A load prediction method considering data enhancement under extreme weather

Abstract

The application discloses a kind of load prediction methods considering data enhancement under extreme weather, it is related to power system load prediction technical field, including: obtaining the meteorological data and historical industry load data of target area according to fixed sampling interval;The correlation between temperature sequence and load sequence in multivariate feature matrix is calculated, and the correlation intensity and stability under extreme scene are combined, and the temperature-sensitive target industry load set is screened;Daily extreme temperature is calculated according to temperature sequence, and different extreme weather days and corresponding similar day set are identified according to temperature threshold value determination;According to similar day set and representative test day, sample expansion and elimination are carried out, and basic training set and enhanced training set are formed;The benchmark model is constructed by two-stage training.The application compared with traditional method simply uses all load data, significantly reduces redundant features, improves model training efficiency and prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, and in particular to a load forecasting method that takes into account data enhancement under extreme weather conditions. Background Technology

[0002] In power systems, short-term load forecasting typically refers to estimating electricity demand for the next few hours to days. Accurate load forecasting plays a crucial supporting role in optimizing power system dispatch, balancing supply and demand, and efficiently allocating energy resources. Electricity load power is influenced by various factors, with some loads sensitive to weather conditions being particularly significantly affected by weather events.

[0003] With the deepening of the electrification of terminal energy, the proportion of weather-sensitive loads such as air conditioning in new power systems is increasing. At the same time, extreme weather events such as cold waves and high temperatures that significantly affect load demand occur frequently. However, the number of daily samples for such extreme weather events is usually small, making data acquisition difficult and samples scarce. As a result, traditional forecasting methods often lack sufficient training data when facing these special meteorological conditions, limiting the model's generalization ability and thus affecting the accuracy of the final forecast.

[0004] Specifically, existing load forecasting methods mostly focus on routine weather and lack targeted handling for extreme weather: on the one hand, they fail to effectively screen load types that are strongly correlated with weather, resulting in redundant input features for the models; on the other hand, the sample size for extreme weather is insufficient, which can easily lead to model overfitting; at the same time, the hyperparameters of traditional LSTM and Bi-LSTM models are mostly set based on experience, making it difficult to adapt to the changing characteristics of different load types.

[0005] Therefore, there is an urgent need for a load forecasting method that can combine weather classification, sample expansion, and hyperparameter optimization to improve the forecasting accuracy under extreme weather conditions. Summary of the Invention

[0006] Therefore, it is necessary to provide a load forecasting method that takes into account data enhancement under extreme weather conditions to address the aforementioned technical problems.

[0007] In a first aspect, the present invention provides a load forecasting method that takes into account data enhancement under extreme weather conditions, comprising:

[0008] Meteorological data and historical industry load data for the target area are acquired at fixed sampling intervals, and after anomaly detection and interpolation repair, a multivariate feature matrix is ​​generated.

[0009] Calculate the correlation between temperature and load sequences in the multivariate feature matrix, and combine the correlation strength and stability under extreme scenarios to screen the target industry load set that is sensitive to temperature.

[0010] The daily extreme temperature is calculated on a daily basis according to the temperature sequence. Different extreme weather days and the corresponding set of similar days are identified based on the temperature threshold. Representative test days are selected from the set of similar days.

[0011] Based on the set of similar days and representative test days, the samples are expanded and removed to form a basic training set and an enhanced training set; a baseline model is built through two-stage training, and the hyperparameters of the baseline model are optimized using the whale optimization algorithm to form a load prediction model;

[0012] The optimized load forecasting model is used to perform daily rolling forecasts on representative test days, and the load forecast values ​​of the corresponding target industry loads in the target industry load set at the next time point are output.

[0013] Furthermore, meteorological data and historical industry load data for the target area are acquired at fixed sampling intervals, and after anomaly detection and interpolation repair, a multivariate feature matrix is ​​generated, including:

[0014] Acquire meteorological data and corresponding historical industry load data for the target region, and align them to the same fixed sampling interval according to the timestamp to form the original multivariate time series;

[0015] The original multivariate time series is sampled for consistency verification. When duplicates, jump points or abnormal intervals are found, deduplication and alignment are performed, and the unaligned segments are marked as intervals to be repaired.

[0016] The Lagrange interpolation method is used to interpolate and repair the points to be repaired in the interval to be repaired. After repair, the continuity and boundary rationality checks are performed to obtain a continuous multivariate time series.

[0017] Various industry load sequences are extracted from multivariate time series, and the total load sequence is obtained by summing the multiple types of industry load sequences according to time. The industry load sequence, total load sequence, temperature sequence and humidity sequence are then spliced ​​together to form a multivariate feature matrix.

[0018] Furthermore, the correlation between temperature and load sequences in the multivariate feature matrix is ​​calculated, and the correlation strength and stability under extreme scenarios are combined to screen the set of temperature-sensitive target industry loads, including:

[0019] Obtain temperature series and industry load series of any type, and perform sliding windowing on temperature series and industry load series using a preset sliding window length; select temperature subsequence and load subsequence within any window, and calculate the Pearson correlation coefficient of the window;

[0020] Based on the pre-defined extreme scenario annotation results, the window set is divided into an extreme window set and a normal window set, and the correlation strength and stability of the temperature subsequence and the load subsequence within the extreme window are statistically analyzed according to the Pearson correlation coefficient.

[0021] Temperature sensitivity is determined based on preset dynamic correlation thresholds and stability thresholds. When the correlation strength is greater than or equal to the dynamic correlation threshold and the stability is less than or equal to the stability threshold, the current type of industry load sequence is determined to be temperature sensitive and included in the target industry load set.

[0022] Furthermore, the formula for correlation strength is:

[0023] ;

[0024] In the formula, This indicates the correlation strength between the temperature subsequence and the load subsequence within the extreme window; Represents an extreme set of windows; This represents the Pearson correlation coefficient within the window; Represents a window;

[0025] The formula for stability is:

[0026] ;

[0027] In the formula, This indicates the stability of the temperature subsequence and the load subsequence within the extreme window.

[0028] Furthermore, the daily extreme temperatures of the temperature series are calculated daily, and different extreme weather days and their corresponding sets of similar days are identified based on temperature thresholds. Representative test days are then selected from the sets of similar days, including:

[0029] The temperature sequence is divided into days to obtain a daily temperature sequence containing a fixed number of sampling points, and the daily extreme temperature, including the daily minimum temperature and the daily maximum temperature, is calculated.

