Power system climbing net load short-term prediction method
By combining Holt-Winters and ARIMA models to process net load data during power system ramp-up, the problems of high dependence on training data and inaccurate processing of multi-dimensional features in existing technologies are solved, resulting in more accurate predictions that can adapt to complex operating conditions.
Patent Information
- Application Number
- CN202511526943.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-24
- Publication Date
- 2026-02-13
AI Technical Summary
In existing technologies, power system ramp net load prediction models have difficulty handling nonlinear features and are highly dependent on training data, resulting in inaccurate prediction results, especially in the processing of multi-dimensional features.
We combined the Holt-Winters and ARIMA prediction models, and constructed a dynamic fusion model to improve prediction accuracy by preprocessing historical net load data, including outlier and missing value repair and feature extraction.
It improves the accuracy of training data and parameter precision of the prediction model, ensures data support for power system operation, adapts to complex operating conditions, and has good robustness.
Smart Images
Figure CN121524909A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a power system load forecasting method, and more particularly to a short-term forecasting method for power system ramp-up net load. Background Technology
[0002] In power systems, net load is used to balance the dispatch of power generation and demand, and to provide a basis for dispatching energy storage systems and distributed power sources. Therefore, it is necessary to accurately predict the net load of the power system, especially the ramp-up net load, which is an extremely important parameter for balancing the processing of new energy sources and load fluctuations.
[0003] In existing technologies, the prediction of net load during power system ramp-up generally adopts a single prediction model, such as a single Holt-Winters prediction model. Under a single model, it is difficult to handle the nonlinear characteristics in the net load data. Similarly, a single neural network, such as an RBF network or an LSTM network, can capture nonlinear characteristics, but it is too dependent on training data and requires a large number of samples to meet the requirements. Moreover, existing technologies are inaccurate in processing the multi-dimensional features of net load data, thus reducing the final prediction results.
[0004] Therefore, in order to solve the above-mentioned technical problems, it is urgent to propose a new technical approach. Summary of the Invention
[0005] In view of this, the purpose of this invention is to provide a short-term forecasting method for power system ramp-up net load. By preprocessing historical net load data, starting from time-period differences and trend change characteristics, the accuracy of the training data for the prediction model is effectively guaranteed, and the training accuracy of the prediction model parameters is improved. Moreover, by simultaneously processing and dynamically fusing the Holt-Winters prediction model and the ARIMA prediction model, the linear and nonlinear characteristics of the data are effectively obtained, thus effectively ensuring the accuracy of the final prediction results and providing accurate and reliable data support for the operation of the power system.
[0006] This invention provides a short-term forecasting method for net load ramping in power systems, comprising the following steps:
[0007] S1. Obtain historical net load data of the power system and preprocess the historical net load data;
[0008] S2. Construct the ARIMA prediction model and the Holt-Winters prediction model, and input the preprocessed historical net load data into the ARIMA prediction model and the Holt-Winters prediction model for training;
[0009] S3. Obtain the current power system load data, and input the current power system load data into the trained ARIMA prediction model and Holt-Winters prediction model to obtain the first prediction result and the second prediction result;
[0010] S4. Correct the first and second prediction results; and dynamically weight and sum the corrected first and second prediction results to obtain the final short-term net load prediction result of the power system.
[0011] Furthermore, in step S1, the preprocessing of historical net load data specifically includes:
[0012] The historical net load data is arranged into a feature vector according to the time series.
[0013] Detect and repair outliers in feature vectors;
[0014] Detect and repair missing values in the feature vector;
[0015] Feature extraction is performed on the repaired feature vector, and the extracted features are used as the final preprocessed data.
[0016] Furthermore, outlier detection and repair of feature vectors specifically includes:
[0017] Determine the historical net load mean and standard deviation;
[0018] For each data point, determine whether it meets the outlier criteria:
[0019] ;in: This represents the net load at time t. This represents the historical average net load. The standard deviation of historical net load;
[0020] If the outlier condition is met, then the net load at time t is... This is an outlier;
[0021] Using step size k as the sliding window, and the net load at time t... Using this as the central point, determine the mean of the seven historical net loads within the window from time t-3 to time t+3. :
[0022] ;
[0023] The average of historical net load Alternate outliers .
[0024] Furthermore, the missing value detection and repair of the feature vector specifically includes:
[0025] If the ratio of the actual data point to the set data point is less than the set value, it is determined that the data point in the current feature vector is missing. For missing data, the average value between the current data point and the next data point is used as the value of the missing point.
