A method, system, device and medium for predicting uncertainties in power load
The original prediction interval is constructed by quantile regression network and corrected by time-segmented and side-segmented calibration. This solves the problem of prediction interval asymmetry caused by the globally uniform calibration amount in the existing technology and achieves a balance between the reliability and coverage of load prediction at different confidence levels.
Patent Information
- Application Number
- CN202611126734.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-28
- Publication Date
- 2026-11-10
AI Technical Summary
Existing load forecasting methods use a globally uniform calibration quantity, which cannot reflect the differences in boundary errors of load at different time periods and ignores the asymmetry of the left and right boundaries of the interval, resulting in insufficient coverage or excessive conservatism of the forecast interval in some time periods.
The original prediction interval is constructed by using a quantile regression network, and then the original prediction interval is corrected by time-segmented and side-segmented calibration to generate the final prediction interval, which reflects the difference in boundary error of load at different time periods and solves the problem of asymmetry between the left and right boundaries.
It achieves a balance between coverage reliability and range width under different confidence levels, providing a boundary reference for load fluctuation range for power system risk dispatch.
Smart Images

Figure CN123073706A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power technology, specifically to a method, system, device, and medium for predicting uncertainties in power load. Background Technology
[0002] With the large-scale grid connection of new energy sources and the continuous integration of new entities such as demand response, the power load exhibits stronger randomness, volatility, and nonlinearity, which places higher demands on the accuracy of load forecasting.
[0003] Currently, load forecasting methods based on statistical learning, traditional machine learning, and deep learning have been extensively studied. However, these methods are mainly based on deterministic point forecasting, outputting a single forecast value. They are difficult to quantify the uncertainty of the forecast results, cannot reflect the range of future load fluctuations, and cannot provide boundary references for power system risk dispatching.
[0004] Research has emerged that uses probability interval prediction to quantify the uncertainty of prediction results. However, existing studies use a globally uniform calibration quantity, which cannot reflect the differences in boundary errors of the load at different time periods. There is also the problem of ignoring the asymmetry of the left and right boundaries of the interval, resulting in insufficient coverage or excessive conservatism of the calibrated prediction interval at certain time periods, affecting the practicality and reliability of interval prediction. Summary of the Invention
[0005] In view of the above-mentioned problems, the present invention provides a method, system, device and medium for predicting power load uncertainty.
[0006] Therefore, the technical problem solved by this invention is: how to solve the problem that existing methods use a globally uniform calibration quantity, which cannot reflect the difference in boundary error of the load at different time periods, ignore the asymmetry of the left and right boundaries of the interval, and cause the calibrated prediction interval to be insufficiently covered or overly conservative in some time periods.
[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Acquire power load data, preprocess the data to obtain standardized data; A candidate feature set is constructed based on the standardized data, and the candidate feature set is filtered to obtain a feature subset; A time series sample is constructed based on the feature subset, and the time series sample is predicted. The prediction is used to obtain the original prediction interval. The original prediction interval is calibrated to obtain the final prediction interval.
[0008] As a preferred embodiment of the power load uncertainty prediction method described in this invention, the step of obtaining the original prediction interval includes: A preset time window length is used to divide the feature subset into multiple time-series samples according to the preset time window length; The time series samples are subjected to quantile regression to obtain a set of predicted values; Based on the predicted value set, construct the original prediction interval.
[0009] As a preferred embodiment of the power load uncertainty prediction method described in this invention, the step of obtaining the final prediction interval includes: Extract the median and geometric center of the original prediction interval, and weight the median and geometric center to obtain the interval center; Extract the left boundary of the original prediction interval, and obtain the first basic span based on the left boundary and the center of the interval; Extract the right boundary of the original prediction interval, and obtain the second basic span based on the right boundary and the center of the interval; Extract the actual load data corresponding to the time series sample, and obtain the first calibration value and the second calibration value based on the actual load data; The final prediction interval is generated based on the interval center, the first basic span, the second basic span, the first calibration value, and the second calibration value.
[0010] As a preferred embodiment of the power load uncertainty prediction method of the present invention, the steps of obtaining the first calibration quantity and the second calibration quantity include: A preset time segmentation rule is used to divide the actual load data into multiple time period subsets according to the preset time segmentation rule; The deviation of the actual load value in each time period subset from the first basic span is statistically analyzed to obtain the first initial calibration value corresponding to each time period. The first initial calibration value is then statistically processed to obtain the first calibration value. The deviation of the actual load value in each time period subset from the second basic span is statistically analyzed to obtain the second initial calibration value corresponding to each time period. The second initial calibration value is then statistically processed to obtain the second calibration value.