[0030] The system sets conditions for determining cold wave days and high temperature days. When a target date meets the cold wave day condition, the corresponding date is marked as a cold wave day; when a target date meets the high temperature day condition, the corresponding date is marked as a high temperature day. The extreme weather days include both cold wave days and high temperature days. The cold wave day condition indicates that the temperature drop of the target date within a preset time range reaches a preset temperature drop threshold, and the lowest temperature of the target date is lower than a preset low temperature threshold. The high temperature day condition indicates that the highest temperature of the target date is higher than a preset high temperature threshold.

[0031] Dates not marked as cold wave days or high temperature days are labeled as regular weather days, and clustering is performed using the range of daily minimum temperature and daily maximum temperature as features to classify regular weather days into multiple weather types.

[0032] Select dates marked as cold wave days or high temperature days, and build sets of similar days for cold waves or high temperature days respectively. Select representative test days that meet the preset test conditions to construct a test set; the representative test days include cold wave test days and high temperature test days.

[0033] Furthermore, dates marked as cold wave days or high-temperature days are selected to form sets of similar cold wave days or sets of similar high-temperature days, respectively. Representative test days meeting preset conditions are then selected to construct a test set, including:

[0034] Dates marked as cold wave days are selected as the set of similar cold wave days;

[0035] Dates marked as high-temperature days are selected as the high-temperature similarity day set. When the high-temperature similarity day set is empty, dates that meet the condition that the daily maximum temperature is greater than the high-temperature threshold are selected to form the high-temperature similarity day set. When the high-temperature similarity day set is still empty, the dates are sorted from high to low according to the daily maximum temperature and a preset number of dates are selected as the high-temperature similarity day set.

[0036] Select cold wave test days that meet the preset conditions from the cold wave similar day set, and select the day with the highest daily maximum temperature from the high temperature similar day set as the high temperature test day; remove all sampling points corresponding to the cold wave test day and the high temperature test day from the training set and use them as the test set.

[0037] Furthermore, based on the set of similar days and representative test days, samples are expanded and removed to form a basic training set and an enhanced training set. A baseline model is constructed through two-stage training, and the hyperparameters of the baseline model are optimized using the whale optimization algorithm to form a load prediction model, including:

[0038] Based on the cold wave test days and high temperature test days, a mapping relationship between the daily index and the sample index is established to locate all sampling points corresponding to the two types of representative test days; the sampling points are removed from the original sample set, and the remaining samples after removing the representative test days are used to construct the basic training set;

[0039] The samples in the basic training set that correspond to the sets of cold wave-similar days and the sets of high temperature-similar days are labeled as extremely similar samples, and the remaining samples are labeled as non-extreme samples.

[0040] Load period data corresponding to cold waves and high temperatures are selected from multivariate time series to construct an extreme weather sample set; and by expanding, enhancing and standardizing the extreme weather sample set, the predicted load sequence of the target industry is output.

[0041] Set the sliding window length, and perform sliding windowing on the normalized time series in chronological order to generate supervised learning sample pairs;

[0042] The first stage of training of the baseline models of long short-term memory neural network and bidirectional long short-term memory neural network was carried out using the basic training set, and the key hyperparameters of the bidirectional long short-term memory neural network baseline model were optimized using the whale optimization algorithm to form a load prediction model.

[0043] Using extremely similar samples as the core, an enhanced training set is constructed by expanding and merging, and the enhanced training set is used to perform the second stage of training, thereby realizing the retraining and fine-tuning of the load prediction model.

[0044] Furthermore, the first stage of training for the baseline models of the Long Short-Term Memory Neural Network and the Bidirectional Long Short-Term Memory Neural Network is performed using the basic training set. The Whale Optimization Algorithm is then used to optimize the key hyperparameters of the Bidirectional Long Short-Term Memory Neural Network baseline model, resulting in a load prediction model including:

[0045] The supervised learning sample pairs are divided into a training subset and a validation subset according to a preset ratio, which serve as the first-stage training set and the first-stage validation set, respectively.

[0046] Using a multivariate sequence fragment composed of feature dimensions and window length as input, and the load prediction value of the target industry load at the next time moment as output, a long short-term memory neural network benchmark model and a bidirectional long short-term memory neural network benchmark model are constructed and trained.

[0047] Key hyperparameters affecting the predictive performance of the bidirectional long short-term memory neural network benchmark model are selected to form a parameter vector to be optimized, and the hyperparameter search range is set.

[0048] The fitness function and fitness index of the whale optimization algorithm are set, and the key hyperparameters are optimized through iterative optimization to generate the optimal parameters and obtain the optimized load prediction model.

[0049] Furthermore, the fitness function and fitness index of the whale optimization algorithm are defined, and key hyperparameters are optimized through iterative optimization to generate optimal parameters, resulting in the optimized load prediction model, including:

[0050] Feature matrices containing multiple types of component load sequences, total load sequences, temperature sequences, and humidity sequences are extracted from the original multivariate time series and normalized to eliminate the influence of dimensions.

[0051] The prediction error of the first-stage validation set is used as the fitness function of the whale optimization algorithm, and the root mean square error is used as the fitness index.

[0052] The key hyperparameters of the bidirectional long short-term memory neural network benchmark model are globally optimized using the whale optimization algorithm. Several hyperparameter combinations are initialized to form a whale population. The fitness value of the bidirectional long short-term memory neural network benchmark model corresponding to each hyperparameter group is calculated in the first stage validation set. The population position is iteratively updated by simulating the behavior of humpback whales, and the hyperparameter combinations are continuously optimized.

[0053] The optimal parameters of the bidirectional long short-term memory neural network baseline model are output after the preset number of iterations or the convergence condition is met, and a load prediction model is constructed based on the optimal parameters.

[0054] Furthermore, using extremely similar samples as the core, an enhanced training set is constructed by expanding and merging them. This enhanced training set is then used to perform a second stage of training, achieving retraining and fine-tuning of the load prediction model, including:

[0055] Using extremely similar samples as the core, synthetic minority class oversampling augmentation is performed on cold wave similar day samples and high temperature similar day samples respectively to generate cold wave synthetic samples and high temperature synthetic samples. These are then combined with the basic training set after removing representative test days to form an enhanced training set.

