mean , and calculating the curve adjustment ratio as: ; that is, the ratio between the daily average load calculated using the manual experience reference curve and the daily average load predicted by the AI algorithm is used as the curve adjustment ratio.

[0083] Based on the 96-point whole network load of the reference curve and the curve adjustment ratio, calculating the adjusted predicted load as: ; that is, using the obtained adjustment ratio to correct the predicted load value of each point in the manual experience reference curve.

[0084] In this embodiment, according to the obtained power load data, the time period of a day is divided into: morning peak period (00:15 - 9:00), noon peak period (9:15 - 17:00), and evening peak period (17:15 - 24:00).

[0085] In this embodiment, the load characteristics are analyzed, including the analysis of load characteristics in different time periods within a day and the analysis of load characteristics in the same time period during the day of different date types.

[0086] The analysis of load characteristics in different time periods within a day is: drawing the load curves of all dates in the current season of the day to be predicted, and observing the change rules of the load at different time periods of the day.

[0087] The analysis of load characteristics in the same time period during the day of different date types is: drawing the load curves of all dates in the current season of the day to be predicted, and observing the change rules between the daily loads considering the date type.

[0088] Based on the analysis results of the load characteristics in different time periods by manual experience, the final predicted load of the day to be predicted is obtained, as shown in the following formula: .

[0089] The core idea of the above method is that the average load of the day to be predicted predicted by the artificial intelligence algorithm is equivalent to predicting the electricity consumption of the day to be predicted. This result comprehensively considers many factors (such as various meteorological conditions, time span, historical load, etc.). This method is more process-oriented and considers more factors than manual prediction, and is less affected by human subjective assumptions. Therefore, the predicted average value result of the artificial intelligence algorithm is more credible, but the result is slightly insufficient in terms of curve shape. This deficiency is mainly reflected in that there are different amplitude deviations between the actual load of 96 points of the whole network obtained at a resolution of 15 minutes and the predicted load in different time periods, resulting in the situation that the curve is locally close to and locally deviated from the actual load at the same time, making the coincidence degree between the predicted load curve shape and the actual load curve shape slightly lower, and it is difficult to reflect the change trend of the power load.

[0090] The key points of the manual experience method are, on the one hand, to find a reference curve that is closer to the day to be predicted. The approximation can be the approximation of the date type (for example, both are rest days, both are working days, or both are dates affected by economic activities), the similarity of precipitation, and the similarity of the perceived temperature. On the other hand, to find the influence of the load characteristics in different time periods on the accuracy of the whole network load.

[0091] According to the ratio of the predicted power generation of the artificial intelligence algorithm to the power generation (load average value) of the reference curve, the amplitude fluctuation of the reference curve is adjusted. At this time, the shapes of the day to be predicted and the reference curve are regarded as the same. This is mainly because the shape of the whole network load has periodicity, and it is affected by various factors such as time, season, and weather. By using the clustering analysis (K-Means) method to calculate the results of meteorological similarity, date type similarity, time span, etc. respectively, and selecting the load curve with the highest meteorological similarity and date type similarity and closest to the day to be predicted as the reference curve for the day to be predicted, its curve shape will also be closer to the load shape of the day to be predicted, further increasing the accuracy of the predicted load curve shape.

[0092] In addition, since the change rules of the daily load in each time period are different, combined with means such as manual experience summary and data mining, the load characteristics in different time periods are analyzed to further obtain the influence of the date type and the load of the first 1 / 3 days on the whole network load.

[0093] To sum up, the method of integrating manual experience and AI technology can not only use the artificial intelligence method to obtain the accurate electricity consumption (load) of the day to be predicted, but also obtain the shape of the load of the day to be predicted based on manual experience, and obtain a more accurate predicted load for the day to be predicted through the integration process.

[0094] In this embodiment, the obtained power load data is the whole network load data.

[0095] When performing data cleaning and preprocessing, the complexity of power load and meteorological data is often faced. These data features are characterized by high fluctuations, multiple noise points, and non-linear characteristics. The useful signals and noise signals are intertwined, resulting in uneven quality of load sample data, which directly affects the accuracy of load forecasting. In addition, affected by automation systems, data acquisition systems, etc., data missing and outliers caused by data mutations often appear in load and meteorological sample data, which masks the true change patterns of historical load and meteorology. Therefore, this embodiment adopts a data preprocessing strategy based on "cross-period comparison and time-series filling".

[0096] For missing data:

[0097] If a single value is missing, interpolation is performed by taking the average of the two values before and after the current data.

[0098] If 2 - 3 consecutive values are missing, linear interpolation is performed.

[0099] When 3 or more consecutive values are missing, the load data of the same date type and the same moment of the same day of last week of this date are used for substitution.

[0100] For abnormal data: Abnormal data correction: An unsupervised anomaly detection algorithm based on trees - Isolation Forest algorithm is adopted. Decision trees are constructed by randomly selecting features and values, and outliers are judged by comprehensively evaluating the anomaly scores of all data points and the scatter plots drawn, where the data points with higher anomaly scores and in the state of isolated points are identified as outliers.