[0011] As a preferred embodiment of the power load uncertainty prediction method of the present invention, the step of generating the final prediction interval includes: A global scale parameter is preset, and the left correction span is obtained by summing the first basic span and the first calibration amount. The left correction span is adjusted based on the global scale parameter to obtain the final left offset. The final left offset is obtained by subtracting the final left offset from the center of the interval. A global scale parameter is preset, and the right correction span is obtained by summing the second basic span and the second calibration amount. The right correction span is adjusted based on the global scale parameter to obtain the final right offset. The right boundary of the final predicted interval is obtained by adding the final right offset to the interval center. The final prediction interval is formed by the left boundary and the right boundary of the final prediction interval.
[0012] This invention constructs an original prediction interval by directly outputting the predicted value set through a quantile regression network, and then corrects the original prediction interval by time-segmented and side-segmented calibration. The resulting final prediction interval can maintain a balance between coverage reliability and interval width at different confidence levels, providing a boundary reference for the load fluctuation range of power system risk dispatch.
[0013] As a preferred embodiment of the power load uncertainty prediction method described in this invention, the step of obtaining the feature subset includes: Extract the contribution degree of each candidate feature in the candidate feature set, and sort the candidate features in the candidate feature set according to the contribution degree to obtain the feature ranking result; In response to the filtering based on preset conditions, the feature subset is obtained from the feature ranking results.
[0014] By extracting the contribution level of each candidate feature, sorting and filtering them, a subset of features that meet the preset conditions is selected, reducing the interference of redundant variables.
[0015] As a preferred embodiment of the power load uncertainty prediction method described in this invention, the step of obtaining standardized data includes: The data is time-aligned to obtain time-synchronized data; The time synchronization data is repaired to obtain repaired data; The repaired data is standardized to obtain the standardized data.
[0016] By performing time alignment, patching, and standardization on the data, the quality and consistency of the input data were ensured.
[0017] Secondly, the present invention provides a power load uncertainty prediction system, comprising: The data processing module is used to acquire power load data, preprocess the data, and obtain standardized data. The feature filtering module is used to construct a candidate feature set based on the standardized data, and to filter the candidate feature set to obtain a feature subset; The interval prediction module is used to construct time series samples based on the feature subset, predict the time series samples, and obtain the original prediction interval; An interval calibration module is used to calibrate the original prediction interval to obtain the final prediction interval.
[0018] Thirdly, the present invention provides a computer device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the power load uncertainty prediction method.
[0019] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the power load uncertainty prediction method.
[0020] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention first constructs the interval center based on a weighted average of the median and geometric center, and then extracts the left and right boundaries of the original prediction interval to determine the first and second basic spans. This achieves independent processing of the left and right sides of the prediction interval, and can perform calibration separately for the asymmetry of the left and right boundaries, solving the problem of existing technologies ignoring the asymmetry of the left and right boundaries of the interval. This invention divides the actual load data into multiple time period subsets by pre-defined time segmentation rules, and statistically analyzes the deviation of the actual load value relative to the first and second basic spans in each time period to obtain the calibration amount for each time period. This can reflect the difference in boundary error of the load in different time periods, avoiding the problem of insufficient coverage or excessive conservatism in some time periods caused by a globally uniform calibration amount. Attached Figure Description
[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 The above is a flowchart of an overall method for predicting the uncertainty of power load, provided as an embodiment of the present invention.
[0023] Figure 2 This is a flowchart of a power load uncertainty prediction method provided in one embodiment of the present invention.
[0024] Figure 3To ensure that, at a 90% confidence level, the prediction interval output by the method of this invention follows the changes in the actual load curve.
[0025] Figure 4 To ensure that, at an 80% confidence level, the prediction interval output by the method of this invention follows the changes in the actual load curve.
[0026] Figure 5 To ensure that, at a 70% confidence level, the prediction interval output by the method of this invention follows the changes in the actual load curve. Detailed Implementation
[0027] To make the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.
[0028] Example 1, referring to Figures 1-3 This is one embodiment of the present invention, which provides a method for predicting the uncertainty of power load, including the following steps S1 to S4: S1. Obtain power load data, preprocess the data to obtain standardized data; Specifically, the power load data includes historical power load data and weather data, with the weather data including environmental data such as temperature, humidity, wind speed, and air pressure. Historical databases of the power system and measured records from meteorological monitoring stations can serve as data sources. The goal of data preprocessing is to provide temporally consistent, numerically continuous, and locally stable input data for subsequent interval prediction models. The raw data may contain missing values, outliers, or inconsistent timestamps from multiple data sources, making the prediction interval boundaries susceptible to noise interference. Therefore, unified preprocessing of the power load data is necessary.
[0029] The specific steps to obtain standardized data include S1.1 to S1.3: S1.1. Time-align the data to obtain time-synchronized data; Specifically, historical power load data and weather data are linked using a unified timestamp to ensure that the load data at each moment corresponds to the environmental data at the same time. Since power load data and meteorological data may come from different acquisition systems, the timestamps of the original data may be inaccurate. The system can use the hourly hour as a unified time reference to map the weather data within the same hour to the hourly time, thus completing timestamp alignment and obtaining time-synchronized data.