[0056] Normalization is performed on the augmented training set and parameters are recorded. Supervision sample pairs are constructed using the same window length as in the first stage and divided into a second-stage training set and a second-stage validation set according to a preset ratio. The long short-term memory neural network benchmark model, the bidirectional long short-term memory neural network benchmark model, and the load prediction model are retrained or fine-tuned on the augmented training set.

[0057] Secondly, the present invention also provides a load forecasting system that takes into account data enhancement under extreme weather conditions, the system comprising:

[0058] The fixed sampling module is used to acquire meteorological data and historical industry load data of the target area at fixed sampling intervals, and generate a multivariate feature matrix after anomaly detection and interpolation repair.

[0059] The load type filtering module is used to calculate the correlation between temperature series and load series in the multivariate feature matrix, and to filter the set of temperature-sensitive target industry loads by combining the correlation strength and stability under extreme scenarios.

[0060] The weather recognition module is used to calculate the daily extreme temperature of the temperature sequence, identify different extreme weather days and the corresponding set of similar days based on the temperature threshold, and select representative test days from the set of similar days.

[0061] The model building module is used to expand and eliminate samples based on the set of similar days and representative test days, forming a basic training set and an enhanced training set; a baseline model is built through two-stage training, and the hyperparameters of the baseline model are optimized using the whale optimization algorithm to form a load prediction model;

[0062] The rolling forecasting module is used to perform daily rolling forecasts on representative test days using the optimized load forecasting model, and outputs the load forecast value of the corresponding target industry load in the target industry load set at the next time moment.

[0063] The beneficial effects of this invention are as follows:

[0064] 1. Based on the Pearson correlation between temperature and load, only load types sensitive to extreme weather are retained, while features with weak correlation to cold waves or high temperatures are removed. Compared to traditional methods that simply use all load data, this invention significantly reduces redundant features and improves model training efficiency and prediction accuracy.

[0065] 2. By combining repeated sampling and the SMOTE algorithm to expand the cold wave and high temperature samples, and generating synthetic samples through nearest-neighbor linear interpolation, the diversity and representativeness of the samples are improved. Existing technologies often suffer from insufficient generalization ability of prediction models due to the scarcity of extreme weather samples. This invention effectively alleviates the sample sparsity problem and improves model stability.

[0066] 3. A Bi-LSTM deep neural network is introduced, which has bidirectional memory capability for time series data and can capture the complex temporal relationship between load and weather changes. The Whale Optimization (WOA) algorithm is used to globally optimize the model's hyperparameters, improving the model's predictive performance and convergence speed under extreme weather conditions. Attached Figure Description

[0067] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0068] Figure 1 This is a flowchart of a load forecasting method considering data enhancement under extreme weather conditions according to an embodiment of the present invention;

[0069] Figure 2 This is a diagram of an LSTM network structure according to an embodiment of the present invention;

[0070] Figure 3 This is a diagram of the Bi-LSTM network structure according to an embodiment of the present invention;

[0071] Figure 4 This is a data-enhanced load forecast under cold wave weather conditions according to an embodiment of the present invention;

[0072] Figure 5 This is a data-enhanced load forecast under cold wave weather according to an embodiment of the present invention;

[0073] Figure 6 This is a load forecast under high-temperature weather conditions before data enhancement, according to an embodiment of the present invention;

[0074] Figure 7 This is a data-enhanced load forecast under high-temperature weather according to an embodiment of the present invention;

[0075] Figure 8 This is an error distribution diagram under different weather conditions before and after high-temperature day data enhancement according to an embodiment of the present invention;

[0076] Figure 9 This is an error distribution map under different weather conditions before and after the cold wave data enhancement according to an embodiment of the present invention. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0078] Please see Figure 1 This paper provides a load forecasting method that considers data augmentation under extreme weather conditions, including:

[0079] S1. Obtain meteorological data and historical industry load data of the target area according to a fixed sampling interval, and generate a multivariate feature matrix after anomaly detection and interpolation repair.

[0080] Specifically, the system acquires one year of NWP meteorological data and corresponding industry load data for the target region, and aligns them at fixed sampling intervals to form the original multivariate time series; it then performs sampling consistency verification, deduplication, and missing data marking on the data.

[0081] To avoid misjudging sudden changes in actual load under extreme weather conditions as anomalies by a uniform cleaning rule, this invention adopts a scenario-adaptive anomaly detection and repair strategy: the 3σ rule is used for anomaly detection during normal weather periods, and the joint constraint of "amplitude change and temporal continuity" is used for anomaly determination during candidate extreme weather periods to distinguish between sensor noise and actual load response; Lagrange interpolation or neighborhood repair is used for missing values ​​or anomalies, and after repair, load non-negativity constraints and meteorological physical range constraints are executed to obtain a cleaned meteorological-load multivariate feature matrix, providing input data for subsequent screening, identification and training.

[0082] In the description of this invention, the step of acquiring meteorological data and historical industry load data of the target area according to a fixed sampling interval, and generating a multivariate feature matrix after anomaly detection and interpolation repair includes:

[0083] S11. Obtain meteorological data and corresponding historical industry load data for the target area, and align them to the same fixed sampling interval according to the timestamp to form the original multivariate time series.

[0084] Specifically, a one-year historical data table of the target region is obtained. This data table includes time segments, temperature segments, humidity segments, and various industry load segments, and is aligned to the same fixed sampling interval by timestamps, preferably 15 minutes. A sampling consistency check is performed on the timestamp sequence. If duplicates, jumps, or abnormal intervals are found, deduplication and alignment are performed, and segments that cannot be aligned are marked as intervals to be repaired. Finally, the various industry load sequences are summed by time intervals to obtain the total load sequence. The industry load series, total load series, temperature series, and humidity series are then concatenated column-wise to form a multivariate feature matrix. .

[0085] S12. Perform sampling consistency verification on the original multivariate time series. When there are duplicates, jump points or abnormal intervals, perform deduplication and alignment processing, and mark the unaligned segments as intervals to be repaired.