[0026] Furthermore, feature extraction from the repaired feature vector specifically includes:
[0027] Constructing the basic load feature vector:
[0028] ;
[0029] in: to The net load at the same time the previous day; to This represents the net load at the same time in the previous week;
[0030] Determine the time period attribute characteristics:
[0031] For the net load during a day and a week, the K-Means clustering algorithm is used to perform time-period clustering to form time-period attribute features;
[0032] Identify the characteristics of trend abrupt changes:
[0033] Given the net load in the previous week, with W as the window step size and W value of 48, calculate the linear trends of window 1 and window 2, where window 1 is t~t+W-1 and window 2 is t+W~t+2W-1.
[0034] ;
[0035] ;
[0036] in: , These are the net load averages for window 1 and window 2, respectively.
[0037] Calculate the trend difference rate:
[0038] ;
[0039] If the trend difference rate is greater than the set value, then t+W is the trend inflection point, and the net load of the trend inflection point is used as the trend inflection feature.
[0040] Furthermore, the specific revisions to the first and second prediction results include:
[0041] Determine the hourly correction factor:
[0042] ; The average of all historical load data over the past h hours. The total mean of all historical load data;
[0043] Determine the weekly correction factor:
[0044] ; for The average of all historical load data within the week;
[0045] Determine the correction factor for the specified time period:
[0046] ; This represents the average of all historical load data within the time period s;
[0047] The outputs of the ARIMA and Holt-Winters prediction models were corrected:
[0048] ;
[0049] ;
[0050] in: These are correction values for the output of the Holt-Winters prediction model. These are correction values for the output of the ARIMA prediction model; The first prediction result output by the Holt-Winters prediction model. This is the second prediction result output by the ARIMA prediction model.
[0051] Furthermore, the dynamic weights for dynamically weighted summing the corrected first and second prediction results are determined as follows:
[0052] When training the ARIMA and Holt-Winters prediction models, the preprocessed historical net load data is divided into training and validation sets.
[0053] Determine the root mean square error of the validation set;
[0054] The basic weights of the first prediction result are calculated based on the root mean square error of the validation set:
[0055] ;
[0056] The base weights of the second prediction result are calculated based on the root mean square error of the validation set:
[0057] ;
[0058] During the daytime period, the weight of the first prediction result will be... Increase by 10%;
[0059] During the nighttime period, the weight of the first prediction result is reduced by 10%.
[0060] The beneficial effects of this invention are as follows: By preprocessing historical net load data, starting from time-period differences and trend change characteristics, the accuracy of the training data for the prediction model is effectively guaranteed, and the training accuracy of the prediction model parameters is improved. Moreover, by simultaneously processing the Holt-Winters prediction model and the ARIMA prediction model and dynamically fusing them, the linear and nonlinear characteristics of the data are effectively obtained, and the accuracy of the final prediction results is effectively guaranteed, providing accurate and reliable data support for the operation of the power system.
[0061] Moreover, the present invention can adapt to complex working conditions and has good robustness. Attached Figure Description
[0062] The present invention will be further described below with reference to the accompanying drawings and embodiments:
[0063] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation
[0064] The present invention will be further described in detail below:
[0065] This invention provides a short-term forecasting method for net load ramping in power systems, comprising the following steps:
[0066] S1. Obtain historical net load data of the power system and preprocess the historical net load data;
[0067] S2. Construct the ARIMA prediction model and the Holt-Winters prediction model, and input the preprocessed historical net load data into the ARIMA prediction model and the Holt-Winters prediction model for training;
[0068] S3. Obtain the current power system load data, and input the current power system load data into the trained ARIMA prediction model and Holt-Winters prediction model to obtain the first prediction result and the second prediction result;
[0069] S4. The first and second prediction results are corrected; and the corrected first and second prediction results are dynamically weighted and summed to obtain the final short-term net load prediction result of the power system. By preprocessing historical net load data, starting from time-period differences and trend change characteristics, the accuracy of the training data of the prediction model is effectively guaranteed, and the training accuracy of the prediction model parameters is improved. Moreover, by simultaneously processing and dynamically fusing the Holt-Winters prediction model and the ARIMA prediction model, the linear and nonlinear characteristics of the data are effectively obtained, and the accuracy of the final prediction result is effectively guaranteed, providing accurate and reliable data support for the operation of the power system.
[0070] Moreover, the present invention can adapt to complex working conditions and has good robustness.