[0030] In this embodiment, the system acquires actual hourly electricity load data and corresponding meteorological measurement data for a certain region. The data collection timestamp for the load data is 10:00:00, and the data collection timestamp for the weather data is 10:00:15. The system uses the hourly hour as a unified timestamp, mapping the weather data from 10:00:00 to 10:59:59 to 10:00:00, ensuring that the load data at each moment corresponds to the environmental data at the same time, thus obtaining time-synchronized data.
[0031] S1.2, Repair the time synchronization data to obtain repaired data; Specifically, missing samples in the time-synchronized data are imputed using linear interpolation and neighboring values to avoid time series breaks caused by direct sample deletion. Outliers in the time-synchronized data are smoothed using a moving window method to reduce the interference of extreme samples on the subsequent quantile boundary learning process. Through missing value imputation and outlier smoothing, numerically continuous repaired data is obtained. The formula for calculating the linear interpolation of missing load data is as follows: in, For the next moment, For the previous moment, The load value at the next moment. This represents the load value at the previous moment.
[0032] In this embodiment, load data at 12:00 is missing from the time synchronization data. The system extracts the load values of 800MW at 11:00 and 850MW at 13:00, and uses linear interpolation to calculate (850-800) / (13-11) to obtain a load value of 825MW at 12:00. The load data at 15:00 is 1200MW, which deviates significantly from the 900MW values at the preceding and following times. The system uses a moving window method with a window length of 3 to calculate the average of 910MW at 14:00 and 16:00, and smoothly corrects the outlier at 15:00 to 910MW, thus obtaining the repaired data.
[0033] S1.3. Standardize the repaired data to obtain the standardized data; Specifically, before standardization, the repaired data is divided into training, calibration, and test sets according to the time exclusivity principle. The standardization process uses the mean and standard deviation of the training set as a benchmark, applying the same set of standardization parameters to the training, calibration, and test sets to avoid information leakage from the test set. The calculation formula for standardization is expressed as follows: in, Let be the value of the standardized i-th input variable at time t; The original data before standardization; Let be the mean of the i-th input variable on the training set; Let be the standard deviation of the i-th input variable on the training set; The input sequence sample corresponds to time t; It is the multidimensional input vector k-th time step forward from the prediction time t.
[0034] In this embodiment, for the i-th input variable in the repaired data, the mean is calculated to be 800MW and the standard deviation is 50MW on the training set. The original data at time t is 850MW. Substituting this into the standardization formula, the standardized value is calculated to be (850-800) / 50=1.0. The above operation is performed on all input variables to obtain the standardized data.
[0035] S2. Construct a candidate feature set based on the standardized data, and filter the candidate feature set to obtain a feature subset; Specifically, based on the standardized data, a candidate feature set is constructed around the sources of short-term load changes. The candidate features include three categories: historical dependency features, local statistical features, and temporal environmental features. Historical dependency features characterize the short-term inertia, daily cycle, and weekly cycle patterns of the load, including load values with lags of 1 hour, 2 hours, 3 hours, 24 hours, 48 hours, and 168 hours. Local statistical features describe the average level and fluctuation degree of the nearest time period, including the rolling mean and rolling standard deviation of short-term and long-term windows. Temporal environmental features introduce periodic patterns and external disturbance information, including hourly sine codes, hourly cosine codes, weekend identifiers, and weather variables such as temperature, humidity, wind speed, and air pressure, as well as their smoothing terms. The candidate feature set is constructed using these three types of features.
[0036] After constructing the candidate feature set, the candidate feature set is filtered. The filtering process removes features with low contribution from the candidate feature set, reducing the interference of redundant variables in subsequent processing. Through this filtering, the final feature subset is obtained.
[0037] The specific steps for obtaining the feature subset include S2.1~S2.2: S2.1 Extract the contribution degree of each candidate feature in the candidate feature set, and sort the candidate features in the candidate feature set according to the contribution degree to obtain the feature sorting result; Specifically, in this embodiment, an extreme gradient boosting tree is used as the ensemble learning model based on gradient boosting decision trees to iteratively train the candidate feature set. During model training, each tree node split leads to a decrease in the objective function, and the decrease in the objective function due to node split, i.e., the split gain, is taken as the contribution of the corresponding candidate feature. By accumulating and statistically analyzing the split gain of each candidate feature in all tree node splits, the contribution of each candidate feature is obtained. Subsequently, the candidate features are sorted in descending order of contribution to obtain the feature ranking result.