[0086] Specifically, the steps for outlier detection and missing value repair of original multivariate time series are as follows: For any sequence to be detected... Calculate the mean with standard deviation When satisfied Time determination These are outliers, and both outliers and missing points are marked as points to be repaired.

[0087] In a preferred embodiment, anomaly detection can be performed by calculating μ and σ on a "day-by-day / time-by-time" basis to reduce the impact of seasonality and daily cycles on the statistical threshold.

[0088] To make the data cleaning strategy more suitable for extreme weather scenarios, this invention adaptively adjusts the anomaly detection threshold according to the weather type of the sample: for samples of similar days with cold waves / high temperatures, a stricter anomaly detection threshold is adopted, such as tightening the 3σ threshold to 2.5σ or using a quantile threshold, to suppress the interference of abnormal spikes under extreme temperature shocks on model training; for samples of normal weather, the 3σ threshold is maintained to balance robustness and data fidelity.

[0089] S13. The Lagrange interpolation method is used to interpolate and repair the points to be repaired in the interval to be repaired, and after repair, the continuity and boundary rationality checks are performed to obtain a continuous multivariate time series.

[0090] Specifically, the Lagrange interpolation method is used to fill the points to be repaired, for those located at time [time value missing]. missing values Select before and after it Valid sampling points Construct the Lagrange interpolation polynomial:

[0091] ;

[0092] and with Replace missing or outlier values, where The optimal value is 3 to 5; after interpolation is completed, a boundary rationality check is performed on the repair sequence, including at least load non-negative constraints and temperature and humidity physical range constraints.

[0093] After interpolation repair is completed, continuity and boundary rationality checks (such as load non-negativity constraints and temperature and humidity physical range constraints) are performed on the repaired meteorological and industry load time series to obtain continuous and reliable meteorological-load multivariate time series data, which are then used as input data for subsequent correlation screening, extreme weather identification, SMOTE enhancement, and deep learning training.

[0094] During the repair process, piecewise Lagrange interpolation is preferred. When the outlier is located in the range of rapid load ramp-up / recession, the order of the interpolation polynomial is limited and neighborhood point constraints are added to avoid excessive smoothing that leads to the loss of peak and valley information.

[0095] S14. Extract various industry load sequences from the multivariate time series, sum the multiple types of industry load sequences according to time to obtain the total load sequence, and integrate the industry load sequence, total load sequence, temperature sequence and humidity sequence to form a multivariate feature matrix.

[0096] Specifically, multiple industry load series, total load series, temperature series, and humidity series are concatenated column-wise to form a feature matrix. This includes: Feature selection: determining the set of multivariate input features for prediction, which at least includes: Industry load vector: extracted from historical industry load data. Industry-specific load sequences form industry-specific load sequences The total load sequence is obtained by summing the loads over time. Its formula is:

[0097] ;

[0098] In the formula, Indicates the first The total load value corresponding to each sampling time. Indicates the first The sampling time of the first sampling moment Load values ​​corresponding to the industry category; Indicates industry category index, Indicates the sampling point index. .

[0099] The industry load series, total load series, temperature series, and humidity series are concatenated column-wise to form a feature matrix:

[0100] ;

[0101] In the formula, Indicates the first The total load value corresponding to each sampling time; This represents the total number of sampling points in the time series; This indicates the total number of industry categories; Indicates the sampling point index. ; Indicates industry category index, ; Indicates the first The sampling time of the first sampling moment Load values ​​corresponding to the industry category; Indicates the first Temperature values ​​corresponding to each sampling time; Indicates the first The humidity value corresponding to each sampling time.

[0102] S2. Calculate the correlation between temperature series and load series in the multivariate feature matrix, and combine the correlation strength and stability under extreme scenarios to screen the target industry load set that is sensitive to temperature.

[0103] Specifically, load types are screened based on the correlation between temperature and load sequences of various industries. To reflect the dynamic changes in correlation under extreme weather conditions, this invention preferably uses a sliding window method to calculate the window correlation coefficient between temperature and load of various industries, and divides the window into an extreme window set covering cold wave / high temperature sampling points and a regular window set; within the extreme window set, the correlation strength and volatility are statistically analyzed, and correlation thresholds and stability thresholds are set as constraints. When the correlation is significant and the volatility does not exceed the threshold, the corresponding industry load is determined to be a temperature-sensitive target load type and included in the target set; otherwise, it is removed. Preferably, the correlation threshold is 0.60, and the stability threshold is within a preset range. More preferably, a rank correlation or nonlinear sensitivity index can be introduced as an auxiliary constraint to improve the adaptability of the screening to nonlinear responses. The target industry load set obtained through step S2 is used to limit the load dimension range of subsequent extreme sample organization, data augmentation, and model training in this invention. Specifically, in step S3, when identifying cold wave days and high temperature days based on daily extreme temperatures and constructing a set of similar days, the load sequences in the "training samples / test samples" are only the industry load sequences corresponding to the target industry load set; in step S4, SMOTE synthesis expansion and two-stage training are also only performed on the training samples corresponding to the target industry load set, thereby outputting the prediction results for the target industry load set.

[0104] In the description of this invention, the calculation of the correlation between the temperature series and the load series in the multivariate feature matrix, and the selection of a set of temperature-sensitive target industry loads based on the correlation strength and stability under extreme scenarios, includes:

[0105] S21. Obtain the temperature series and any type of industry load series, and use a preset sliding window length to perform sliding windowing on the temperature series and industry load series; select the temperature subsequence and load subsequence within any window, and calculate the Pearson correlation coefficient of the window.

[0106] Specifically, obtain the temperature sequence With the Industry-specific load series Set the sliding window length to (Preferably 96 points / day or 48 points / half-day), with a step size (Preferably 1) Perform sliding windowing on the sequence; in the first... Within each window, the temperature subsequence and the load subsequence are recorded as follows: , Calculate the Pearson correlation coefficient of the window. The formula is as follows:

[0107] ;

[0108] in, The average temperature within the window. This represents the average industry load within the window.

[0109] S22. Based on the preset extreme scene annotation results, the window set is divided into an extreme window set and a normal window set, and the correlation strength and stability of the temperature subsequence and the load subsequence within the extreme window are statistically analyzed according to the Pearson correlation coefficient.