[0071] Both the ARIMA and Holt-Winters prediction models can use existing models. Specifically:
[0072] The Holt-Winters model fits the variation pattern of load using three components: "horizontal, trend, and seasonal." Its core principle is optimizing three smoothing coefficients. This ensures that the model can accurately capture intraday seasonal fluctuations.
[0073] The horizontal component representing the "baseline level" of the load. :
[0074]
[0075] in, The seasonal component S(t-24) from the same time the previous day is used to eliminate the seasonal fluctuations of the current load, resulting in a "de-seasonalized" baseline level; L(t-1)+T(t-1) is the superposition of the level and trend from the previous hour, representing the continuation of the historical baseline level. (0.05-0.95) is the horizontal smoothing coefficient, which controls the weight of current data and historical data. The larger the value, the stronger the impact of the current data.
[0076] The trend component representing the "long-term direction of change" of the load. :
[0077]
[0078] Where L(t)-L(t-1) is the difference between the current level and the level of the previous hour, reflecting short-term trend changes; (0.05-0.95) is the trend smoothing coefficient, which controls the weight of short-term changes and historical trends.
[0079] "Intraday periodic fluctuations" representing load :
[0080]
[0081] in, It is the ratio of the current load to the level of the previous hour, reflecting the current seasonal fluctuation; S(t-24) is the seasonal component at the same time the previous day, representing historical seasonal patterns; (0.05-0.95) is the seasonal smoothing coefficient, which controls the weight of current fluctuations and historical patterns.
[0082] Three smoothing coefficients The value of directly affects the accuracy of the model. If it is set based on experience (e.g.) This might not be suitable for specific load data. Therefore, a genetic algorithm is used to optimize the parameters: with the objective of minimizing the root mean square error (RMSE) of prediction, the optimal coefficients are searched within the range of 0.05-0.95. .
[0083] After obtaining the optimal parameters, calculate the predicted values for future times. Assuming the current time is t, the formula for predicting the load in the next k hours (where k is the number of prediction steps) is as follows:
[0084]
[0085] in, S(t+k) is the combined level of "level + trend" after k hours, and S(t+k) is the seasonal component after k hours (derived from historical seasonal patterns). Multiplying the two together gives the basic forecast value of Holt-Winters.
[0086] ARIMA model:
[0087] The ARIMA model processes non-stationary load sequences through three steps: autoregression (AR), differencing (I), and moving average (MA). The core is to optimize the order (p, d, q) to ensure that the model can effectively handle linear fluctuations in load.
[0088] (1) Core form of the model
[0089] To transform a non-stationary series (such as a long-term increase in load) into a stationary series (meeting the requirements of the ARIMA model), the original load series is... Perform d-order difference Let d be the difference order (usually between 0 and 2), then the mathematical form of the ARIMA model is:
[0090]
[0091] Where c is a constant term, representing the mean level of the differenced sequence; For the autoregressive (AR) part—the first p values of the differencing sequence. Predict the current value, where p is the AR order (0-3). The AR coefficient (reflecting the influence of the first i values); For the moving average (MA) part—using the first q error terms Revise the current prediction, where q is the order of MA (0-3). The MA coefficient (reflecting the degree of correction for the first j errors); The white noise error term represents random fluctuations that the model cannot explain (mean is 0, variance is constant).
[0092] (2) Order optimization and prediction
[0093] The choice of order (p, d, q) is crucial for the ARIMA model. Too high an order leads to overfitting (the model remembers noise), while too low an order leads to underfitting (it fails to capture patterns). Therefore, a "rolling window validation method" is used to optimize the order: historical data is divided into a training set (first 80%) and a validation set (last 20%); all combinations of (p=0-3), (d=0-2), and (q=0-3) are iterated through, and an ARIMA model is fitted to each combination, with the training set used to predict the validation set; the RMSE of each combination on the validation set is calculated, and the combination with the smallest RMSE is selected as the optimal order.
[0094] During forecasting, the load is decomposed into three parts: "trend, seasonality, and residual," which are forecasted separately and then superimposed.
[0095]
[0096] in The trend component after k hours is extracted using moving median smoothing (24-hour window) to reflect long-term changes; The seasonal component after k hours is extracted using the median of the period (24-hour period) to reflect intraday periodicity; The residual components after k hours are predicted using the optimized ARIMA model, correcting for fluctuations that were not captured by the trend and seasonal components.