[0038] In this embodiment, the system acquires a candidate feature set containing historical dependency features, local statistical features, and temporal environment features. After iterative training and normalization using an extreme gradient boosting tree model, the contribution level of each feature is extracted. Local statistical features: the feature contribution of the rolling average over the past 3 hours is 0.5073; temporal environment features: the feature contribution of hourly cosine encoding is 0.1144; temporal environment features: the feature contribution of hourly sine encoding is 0.0266; historical dependency features: the feature contribution of features lagged by 1 hour is 0.2635. The features are then sorted from largest to smallest contribution level to obtain the feature ranking result.
[0039] S2.2 In response to the filtering based on preset conditions, the feature subset is obtained from the feature sorting results.
[0040] Specifically, the preset conditions are that the contribution level meets a preset threshold or the ranking position meets a preset ranking range. Candidate features that meet the preset conditions are extracted from the feature ranking results, and these extracted candidate features are combined into the feature subset. This step removes redundant features with low contribution levels through screening.
[0041] In this embodiment, the preset condition is set to select the top 5 candidate features in terms of contribution from the feature ranking results. Based on the feature ranking results obtained in S2.1, the system extracts the top 5 features, including the 3-hour rolling average, 1-hour lag, hourly cosine code, hourly sine code, and time sequence number. These 5 features are combined into the feature subset for use in the construction of subsequent time-series samples.
[0042] In an optional implementation, the screening can also employ a feature importance evaluation mechanism based on random forests. Specifically, a random forest model is constructed by building multiple decision trees, and the reduction in Gini impurity brought about by node splitting for each candidate feature is calculated using out-of-bag data, which is then used as the corresponding contribution level. Subsequently, the candidate features are sorted in descending order of contribution level to obtain the feature ranking results, and a subset of candidate features that meet the preset ranking range is selected to reduce the interference of redundant variables on subsequent time series modeling.
[0043] S3. Construct time series samples based on the feature subset, predict the time series samples, and obtain the original prediction interval; Specifically, time-series samples are constructed using a sliding time window approach based on the aforementioned feature subset. These constructed time-series samples are then input into a quantile regression network to learn conditional quantiles at different quantiles, outputting predicted values for multiple quantiles. These predicted values constitute the original prediction interval.
[0044] The specific steps to obtain the original prediction interval include S3.1 to S3.3: S3.1. A preset time window length is used to divide the feature subset into multiple time-series samples according to the preset time window length; Specifically, a sliding time window method is used to reconstruct the sequence of the feature subset. A fixed time window length is set, and feature subset data from multiple consecutive time points are extracted backward from the current time point as input time-series samples. The actual load value of the next time point is used as the corresponding prediction target. Multiple time-series samples are obtained by traversing the entire dataset through the sliding window.
[0045] In this embodiment, the preset time window length is set to 24 hours. The system uses time t as the prediction target time and backtracks to extract a subset of feature data from time t-23 to time t, forming a time series sample containing 24 time steps. The corresponding prediction target is the power load value at time t+1. Multiple time series samples are obtained by traversing through a sliding window with a step size of 1.
[0046] S3.2 Perform quantile regression processing on the time series samples to obtain a predicted value set; Specifically, in response to the acquisition of the time-series samples, an initial lower boundary value, an initial center reference value, and an initial upper boundary value are obtained through a first bidirectional time-series network. The first bidirectional time-series network is a bidirectional long short-term memory network, and its basic unit is a long short-term memory network. The long short-term memory network selectively retains useful information and suppresses irrelevant information in long-term sequences through a forget gate, an input gate, an output gate, and a memory state update mechanism. Specifically, the forget gate determines the information discarded at the previous time step; the input gate determines the information newly added to the memory state; and the output gate determines the output information of the memory state. Furthermore, the first bidirectional temporal network employs a bidirectional propagation mechanism. The forward network updates the state gradually from the start point to the end point of the window, learning the evolutionary pattern from far to near; the reverse network updates in reverse from the end point to the start point of the window, learning the reverse dependencies from near to far. After fusing the forward and reverse hidden states, a bidirectional output is obtained.
[0047] Within the quantile regression framework, the first bidirectional temporal network predicts values for three target quantiles: 0.05, 0.50, and 0.95. The bidirectional hidden states at the final time step are mapped to the predicted values for the corresponding quantiles via a fully connected layer, serving as the initial lower boundary value, the initial center reference value, and the initial upper boundary value, respectively. Model training employs a quantile loss function, which assigns different weights to errors where the true value is higher or lower than the predicted quantile, enabling the network to learn the lower boundary, median, and upper boundary respectively.
[0048] The initial lower boundary value, the initial center reference value, and the initial upper boundary value are subjected to monotonicity correction to obtain the predicted value set. The monotonicity correction is used to ensure that the lower boundary value of the predicted value set is not higher than the center reference value, and the center reference value is not higher than the upper boundary value, so as to avoid quantile crossover and ensure that the original prediction interval has basic geometric rationality.