[0110] Specifically, based on the extreme weather annotation results, the window set is divided into extreme window sets. (Sampling points covering cold or hot days) and a set of regular sampling windows Statistically analyze the correlation strength and stability within extreme windows, including at least:

[0111]

[0112] ;

[0113] S23. Based on the preset dynamic correlation threshold and stability threshold, temperature sensitivity is determined. When the correlation strength is greater than or equal to the dynamic correlation threshold and the stability is less than or equal to the stability threshold, the current type of industry load sequence is determined to be temperature sensitive and included in the target industry load set.

[0114] Specifically, set a dynamic correlation threshold. With stability threshold When satisfied and At that time, the judgment of the first Loads exhibiting a stable and significant correlation with temperature under extreme weather conditions are included in the target industry load set; otherwise, they are excluded. The preferred value is 0.60. The preferred value is 0.15 to 0.25.

[0115] S3. Calculate the daily extreme temperature of the temperature sequence by day, identify different extreme weather days and the corresponding set of similar days based on the temperature threshold, and select representative test days from the set of similar days.

[0116] Specifically, for the industry load sequence corresponding to the target industry load set, the daily minimum and maximum temperatures are calculated based on the temperature sequence, and historical dates are labeled according to cold wave and high temperature criteria to obtain cold wave day sets and high temperature day sets. Based on this, cold wave similar day sets and high temperature similar day sets are determined, and representative test days are selected from each similar day set. The sampling points corresponding to the test days are removed from the training samples to avoid information leakage. Preferably, daily features can be constructed by combining multiple factors such as humidity and irradiance to perform secondary verification on candidate extreme days, thereby reducing the risk of misjudgment and missed judgment in extreme situation identification. Similar days can be selected based on multi-factor weighted similarity distance to ensure that the distribution of similar day samples is consistent with the target extreme scenario.

[0117] In the description of this invention, the step of calculating the daily extreme temperature of the temperature sequence by day, identifying different extreme weather days and corresponding sets of similar days based on temperature thresholds, and selecting representative test days from the set of similar days includes:

[0118] S31. Divide the temperature sequence by day to obtain a daily temperature sequence containing a fixed number of sampling points, and calculate the daily extreme temperature, including the daily minimum temperature and the daily maximum temperature.

[0119] S32. Set cold wave day determination conditions and high temperature day determination conditions. When the target date meets the cold wave day determination conditions, the corresponding date is marked as a cold wave day; when the target date meets the high temperature day determination conditions, the corresponding date is marked as a high temperature day. Among them, extreme weather days include cold wave days and high temperature days. The cold wave day determination condition means that the temperature drop of the target date within a preset time range reaches a preset temperature drop threshold, and the lowest temperature of the target date is lower than a preset low temperature threshold. The high temperature day determination condition means that the highest temperature of the target date is higher than a preset high temperature threshold.

[0120] Specifically, the temperature series is divided into days to obtain the daily minimum temperature. Daily maximum temperature Date will be changed when any of the following conditions are met. Date marked as a cold wave:

[0121] 1. Temperature drop exceeding 8°C within 24 hours and daily minimum temperature ≤4℃;

[0122] 2. Temperature drop exceeding 10℃ within 48 hours and the daily minimum temperature ≤4℃;

[0123] 3. Temperature drop exceeding 12°C within 72 hours and the daily minimum temperature ≤4℃;

[0124] when At that time, the date Days marked as high-temperature days; sets of similar days for cold waves and sets of similar days for high temperatures are determined separately, and cold wave test days are selected from the set of similar days for cold waves; when the set of similar days for cold waves is empty, the minimum daily temperature is selected. The date with the lowest temperature was used as the test date for the cold wave. The date was selected from the set of similar high-temperature days. The highest date was used as the high-temperature test day, and the sampling points corresponding to the two types of test days were removed from the training samples.

[0125] S33. Mark the dates that are not marked as cold wave days or high temperature days as regular weather, and perform clustering using the range of daily minimum temperature and daily maximum temperature as features to classify regular weather into multiple weather types.

[0126] Specifically, dates not marked as cold waves or high temperatures will be labeled as regular weather and... Clustering is performed as a feature to further divide regular weather into at least two weather types (e.g., sunny and cloudy) to improve the stability of similar day organization.

[0127] S34. Select dates marked as cold wave days or high temperature days, respectively form a set of similar cold wave days or a set of similar high temperature days, and select representative test days that meet the preset test conditions to construct a test set; the representative test days include cold wave test days and high temperature test days.

[0128] In the description of this invention, the step of selecting dates marked as cold wave days or high temperature days, respectively forming a set of similar cold wave days or a set of similar high temperature days, and selecting representative test days that meet preset conditions to construct a test set includes:

[0129] S341. Select the dates marked as cold wave days as the set of similar cold wave days.

[0130] S342. Select the dates marked as high-temperature days as the high-temperature similarity day set. When the high-temperature similarity day set is empty, select the dates that meet the condition that the daily maximum temperature is greater than the high-temperature threshold to form the high-temperature similarity day set. When the high-temperature similarity day set is still empty, sort the dates from high to low according to the daily maximum temperature and select a preset number of dates as the high-temperature similarity day set.

[0131] Specifically, dates marked as high temperatures are prioritized as the set of similar high-temperature days; when the set of similar high-temperature days is empty, dates that meet the following criteria are selected: A collection of days with similar high temperatures; when still empty, according to... Sort the dates from highest to lowest and select a preset number of dates as the set of similar high-temperature days; if the number of similar high-temperature days is insufficient, continue sorting by... The maximum value direction is supplemented to at least two days.

[0132] S343. Select cold wave test days that meet the preset conditions from the cold wave similar day set, and select the day with the highest daily maximum temperature from the high temperature similar day set as the high temperature test day; and remove all sampling points corresponding to the cold wave test day and the high temperature test day from the training set to use as the test set.

[0133] The test set includes multivariate sample points corresponding to all sampling times within the cold wave and high temperature test days, used for daily rolling prediction evaluation of the prediction model; the basic training set is the set of remaining sample points obtained after removing the test set; the augmented dataset is obtained by repeatedly sampling and synthesizing SMOTE on the training subset of the basic training set for extremely similar samples. The test set does not participate in data augmentation, model training, and hyperparameter optimization to avoid information leakage and ensure comparability of evaluations before and after augmentation.