[0101] The outliers in the dataset are identified and corrected by interpolating the missing data, so as to achieve the purpose of correction, thereby maintaining the consistency and reliability of the data.

[0102] The corrected data is verified. By drawing a box plot of the data to compare the distribution characteristics before and after data correction, it is ensured that all outliers have been correctly processed and the overall distribution of the data has not changed significantly.

[0103] Data filtering: Apply Savitzky - Golay filtering to the data. Savitzky - Golay filtering smooths the data by locally fitting polynomials to data points, removes noise and unnecessary fluctuations in the data, and makes the data smoother and easier to analyze.

[0104] Standardization processing: Given the significant scale differences among data features, scale unification measures need to be implemented in the data preprocessing stage. This process adopts the Min - Max Scaling method to adjust the scale of the data to the range of 0 to 1 to ensure the comparability and consistency of each feature in the subsequent analysis and comparison processes. The processing method of the standardized value is as follows:

[0105] Normalized value = (Original value - Minimum value) / (Maximum value - Minimum value).

[0106] Numerical conversion of non-numerical data: When dealing with non-numerical data, if information such as date type is encountered, it is converted into a label encoding format as an input feature during the model training process. This conversion method specifically maps the day-of-week information in the date (from Monday to Sunday) to the numbers 1 to 7 respectively, enabling the model to recognize and effectively utilize these date features for training.

[0107] Furthermore, it should be noted that the Spearman correlation coefficient analysis technique is adopted to explore the key factors affecting the performance of various load forecasting models, and feature factors are extracted accordingly. This process involves calculating the correlation coefficients between load characteristics and each potential influencing factor to identify the factors with the strongest influence on load changes.

[0108] Furthermore, it should be noted that a customized convolutional neural network (CNN) architecture is used, which incorporates parallel attention branches to efficiently extract key significant features. On this basis, combined with the capabilities of the long short-term memory network (LSTM), the temporal dependence relationships in the action sequence are modeled. By introducing the attention mechanism into the CNN-LSTM model, the fusion of features from coarse-grained to fine-grained is achieved, comprehensively and deeply capturing the dynamic characteristics of time-series data, and finally building a short-term load forecasting model for time series based on the attention mechanism of the CNN-LSTM method.

[0109] Furthermore, it should be noted that in the research of short-term load forecasting methods, the acquisition of artificial experience is crucial. The specific ways to obtain artificial experience are as follows: mainly through expert interviews in multiple places to deeply understand the characteristics of power markets in different regions and the load change rules; at the same time, designing questionnaires to widely collect insights and predictions on load change trends inside and outside the industry. In addition, data analysis and mining techniques are used to extract potential rules and trends that are difficult for experts to intuitively capture from historical load data.

[0110] Example 2:

[0111] A short-term power load forecasting system, including:

[0112] A data collection and preprocessing module, configured to: obtain power load data, meteorological data, and social activity data affecting power load, and form a data set through preprocessing;

[0113] An AI prediction module, configured to: obtain the daily average load value corresponding to the day to be predicted using the trained artificial intelligence model for the obtained data set ;

[0114] An artificial experience prediction module, configured to: use the artificial experience method for the obtained data set to determine the reference curve of the power load on the same day to be predicted;

[0115] A fusion processing module, configured to: calculate the daily average load value corresponding to the reference curve according to the power load values corresponding to all time points in the power load reference curve , taking the ratio with as the curve adjustment ratio , and use the obtained curve adjustment ratio to correct the load value corresponding to each time point in the power load reference curve, and use the corrected load value as the prediction result.

[0116] Through the powerful data analysis ability of the AI algorithm, deeply explore the potential laws between power load and various influencing factors (such as weather, holidays, economic activities, etc.), and at the same time combine the practical experience and professional knowledge of humans to verify and optimize the prediction results of artificial experience. By combining the wisdom of the human brain with the computing power of the computer, the "computerization" of the human brain's wisdom and more accurate and reliable short-term load prediction are realized.