[0049] In this embodiment, the system predicts the quantiles of 0.05, 0.50, and 0.95 using the first bidirectional time-series network. Assume that for a time-series sample at a certain moment, the network's initial lower boundary value is 810MW, the initial center reference value is 850MW, and the initial upper boundary value is 830MW. Since the initial upper boundary value is lower than the initial center reference value, a quantile crossover occurs. The system performs the monotonicity correction, adjusting the initial upper boundary value to 860MW and the initial center reference value to 840MW, ensuring that the lower boundary value is not higher than the center reference value and the center reference value is not higher than the upper boundary value, thus obtaining the predicted value set.
[0050] S3.3 Construct the original prediction interval based on the predicted value set; Specifically, the lower boundary value, the center reference value, and the upper boundary value are extracted from the predicted value set. The lower boundary value and the upper boundary value are used as the left and right boundaries of the original prediction interval, respectively, and the center reference value is used as the center position reference of the original prediction interval. The original prediction interval is then constructed by combining these values.
[0051] In this embodiment, based on the predicted value set obtained in S3.2, the lower boundary value of 810MW, the center reference value of 840MW, and the upper boundary value of 860MW are extracted. The system constructs the original prediction interval as [810MW, 860MW], and uses the center reference value of 840MW as the center position reference of the original prediction interval for the interval center calculation in subsequent interval calibration operations.
[0052] In an optional implementation, prediction can also employ a quantile regression mechanism based on a temporal convolutional network. Specifically, in response to the acquisition of temporal samples, local temporal dependency features of the temporal samples are extracted through causal convolution and dilated convolution structures in the temporal convolutional network, and initial lower boundary values, initial center reference values, and initial upper boundary values are obtained by mapping them using fully connected layers. The model training uses a quantile loss function, enabling the network to learn conditional quantiles at different quantile points. Monotonicity corrections are performed on the initial lower boundary values, initial center reference values, and initial upper boundary values to ensure that the lower boundary value is not higher than the center reference value and the center reference value is not higher than the upper boundary value, thus obtaining the predicted value set.
[0053] Please refer to Figure 2 S4. The original prediction interval is calibrated to obtain the final prediction interval; Specifically, the original prediction interval, output by the quantile regression network, may have boundary biases, leading to a gap between the actual coverage and the target confidence level. This step introduces calibration based on real load data to asymmetrically correct the boundaries of the original prediction interval, ensuring that the final prediction interval maintains an appropriate width while meeting the target confidence level.
[0054] The specific steps to obtain the final prediction interval include S4.1~S4.5: S4.1 Extract the median and geometric center of the original prediction interval, and weight the median and geometric center to obtain the interval center; Specifically, the median obtained in S3 is extracted as the center reference value of the original prediction interval. The geometric center calculated using the left and right boundaries is then weighted and fused with the center reference value using a preset center fusion coefficient to obtain the interval center. The interval center reflects the center position of the original prediction interval and serves as the reference point for subsequent asymmetric calibration.
[0055] In this embodiment, the left boundary of the original prediction interval obtained by S3 is 810MW, the center reference value is 840MW, the right boundary is 860MW, the calculated geometric center is (810+860) / 2=835MW, and the center fusion coefficient is set to 0.5, then the interval center is 0.5×835+0.5×840=837.5MW.
[0056] S4.2 Extract the left boundary of the original prediction interval, and obtain the first basic span based on the left boundary and the center of the interval; Specifically, the left boundary of the original prediction interval is extracted, and the difference between the interval center and the left boundary is calculated. This difference is used as the first basic span. The first basic span is used to characterize the initial distance from the interval center to the original left boundary.
[0057] In this embodiment, the center of the interval is 837.5MW, the left boundary is 810MW, and the first basic span is calculated to be 837.5-810=27.5MW.
[0058] S4.3 Extract the right boundary of the original prediction interval, and obtain the second basic span based on the right boundary and the center of the interval; Specifically, the right boundary of the original prediction interval is extracted, and the difference between the right boundary and the center of the interval is calculated. This difference is used as the second basic span. The second basic span is used to characterize the initial distance from the center of the interval to the original right boundary.
[0059] In this embodiment, the right boundary is 860MW, the center of the interval is 837.5MW, and the calculated second basic span is 860-837.5=22.5MW.
[0060] S4.4 Extract the actual load data corresponding to the time series sample, and obtain the first calibration value and the second calibration value based on the actual load data; Specifically, the actual load data corresponding to the calibration set is extracted. The calibration set is a subset of data specifically used for estimating boundary deviations, partitioned during the data preprocessing stage according to the principle of time exclusivity. Based on the boundary deviations between the actual load data and the original prediction interval, the left-side deviation and right-side deviation are calculated respectively, and statistically analyzed according to time periods to obtain the first calibration quantity and the second calibration quantity. The first calibration quantity characterizes the degree to which the actual load data deviates to the left from the lower bound of the original prediction interval, and the second calibration quantity characterizes the degree to which the actual load data deviates to the right from the upper bound of the original prediction interval.