[0134] S4. Based on the set of similar days and representative test days, expand and eliminate samples to form a basic training set and an enhanced training set. Construct a baseline model through two-stage training, and use the whale optimization algorithm to optimize the hyperparameters of the baseline model to form a load forecasting model. The load sequences in both the basic and enhanced training sets only include the industry load sequences corresponding to the target industry load set obtained in step S2, to ensure that data augmentation and model training focus on the load forecasting objects in temperature-sensitive industries.

[0135] Specifically, after dividing the sample set into training and validation sets according to a preset ratio, the preferred ratio is 8:2. SMOTE data augmentation is performed only on extremely similar samples within the training set. Based on the cold wave and high temperature test days determined in S3, the corresponding sampling points are removed from the training samples. In the basic training set after removal, the samples corresponding to the cold wave / high temperature similar days are marked as extremely similar samples. Only the extremely similar samples are used as seed samples for SMOTE synthesis augmentation to generate synthetic samples, which are then merged with the basic training samples to form an enhanced training set. This avoids leakage of test information and maintains the distribution of enhanced samples consistent with extreme scenarios. Furthermore, when the number of extreme samples n is insufficient to construct a K-nearest neighbor set, it is preferable to first use repeated sampling to ensure the number of samples meets the minimum requirement for nearest neighbor calculation before performing SMOTE synthesis to ensure the effectiveness of the synthetic samples and the stability of training.

[0136] In the description of this invention, the process of expanding and eliminating samples based on a set of similar days and representative test days to form a basic training set and an enhanced training set; constructing a benchmark model through two-stage training; and optimizing the hyperparameters of the benchmark model using the whale optimization algorithm to form a load prediction model includes:

[0137] S41. Based on the cold wave test days and high temperature test days, establish a mapping relationship between the daily index and the sample index to locate all sampling points corresponding to the two types of representative test days; remove the sampling points from the original sample set, and construct the basic training set based on the remaining samples after removing the representative test days.

[0138] Specifically, a "daily index - sample index" mapping relationship is established based on the cold wave test days and the high temperature test days to locate all sampling points corresponding to the test days, and these points are removed from the original sample set to obtain the basic training sample set. The It also includes regular weather samples as well as samples from similar days of cold waves / high temperatures, excluding the test day; The samples corresponding to the sets of cold wave similar days and the sets of high temperature similar days are marked as extremely similar samples, and the remaining samples are marked as non-extreme samples. The extremely similar samples are used for the second stage of SMOTE (Synthetic Minority Oversampling) synthesis expansion. The basic training sample set is divided into training subset and validation subset according to a preset ratio, preferably 8:2.

[0139] S42. Mark the samples in the basic training set that correspond to the sets of days similar to cold waves and days similar to high temperatures as extremely similar samples, and mark the remaining samples as non-extreme samples.

[0140] Specifically, in the basic training set Within this framework, samples corresponding to the sets of cold wave-like days and high temperature-like days are labeled as "extremely similar samples," while the remaining samples are labeled as "non-extreme samples." The "extremely similar samples" serve as seed samples for the second-stage SMOTE synthesis and expansion, while the first-stage training used... The entire dataset is used for training (i.e., it includes both extremely similar and non-extreme samples). When n < 2, it is impossible to construct an effective K-nearest neighbor set. It is preferable to first use repeated sampling to ensure that the number of samples meets the minimum requirement for nearest neighbor calculation, and then perform SMOTE synthesis to ensure the effectiveness of the synthesized samples and the stability of training.

[0141] S43. Select load period data corresponding to cold waves and high temperatures from multivariate time series to construct an extreme weather sample set; and output the prediction result sequence of target industry load by expanding, enhancing and standardizing the extreme weather sample set.

[0142] Among them, samples located in the training subset of the basic training set and belonging to the cold wave similar day set or the high temperature similar day set are defined as extreme similar samples in step S42. When the number of extreme similar samples is insufficient to meet the requirements of nearest neighbor calculation or model training stability, the extreme similar samples are first resampled to supplement the sample size, and then SMOTE synthesis expansion is performed using the extreme similar samples as seeds to obtain synthetic samples; the synthetic samples are merged with the training samples of the basic training set to form an enhanced training set. Standardization processing is performed by fitting normalized / standardized parameters to the training subset of the basic training set, and the same parameters are applied to the validation set and the test set to avoid information leakage. The final output is the prediction result sequence of the target industry load, that is, the load prediction time series for the next moment (or multiple future steps) obtained according to the sampling interval.

[0143] Specifically, load periods corresponding to cold waves and high temperatures are selected from the raw load and meteorological data to construct an extreme weather sample set. This sample set is represented in matrix form, denoted as:

[0144] ;

[0145] Where n represents the number of samples (number of extreme weather days), and m=17 represents the number of features; This represents the value of feature j for the i-th sample. The feature vector includes: 14 categories of individual loads, 1 category of total load, temperature, and humidity, for a total of 17 features. , This indicates the number of original extreme weather samples. This represents the total number of sampling points obtained within the target time range at fixed sampling intervals. This indicates the number of enhanced samples obtained after repeating the original extreme weather sample and performing SMOTE synthesis augmentation. .

[0146] Due to the limited number of extreme weather days, sample scarcity may occur. To ensure the stability of clustering and subsequent model training, it is necessary to determine whether the sample size is sufficient. If the number of extreme weather days (n) is small (e.g., n < 2), the sample size can be expanded through replication sampling. The `repmat` function in MATLAB is used to replicate the original samples, constructing a new sample matrix.

[0147] ;

[0148] Where N is the target sample multiple, satisfying , This is the expanded sample matrix.

[0149] If the sample size meets the requirements (e.g.) If the sampling is successful, the resampling step can be skipped and the process can proceed directly to the next step. For sample synthesis and enhancement, the nearest neighbor set of each sample needs to be obtained. The distance between any two samples is calculated using Euclidean distance, defined as:

[0150] ;

[0151] The distance matrix between samples is calculated as follows: To eliminate the influence of sample self-distance on the sorting results, the elements on the main diagonal of the matrix are set to infinity. Then, each row (i.e. each sample) is sorted from smallest to largest distance, and the k samples with the smallest distance are selected as its k-Nearest Neighbors (KNN) sample set. After obtaining the nearest neighbor samples, new synthetic samples are generated by using the feature differences between the neighbor samples to improve the representativeness and diversity of the samples.