[0117] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A short-term electric load forecasting method, characterized in that It includes the following steps: Obtain power load data, meteorological data, and social activity data that affect power load, and form a data set through preprocessing; For missing data, if a single value is missing, interpolation is performed by taking the average of the previous and next 2 values of the current data; if 2-3 values are continuously missing, linear interpolation is performed; when 3 or more consecutive values are missing, the load data at the same time of the same date type in the same week of the current date is used for substitution; For abnormal data, an unsupervised anomaly detection algorithm based on trees - Isolation Forest algorithm is used. Decision trees are constructed by randomly selecting features and values. Outliers are determined by comprehensively evaluating the anomaly scores of all data points and the plotted scatter plots, and the anomaly scores are relatively high and the data points are in an isolated state. The interpolation method for missing data is used for correction to maintain the consistency and reliability of the data; Verify the corrected data by plotting a data box plot to compare the distribution characteristics before and after data correction, ensuring that all outliers have been correctly processed and the overall distribution of the data has not changed significantly; Apply Savitzky-Golay filtering to the data. Savitzky-Golay filtering smooths the data by performing local polynomial fitting on data points, removing noise and unnecessary fluctuations in the data; The obtained dataset uses the trained artificial intelligence model to obtain the daily average load value P corresponding to the day to be predicted AIave ; Using the artificial experience method, determine the power load reference curve for the same day to be predicted from the obtained data set; According to the power load values corresponding to all time points in the power load reference curve, calculate the daily average load value P corresponding to the reference curve ave , taking P AIave and P ave ratio as the curve adjustment ratio ρ, using the obtained curve adjustment ratio ρ to correct the load value corresponding to each time point in the power load reference curve, and taking the corrected load value as the prediction result; Calculate the perceived temperature and daily cumulative precipitation involved in load forecasting respectively; The calculation principle of the perceived temperature is as shown in the following formula: In the formula, Tem represents temperature; Hum represents humidity; Wind represents the perceived temperature; The calculation principle of the daily cumulative precipitation is: Obtain the measured meteorological data of a single meteorological station, and calculate the daily cumulative precipitation according to the 96-point cumulative measured precipitation / N, where N is the number of measured meteorological stations; When multiple meteorological stations provide measured meteorological data, first determine the average precipitation data at each moment of the multiple meteorological stations, and then calculate the daily cumulative precipitation according to the 96-point cumulative precipitation / N; Using the artificial experience method to determine the power load reference curve for the day to be predicted includes the following steps: Obtain power load data, meteorological data, and social activity data that affect power load; Determine the daily cumulative precipitation for the day to be predicted according to the meteorological data; Determine the date type for the day to be predicted according to the social activity data; When the daily cumulative precipitation for the day to be predicted is not greater than the set value X, select the day with the same date type as the day to be predicted and the maximum similarity of perceived temperature from all the dates in the same month of the previous year of the day to be predicted and the first 30 days of the current year of the day to be predicted, and obtain the power load curve of the selected day as the first artificial experience output curve; When the daily cumulative precipitation for the day to be predicted is greater than the set value X, select the day with the same date type as the day to be predicted and the maximum similarity of precipitation from the dates with daily cumulative precipitation exceeding the set value X in all the months of the current year and the previous year of the day to be predicted, and obtain the power load curve of the selected day as the second artificial experience output curve; The obtained first artificial experience output curve or second artificial experience output curve is the power load reference curve for the day to be predicted.

2. The short-term power load forecasting method according to claim 1, wherein Social activity data that affects power load, including at least holiday information, production start / stop information, and electricity price adjustment information, is used to determine that the date type of the day to be predicted is at least one of a working day, a rest day, a compensatory leave day, a major holiday, and a day affected by economic activities.

3. The short-term electric load forecasting method according to claim 1, wherein Preprocessing includes at least one of missing data filling, abnormal data detection and processing, multi-source data fusion and cleaning, and filtering and standardization processing.

4. The short-term power load forecasting method according to claim 1, wherein The power load reference curve of the day to be predicted reflects power load data every 15 minutes and has power load values corresponding to 96 time points.

5. The short-term power load forecasting method according to claim 1, characterized in that The similarity is the largest, specifically: the root mean square error is the smallest.

6. The short-term power load forecasting method according to claim 1, characterized in that According to the power load values corresponding to all time points in the power load reference curve, calculate the daily average load value P corresponding to the reference curve ave , as shown in the following formula: Where N is the number of time points of the load; P i is the load corresponding to the i-th point at the corresponding moment.

7. The short-term electric load forecasting method according to claim 1, characterized in that Take P AIave The ratio with P ave is used as the curve adjustment ratio ρ. Using the obtained curve adjustment ratio ρ, the load value corresponding to each time point in the power load reference curve is corrected, and the corrected load value is used as the prediction result, as shown in the following formula: Where, P ave is the daily average load value corresponding to the reference curve, and P AIave is the daily average load value corresponding to the day to be predicted, and P i is the load at the time corresponding to the i-th point in the reference curve.

8. The short-term power load forecasting method according to claim 1, wherein, The artificial intelligence model uses historical power load data, historical meteorological data, and social activity data affecting power load as the training set. Through optimization and training, the daily average load value P corresponding to the day to be predicted is obtained. AIave .

9. A short-term power load forecasting system for implementing the method according to any one of claims 1-8, characterized in that, Including: A data acquisition and preprocessing module, configured to: obtain power load data, meteorological data, and social activity data that affects power load, and form a data set through preprocessing; The AI prediction module is configured to: obtain the daily average load value P corresponding to the day to be predicted by using the trained artificial intelligence model for the obtained data set AIave ; An artificial experience prediction module, configured to: use the artificial experience method for the obtained data set to determine the power load reference curve of the same day to be predicted; The fusion processing module is configured to calculate the daily average load value P corresponding to the reference curve based on the power load values corresponding to all time points in the power load reference curve. ave , taking P AIave and P ave as the curve adjustment ratio ρ, using the obtained curve adjustment ratio ρ to correct the load value corresponding to each time point in the power load reference curve, and taking the corrected load value as the prediction result.

Images