[0097] In this embodiment, step S1, which involves preprocessing the historical net load data, specifically includes:
[0098] Generate a feature vector from historical net load data based on time series:
[0099] ;in The net load value for hour t (unit: MW) is given, with a time granularity of 1 hour, and the data integrity must meet the requirement of ≥95%.
[0100] Including time and hours Weekly attributes (0 = weekday, 1 = weekend), must be perfectly aligned with the net load data timestamp.
[0101] Detect and repair outliers in feature vectors;
[0102] Detect and repair missing values in the feature vector;
[0103] Feature extraction is performed on the repaired feature vector, and the extracted features are used as the final preprocessed data.
[0104] Specifically: outlier detection and repair of feature vectors includes:
[0105] Determine the historical net load mean and standard deviation;
[0106] For each data point, determine whether it meets the outlier criteria:
[0107] ;in: This represents the net load at time t. This represents the historical average net load. The standard deviation of historical net load;
[0108] If the outlier condition is met, then the net load at time t is... This is an outlier;
[0109] Using step size k as the sliding window, and the net load at time t... Using this as the central point, determine the mean of the seven historical net loads within the window from time t-3 to time t+3. :
[0110] ;
[0111] The average of historical net load Alternate outliers .
[0112] The specific steps of missing value detection and repair for feature vectors include:
[0113] If the ratio of the actual data point to the set data point is less than the set value, it is determined that a data point is missing in the current feature vector. For missing data, the average value between the current data point and the next data point is used as the value of the missing point. Specifically:
[0114] For missing points If it is known and Then the net load value at time t after completion :
[0115] The above methods can effectively ensure the accuracy of training data and provide reliable data support for the training of subsequent prediction models.
[0116] In this embodiment, feature extraction of the repaired feature vector specifically includes:
[0117] Constructing the basic load feature vector:
[0118] ;
[0119] in: to The net load at the same time the previous day; to This represents the net load at the same time in the previous week;
[0120] Determine the time period attribute characteristics:
[0121] For the net load within a day and a week, the K-Means clustering algorithm is used to perform time-period clustering to form time-period attribute features. The K-Means clustering algorithm is an existing technology. Through this step, the net load within a day and a week is divided into data of different time-period categories. Because the net load corresponding to different time-period categories within a day or a week is different, it reflects the load characteristics of different time-period attributes.
[0122] Identify the characteristics of trend abrupt changes:
[0123] Given the net load in the previous week, with W as the window step size and W value of 48, calculate the linear trends of window 1 and window 2, where window 1 is t~t+W-1 and window 2 is t+W~t+2W-1.
[0124] ;
[0125] ;
[0126] in: , These are the net load averages for window 1 and window 2, respectively.
[0127] Calculate the trend difference rate:
[0128] ;
[0129] If the trend difference rate is greater than a set value, then t+W is the trend inflection point, and the net load of the trend inflection point is used as the trend inflection feature. Through the above, the basic features, time-period attribute features, and trend inflection features are used as training inputs for the prediction model, thereby effectively ensuring the training accuracy of the prediction model.
[0130] In this embodiment, the correction of the first prediction result and the second prediction result specifically includes:
[0131] Determine the hourly correction factor:
[0132] ; The average of all historical load data over the past h hours. The total mean of all historical load data;
[0133] Determine the weekly correction factor:
[0134] ; for The average of all historical load data within the week;
[0135] Determine the correction factor for the specified time period:
[0136] ; This represents the average of all historical load data within the time period s;
[0137] The outputs of the ARIMA and Holt-Winters prediction models were corrected:
[0138] ;
[0139] ;
[0140] in: These are correction values for the output of the Holt-Winters prediction model. These are correction values for the output of the ARIMA prediction model; The first prediction result output by the Holt-Winters prediction model. This is the second prediction result output by the ARIMA prediction model. By correcting it using the methods described above, the accuracy of the final prediction result can be effectively ensured.
[0141] In this embodiment, the dynamic weights for dynamically weighted summation of the corrected first and second prediction results are determined as follows:
[0142] When training the ARIMA and Holt-Winters prediction models, the preprocessed historical net payload data is divided into training and validation sets; the preprocessing here includes outlier repair and missing value repair, but does not include the feature extraction process.