[0061] The specific steps for obtaining the first and second calibration values include S4.4.1 to S4.4.3: S4.4.1. Preset time segmentation rules: Divide the actual load data into multiple time period subsets according to the preset time segmentation rules; Specifically, a preset time segmentation rule divides a day into multiple consecutive time period subsets. This preset time segmentation rule can be set according to actual power grid dispatching needs. Based on the preset time segmentation rule, the actual load data in the calibration set and its corresponding original prediction intervals are divided into multiple time period subsets to allow for the separate statistical analysis of boundary error differences in different time periods.
[0062] In this embodiment, the preset time segmentation rule is set to divide a 24-hour day into four time period subsets, namely 0:00 to 6:00, 6:00 to 12:00, 12:00 to 18:00, and 18:00 to 24:00.
[0063] S4.4.2. Calculate the deviation of the actual load value in each time period subset from the first basic span to obtain the first initial calibration amount corresponding to each time period. Perform statistical processing on the first initial calibration amount to obtain the first calibration amount. Specifically, for each time period subset, the deviation of the actual load value from the left boundary of the original prediction interval is statistically analyzed, i.e., the deviation from the left boundary position determined by the first basic span, to obtain the first initial calibration value corresponding to each time period. If the actual load value is higher than or equal to the left boundary, the deviation is zero; if the actual load value is lower than the left boundary, the deviation is the difference between the left boundary and the actual load value. The first initial calibration value of all samples within the time period subset is then statistically processed using empirical quantile values to obtain the first calibration value.
[0064] In this embodiment, within the time period subset from 0:00 to 6:00, the actual load value of a certain calibration sample is 790MW, and the left boundary of its corresponding original prediction interval is 810MW. Since the actual load value is lower than the left boundary, the first initial calibration amount is calculated to be 810-790=20MW. If the actual load value of another sample is 820MW, which is higher than the left boundary of 810MW, then its first initial calibration amount is 0. The 90th percentile value of the first initial calibration amount set of all samples within this time period subset is calculated, resulting in a first calibration amount of 5.0MW for this time period.
[0065] S4.4.3. Calculate the deviation of the actual load value in each time period subset from the second basic span to obtain the second initial calibration value corresponding to each time period. Perform statistical processing on the second initial calibration value to obtain the second calibration value.
[0066] Specifically, for each time period subset, the deviation of the actual load value from the right boundary of the original prediction interval is statistically analyzed, i.e., the deviation from the right boundary position determined by the second basic span, to obtain the second initial calibration value corresponding to each time period. If the actual load value is lower than or equal to the right boundary, the deviation is zero; if the actual load value is higher than the right boundary, the deviation is the difference between the actual load value and the right boundary. The second initial calibration value of all samples within the time period subset is then statistically processed using empirical quantile values to obtain the second calibration value.
[0067] In this embodiment, within the time period subset from 12:00 to 18:00, the actual load value of a certain calibration sample is 880MW, and the right boundary of its corresponding original prediction interval is 860MW. Since the actual load value is higher than the right boundary, the calculated second initial calibration amount is 880-860=20MW. The 90th percentile value is calculated for the set of second initial calibration amounts for all samples within this time period subset, yielding a second calibration amount of 3.0MW for this time period.
[0068] S4.5. Generate the final prediction interval based on the interval center, the first basic span, the second basic span, the first calibration value, and the second calibration value; Specifically, the center of the interval is used as the benchmark. The first calibration value and the second calibration value corresponding to the time period are superimposed on the first basic span and the second basic span respectively to obtain the corrected left boundary and right boundary, thereby generating the final prediction interval.
[0069] The specific steps for generating the final prediction interval include S4.5.1 to S4.5.3: S4.5.1. Preset global scale parameters, obtain the left correction span by the sum of the first basic span and the first calibration amount, adjust the left correction span based on the global scale parameters to obtain the final left offset, and obtain the final prediction interval left boundary by subtracting the final left offset from the interval center. Specifically, the first base span is added to the first calibration value to obtain the left-side corrected span. The left-side corrected span is multiplied by the global scale parameter to obtain the final left-side offset. The final left-side offset is subtracted from the interval center to obtain the left boundary of the final predicted interval.
[0070] In this embodiment, the first basic span is 27.5MW, the first calibration amount is 5.0MW, the left correction span is 27.5+5.0=32.5MW, the global scale parameter is set to 1.0, the final left offset is 32.5MW, the center of the interval is 837.5MW, and the calculated left boundary of the final predicted interval is 837.5-32.5=805.0MW.