[0152] The expression for generating new samples is:

[0153] ;

[0154] in, For the original sample, For samples randomly selected from its nearest neighbor set, This represents the random disturbance coefficient. New data points are generated in the sample space through linear interpolation, thereby enhancing the representativeness of the sample set under extreme weather conditions.

[0155] Finally, to eliminate the influence of different units on the clustering results, all features are standardized using Z-score standardization: the standardization formula is:

[0156] ;

[0157] in, and The first The mean and standard deviation of each feature; for the basic training sample set Each feature is processed using a uniform dimensionality, preferably using min-max normalization to map each feature to the [0, 1] interval, and the normalization parameter set is recorded. It is used for consistent normalization of the first-stage training / validation, consistent normalization of the input on the test day, and denormalization recovery of the prediction results.

[0158] S44. Set the sliding window length and perform sliding windowing on the normalized time series in chronological order to generate supervised learning sample pairs.

[0159] Specifically, set the window length to The normalized time series is subjected to sliding windowing in chronological order to generate supervised learning sample pairs. Specifically, the input sample is of length [length missing]. Multivariate sequence fragments:

[0160] ;

[0161] The output label represents the true value of the target industry load at the next time step after the window.

[0162] ;

[0163] in, This represents the total number of sampling points in the time series; Indicates the length of the sliding window; Indicates the target industry load in the first... The actual load value at each sampling time.

[0164] S45. The first stage of training of the baseline models of Long Short-Term Memory Neural Network (LSTM) and Bi-Long Short-Term Memory Neural Network (Bi-LSTM) is performed using the basic training set, and the key hyperparameters of the Bi-Long Short-Term Memory Neural Network baseline model are optimized using the Whale Optimization Algorithm (WOA) to form the load prediction model (WOA-Bi-LSTM).

[0165] In the description of this invention, the first-stage training of the Long Short-Term Memory Neural Network (LSTM) and Bidirectional LSM benchmark models using a basic training set, and the optimization of key hyperparameters of the Bidirectional LSM benchmark model using the whale optimization algorithm to form a load prediction model, includes:

[0166] S451. Divide the supervised learning sample pairs into a training subset and a validation subset according to a preset ratio, which will be used as the first-stage training set and the first-stage validation set, respectively.

[0167] Specifically, the basic training set is a set of multivariate time series samples after removing the test set. For supervised learning training, a set of supervised learning sample pairs is constructed in the basic training set using a sliding window and time alignment method, that is, the multivariate input time series within the window are used as features. The actual value of the target industry load at the corresponding prediction step size is used as the label. In the first stage of baseline model training, the supervised learning sample pairs constructed from the basic training set are divided into training subsets and validation subsets according to a preset ratio, preferably an 8:2 ratio, to obtain the first stage training set. With the first phase validation set .

[0168] S452. Using a multivariate sequence fragment composed of feature dimensions and window length as input, and the load prediction value of the target industry load at the next time moment as output, construct and train a long short-term memory neural network benchmark model and a bidirectional long short-term memory neural network benchmark model respectively; the input and output forms and network structures of the models are described in step S453.

[0169] S453. Select key hyperparameters that affect the predictive performance of the bidirectional long short-term memory neural network benchmark model to form a parameter vector to be optimized, and set the hyperparameter search range. In this invention, the load forecasting model adopts a supervised learning multivariate time series forecasting form; the model input is composed of... The feature variables have a length of The model outputs a multivariate sequence segment formed within a time window, and the model outputs the predicted value of the target industry load at the next time step; where, Indicates the input feature dimension. Indicates the length of the sliding window; in this embodiment, When the sampling interval is 15 minutes, the corresponding time span is 4 hours.

[0170] Specifically, the constructed multivariate feature matrix is ​​as follows:

[0171] ;

[0172] in, Indicates the length of the time series. This represents the number of feature dimensions. The sliding window length is set accordingly. The feature matrix is ​​constructed into supervised learning sample pairs in chronological order, and its input form is:

[0173] ;

[0174] The corresponding output is: That is, the actual load value of the target industry at the next moment. Preferably, the feature dimension... This includes 14 industry loads, total load, temperature, and humidity; the window length In the first phase, the settings were uniformly configured across all benchmark models to ensure model comparability.

[0175] Based on this, LSTM and Bi-LSTM benchmark models were constructed and trained respectively. For example... Figure 2 As shown, the LSTM baseline model includes: a sequence input layer (input feature dimension is 17, input data is...) The algorithm includes: a multivariate time series segment; two LSTM layers (10 hidden units in the first layer and 5 hidden units in the second layer); a Dropout layer (dropout rate of 0.2); a fully connected layer; and a regression output layer (without an activation function, outputting continuous value prediction results). The LSTM layers internally use sigmoid and tanh as gating activation functions to control the input, forget, and output gates, enabling selective memorization and updating of historical information. Figure 3 As shown, the Bi-LSTM baseline model includes: a sequence input layer (input feature dimension is 17, input data is...) The network structure consists of: a multivariate sequence fragment; a Bi-LSTM layer (bidirectional structure with 8 hidden units); a Dropout layer (dropout rate of 0.3); a fully connected layer; and a regression output layer (outputting the predicted value for the next time step). The WOA-Bi-LSTM model maintains the same network topology as the Bi-LSTM baseline model, but differs in that the number of hidden units, initial learning rate, and maximum number of training epochs are adaptively determined by the whale optimization algorithm. By maintaining a consistent network structure and optimizing only key hyperparameters, structural comparability between different models is ensured.

[0176] To ensure comparability in the future, all benchmark models in the first phase adopted the same input feature set and the same window length. The same normalization method and the same 8:2 partitioning strategy are used; preferably, the basic training set is divided into a training subset and a validation subset in chronological order. The validation subset is used for early stopping / validation monitoring and hyperparameter fitness evaluation to improve training stability.

[0177] Furthermore, in this invention, the LSTM, Bi-LSTM, and WOA-Bi-LSTM models all employ a unified training process and optimization strategy: the data source is the full year data of 2020 from the public dataset "ScientificData:EWELD"; all input features are normalized to intervals using Min-Max normalization. The regression output layer uses mean squared error (MSE) as the loss function.