[0143] Determine the root mean square error of the validation set;
[0144] The basic weights of the first prediction result are calculated based on the root mean square error of the validation set:
[0145] ;
[0146] The base weights of the second prediction result are calculated based on the root mean square error of the validation set:
[0147] ;
[0148] During the daytime period, the weight of the first prediction result will be... Increase by 10%;
[0149] During the nighttime period, the weight of the first prediction result is reduced by 10%. By using the above method, the weight can be dynamically adjusted according to the different time periods during the fusion process to determine an accurate prediction result. This result also reflects the characteristics of the time period and provides accurate data support for subsequent power dispatch.
[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A short-term forecasting method for net load ramping in a power system, characterized in that: Includes the following steps: S1. Obtain historical net load data of the power system and preprocess the historical net load data; S2. Construct the ARIMA prediction model and the Holt-Winters prediction model, and input the preprocessed historical net load data into the ARIMA prediction model and the Holt-Winters prediction model for training; S3. Obtain the current power system load data, and input the current power system load data into the trained ARIMA prediction model and Holt-Winters prediction model to obtain the first prediction result and the second prediction result; S4. Correct the first and second prediction results; and dynamically weight and sum the corrected first and second prediction results to obtain the final short-term net load prediction result of the power system.
2. The short-term forecasting method for power system ramp-up net load according to claim 1, characterized in that: Step S1, the preprocessing of historical net load data specifically includes: The historical net load data is arranged into a feature vector according to the time series. Detect and repair outliers in feature vectors; Detect and repair missing values in the feature vector; Feature extraction is performed on the repaired feature vector, and the extracted features are used as the final preprocessed data.
3. The short-term forecasting method for power system ramp-up net load according to claim 2, characterized in that: Outlier detection and repair of feature vectors specifically includes: Determine the mean and standard deviation of historical net load; For each data point, determine whether it meets the outlier criteria: ;in: This represents the net load at time t. This represents the historical average net load. The standard deviation of historical net load; If the outlier condition is met, then the net load at time t is... This is an outlier; Using step size k as the sliding window, and the net load at time t... Using this as the central point, determine the mean of the seven historical net loads within the window from time t-3 to time t+3. : ； The average of historical net load Alternate outliers .
4. The short-term forecasting method for power system ramp-up net load according to claim 2, characterized in that: The specific steps of missing value detection and repair for feature vectors include: If the ratio of the actual data point to the set data point is less than the set value, it is determined that the data point in the current feature vector is missing. For missing data, the average value between the current data point and the next data point is used as the value of the missing point.
5. The short-term forecasting method for power system ramp-up net load according to claim 2, characterized in that: Feature extraction from the repaired feature vector specifically includes: Constructing the basic load feature vector: ； in: to The net load at the same time the previous day; to This represents the net load at the same time in the previous week; Determine the time period attribute characteristics: For the net load during a day and a week, the K-Means clustering algorithm is used to perform time-period clustering to form time-period attribute features; Identify the characteristics of trend abrupt changes: Given the net load in the previous week, with W as the window step size and W value of 48, calculate the linear trends of window 1 and window 2, where window 1 is t~t+W-1 and window 2 is t+W~t+2W-1. ； ； in: , These are the net load averages for window 1 and window 2, respectively. Calculate the trend difference rate: ； If the trend difference rate is greater than the set value, then t+W is the trend inflection point, and the net load of the trend inflection point is used as the trend inflection feature.
6. The short-term forecasting method for power system ramp-up net load according to claim 1, characterized in that: The specific corrections to the first and second prediction results include: Determine the hourly correction factor: ; The average of all historical load data over the past h hours. The total mean of all historical load data; Determine the weekly correction factor: ; for The average of all historical load data within the week; Determine the correction factor for the specified time period: ; This represents the average of all historical load data within the time period s; The outputs of the ARIMA and Holt-Winters prediction models were corrected: ； ； in: These are correction values for the output of the Holt-Winters prediction model. These are correction values for the output of the ARIMA prediction model; The first prediction result output by the Holt-Winters prediction model. This is the second prediction result output by the ARIMA prediction model.
7. The short-term forecasting method for power system ramp-up net load according to claim 1, characterized in that: The dynamic weights for dynamically weighted summation of the corrected first and second prediction results are determined as follows: When training the ARIMA and Holt-Winters prediction models, the preprocessed historical net load data is divided into training and validation sets. Determine the root mean square error of the validation set; The basic weights of the first prediction result are calculated based on the root mean square error of the validation set: ； The base weights of the second prediction result are calculated based on the root mean square error of the validation set: ； During the daytime period, the weight of the first prediction result will be... Increase by 10%; During the nighttime period, the weight of the first prediction result is reduced by 10%.