[0071] S4.5.2. Preset global scale parameters, obtain the right correction span by the sum of the second basic span and the second calibration amount, adjust the right correction span based on the global scale parameters to obtain the final right offset, and obtain the right boundary of the final prediction interval by adding the final right offset to the interval center. Specifically, the second basic span is added to the second calibration value to obtain the right-side corrected span. The right-side corrected span is multiplied by the global scale parameter to obtain the final right-side offset. The final right-side offset is then added to the interval center to obtain the right boundary of the final predicted interval.
[0072] In this embodiment, the second basic span is 22.5MW, the second calibration amount is 3.0MW, the right-side correction span is 22.5 + 3.0 = 25.5MW, and the global scale parameter is set to 1.0, so the final right-side offset is 25.5MW. The center of the interval is 837.5MW, and the calculated right boundary of the final predicted interval is 837.5 + 25.5 = 863.0MW.
[0073] S4.5.3. The final prediction interval is formed by the left boundary and the right boundary of the final prediction interval.
[0074] Specifically, the final prediction interval is constructed based on the calculated left and right boundaries of the final prediction interval. This final prediction interval is the final output result of the quantitative analysis and prediction of power load uncertainty, used to provide a risk boundary reference for power system dispatching and operation.
[0075] In this embodiment, the left boundary of the final prediction interval is 805.0MW and the right boundary of the final prediction interval is 863.0MW, forming the final prediction interval [805.0MW, 863.0MW].
[0076] The specific numerical values mentioned above are hypothetical data used to better explain the content and are not related to the actual demonstration below. Because the actual data fluctuates greatly, it is inconvenient to provide examples, so hypothetical data is used.
[0077] Please refer to Figures 3 to 5 To verify the effectiveness of the method of the present invention, simulation examples were designed using actual hourly electricity load data of a certain region as input data. The examples were set with three confidence levels: 90%, 80%, and 70%. The method of the present invention (…) Figures 3 to 5 The text is a mix of seemingly unrelated fragments and incomplete sentences, making it impossible to translate accurately. It appears to be a collection of snippets from various sources, possibly related to a network, a 2GBoost-BiLSTM-CQR, and bidirectional long short-term memory networks. Figures 3 to 5 The text is labeled as BiLSTM and convolutional neural networks. Figures 3 to 5 The text is labeled as CNN, gated neural network (CNN). Figures 3 to 5 Three comparative models (labeled as GRU) were compared. All models were trained and tested under the same dataset, the same time window length, the same training and testing partition, and the same confidence level.
[0078] like Figure 3As shown, at a 90% confidence level, the prediction interval output by the method of this invention can follow the peak and trough changes of the actual load curve well, and the interval boundary can be dynamically adjusted according to load changes. All actual load values are within the prediction interval. In contrast, although the bidirectional long short-term memory network model and the gated recurrent unit network model can also cover the actual load values, their prediction intervals are generally wider, with the gated recurrent unit network model showing more obvious conservatism. The convolutional neural network model has a relatively compact interval at this confidence level and can cover the actual load values within the current period. However, combined with the results at lower confidence levels, it can be seen that its coverage stability after interval shrinkage is relatively insufficient.
[0079] like Figure 4 As shown, at an 80% confidence level, the prediction interval of the method of this invention shrinks significantly compared to the 90% confidence level, while still effectively covering the vast majority of the actual load values, with only slight coverage deviations at individual moments, demonstrating good interval compactness and coverage reliability. Although the bidirectional long short-term memory network model maintains a high degree of coverage, its overall interval width is larger than that of the method of this invention, exhibiting a certain degree of over-conservatism. The gated recurrent unit network model has a wider prediction interval, but still fails to cover the actual load values at individual moments. The convolutional neural network model has a relatively small interval width, but continuous coverage gaps occur in some time periods, indicating that it is difficult to stably maintain coverage after interval shrinkage.
[0080] like Figure 5 As shown, at a 70% confidence level, the prediction interval of the method of this invention further shrinks, while the interval boundary can still be adjusted according to actual load changes, and it can cover the vast majority of the actual load values, with coverage deviations only occurring at a few moments. The interval widths of the bidirectional long short-term memory network model and the gated recurrent unit network model are still significantly larger than those of the method of this invention, exhibiting strong conservatism; although the convolutional neural network model has a narrower interval, the actual load values exceed the prediction interval significantly more often, resulting in a decrease in the continuity and stability of interval coverage. The results at the three confidence levels indicate that the method of this invention can gradually shrink the prediction interval as the target confidence level decreases, maintaining a relatively stable coordination between interval compactness and coverage reliability, avoiding the problems of excessive expansion of the prediction interval and a significant decrease in coverage performance after interval shrinkage.
[0081] Example 2, an embodiment of the present invention, provides a power load uncertainty prediction system based on the previous embodiment, comprising: The data processing module is used to acquire power load data, preprocess the data, and obtain standardized data. The feature filtering module is used to construct a candidate feature set based on the standardized data, and to filter the candidate feature set to obtain a feature subset; The interval prediction module is used to construct time series samples based on the feature subset, predict the time series samples, and obtain the original prediction interval; An interval calibration module is used to calibrate the original prediction interval to obtain the final prediction interval.
[0082] This embodiment also provides a computer device applicable to a power load uncertainty prediction method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the steps of the power load uncertainty prediction method proposed in the above embodiment.
[0083] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the steps of a power load uncertainty prediction method as described in the above embodiments.
[0084] The storage medium proposed in this embodiment and the method for predicting power load uncertainty proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.
[0085] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory, random access memory, flash memory, hard disk, or optical disk, and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.
[0086] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not 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 method for predicting uncertainties in power load, characterized in that, Includes the following steps: Acquire power load data, preprocess the data to obtain standardized data; A candidate feature set is constructed based on the standardized data, and the candidate feature set is filtered to obtain a feature subset; A time series sample is constructed based on the feature subset, and the time series sample is predicted. The prediction is used to obtain the original prediction interval. The original prediction interval is calibrated to obtain the final prediction interval.
2. The method for predicting power load uncertainty as described in claim 1, characterized in that, The steps to obtain the original prediction interval include: A preset time window length is used to divide the feature subset into multiple time-series samples according to the preset time window length; The time series samples are subjected to quantile regression to obtain a set of predicted values; Based on the predicted value set, construct the original prediction interval.
3. The method for predicting power load uncertainty as described in claim 1, characterized in that, The steps to obtain the final prediction interval include: Extract the median and geometric center of the original prediction interval, and weight the median and geometric center to obtain the interval center; Extract the left boundary of the original prediction interval, and obtain the first basic span based on the left boundary and the center of the interval; Extract the right boundary of the original prediction interval, and obtain the second basic span based on the right boundary and the center of the interval; Extract the actual load data corresponding to the time series sample, and obtain the first calibration value and the second calibration value based on the actual load data; The final prediction interval is generated based on the interval center, the first basic span, the second basic span, the first calibration value, and the second calibration value.
4. The method for predicting uncertainties in power load as described in claim 3, characterized in that, The steps to obtain the first calibration value and the second calibration value include: A preset time segmentation rule is used to divide the actual load data into multiple time period subsets according to the preset time segmentation rule; The deviation of the actual load value in each time period subset from the first basic span is statistically analyzed to obtain the first initial calibration value corresponding to each time period. The first initial calibration value is then statistically processed to obtain the first calibration value. The deviation of the actual load value in each time period subset from the second basic span is statistically analyzed to obtain the second initial calibration value corresponding to each time period. The second initial calibration value is then statistically processed to obtain the second calibration value.
5. The method for predicting power load uncertainty as described in claim 4, characterized in that, The steps for generating the final prediction interval include: A global scale parameter is preset, and the left correction span is obtained by summing the first basic span and the first calibration amount. The left correction span is adjusted based on the global scale parameter to obtain the final left offset. The final left offset is obtained by subtracting the final left offset from the center of the interval. A global scale parameter is preset, and the right correction span is obtained by the sum of the second basic span and the second calibration amount. The right correction span is adjusted based on the global scale parameter to obtain the final right offset. The right boundary of the final predicted interval is obtained by adding the final right offset to the interval center. The final prediction interval is formed by the left boundary and the right boundary of the final prediction interval.
6. The method for predicting power load uncertainty as described in claim 1, characterized in that, The steps to obtain a feature subset include: Extract the contribution degree of each candidate feature in the candidate feature set, and sort the candidate features in the candidate feature set according to the contribution degree to obtain the feature ranking result; In response to the filtering based on preset conditions, the feature subset is obtained from the feature ranking results.
7. The method for predicting power load uncertainty as described in claim 1, characterized in that, The steps to obtain standardized data include: The data is time-aligned to obtain time-synchronized data; The time synchronization data is repaired to obtain repaired data; The repaired data is standardized to obtain the standardized data.
8. A power load uncertainty prediction system, employing the power load uncertainty prediction method as described in any one of claims 1 to 7, characterized in that, include: The data processing module is used to acquire power load data, preprocess the data, and obtain standardized data. The feature filtering module is used to construct a candidate feature set based on the standardized data, and to filter the candidate feature set to obtain a feature subset; The interval prediction module is used to construct time series samples based on the feature subset, predict the time series samples, and obtain the original prediction interval; An interval calibration module is used to calibrate the original prediction interval to obtain the final prediction interval.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the power load uncertainty prediction method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the power load uncertainty prediction method according to any one of claims 1 to 7.