[0178] ;

[0179] The model was trained using the Adam optimizer, with the initial learning rate set to a range of [value missing]. An early stopping strategy is introduced during training. Training is automatically terminated when the validation set error no longer decreases within a certain number of consecutive rounds, in order to prevent overfitting and improve training stability.

[0180] S4531. Setting the hyperparameter vector to be optimized and the search range. Select key hyperparameters that significantly affect the prediction performance of Bi-LSTM to construct the parameter vector to be optimized:

[0181] ;

[0182] in, The number of hidden units in the Bi-LSTM. The initial learning rate, This represents the maximum number of training epochs. To ensure search efficiency and model trainability, the search range is set as follows:

[0183]

[0184]

[0185] ;

[0186] The above range can be adjusted to an equivalent range based on hardware resources and data scale. Discretization and constraint correction: for and Perform rounding operation on the... Perform upper and lower bound truncation to ensure that the candidate parameters generated in each iteration are all trainable and effective parameters.

[0187] S454. Set the fitness function and fitness index of the whale optimization algorithm, and perform key hyperparameter optimization through iterative optimization to generate the optimal parameters and obtain the optimized load prediction model.

[0188] In the description of this invention, the process of setting the fitness function and fitness index of the whale optimization algorithm, and performing key hyperparameter optimization through iterative optimization to generate optimal parameters and obtain the optimized load prediction model includes:

[0189] S4541. To eliminate differences in the dimensions of different features, Min-Max normalization is performed on the feature matrix, scaling each feature to the range. For any feature dimension raw values ​​on Normalization results Defined as:

[0190] ;

[0191] in, and The first The minimum and maximum values ​​of each feature in the training data. To avoid information leakage, the... , It is obtained only from the statistics of the first-stage training subset, and the same set of normalization parameters is reused for normalization and denormalization in the validation set, test set and rolling prediction stage.

[0192] S4542. The prediction error of the first-stage validation set is used as the fitness function of the Whale Optimization Algorithm (WOA), preferably with the root mean square error (RMSE) as the fitness metric. Let the true value of the validation set be... The predicted value is The sample size is Then RMSE is defined as:

[0193] ;

[0194] In each iteration, WOA is based on the candidate hyperparameter vector. Train the corresponding Bi-LSTM model and compute its performance on the validation set. As a fitness value; to minimize Update the population position for the target and output the optimal combination of hyperparameters.

[0195] S4543. The key hyperparameters of the bidirectional long short-term memory neural network benchmark model are globally optimized using the whale optimization algorithm. Several hyperparameter combinations are initialized to form a whale population. The fitness value of the bidirectional long short-term memory neural network benchmark model corresponding to each hyperparameter group is calculated in the first stage validation set. The population position is iteratively updated by simulating the behavior of humpback whales, and the hyperparameter combinations are continuously optimized.

[0196] Specifically, the key hyperparameters of Bi-LSTM are encoded into parameter vectors:

[0197] ;

[0198] in, To hide the number of units, The initial learning rate, To determine the maximum number of training rounds; define the search boundary for the parameter vector and perform discretization and constraint correction, where... and Rounding down, for Upper and lower bounds are truncated; the prediction error on the validation set is used as the WOA fitness function, with RMSE preferably used as the fitness metric. The WOA iteratively updates the parameter vector and outputs the optimal parameters. Based on optimal parameters The WOA-Bi-LSTM model is trained, and the optimal parameters are fixed for the second stage of retraining or fine-tuning to ensure consistency before and after enhancement.

[0199] S4544. After reaching the preset number of iterations or meeting the convergence condition, output the optimal parameters of the bidirectional long short-term memory neural network benchmark model, and construct the load prediction model based on the optimal parameters.

[0200] S46. Using extremely similar samples as the core, an enhanced training set is constructed by expanding and merging, and the enhanced training set is used to perform the second stage of training to achieve retraining and fine-tuning of the load prediction model.

[0201] In the description of this invention, the step of using extremely similar samples as the core, expanding and merging to construct an enhanced training set, and using the enhanced training set to perform a second-stage training to achieve retraining and fine-tuning of the load prediction model includes:

[0202] S461. Using extremely similar samples as the core, perform synthetic minority class oversampling augmentation on cold wave similar day samples and high temperature similar day samples respectively to generate cold wave synthetic samples and high temperature synthetic samples, and combine them with the basic training set after removing representative test days to form an enhanced training set.

[0203] Specifically, the steps for SMOTE synthesis augmentation to form an enhanced training set are as follows:

[0204] Take any sample from extremely similar samples. Selecting its based on distance metric A set of nearest neighbor samples And randomly select the nearest neighbor samples from them. Generate synthetic samples using the following formula:

[0205] ;

[0206] in, The random perturbation coefficients are used to form the enhanced training set by merging the synthetic samples with the basic training sample set. .

[0207] S462. Normalize the augmented training set and record the parameters. Construct supervised sample pairs using the same window length as in the first stage, and divide them into a second-stage training set and a second-stage validation set according to a preset ratio. Retrain or fine-tune the long short-term memory neural network benchmark model, the bidirectional long short-term memory neural network benchmark model, and the load prediction model on the augmented training set.

[0208] S5. Using the optimized load forecasting model, perform daily rolling forecasts for representative test days and output the load forecast values ​​for the corresponding target industry loads in the target industry load set at the next time moment.

[0209] Specifically, the optimized load forecasting model (WOA-Bi-LSTM) is used to perform daily rolling forecasts for cold wave test days and high temperature test days, outputting industry load forecast curves, and calculating forecast evaluation indicators including at least root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) to compare the forecasting effects before and after enhancement and different models.

[0210] Extract daily multivariate sequence data corresponding to cold wave and high temperature test days. The preferred number of sampling points per day is 96, and the sampling interval is 15 minutes. Normalize the test day data using the same normalization parameters as during model training, and set the window length. Construct a sliding window input sequence; input the window input sequence into the pre-amplification model and the post-amplification model respectively to output the prediction sequence, and perform inverse normalization and time alignment on the prediction sequence;

[0211] The evaluation indicators used are root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). The formulas for RMSE and MAPE are as follows:

[0212] ;

[0213] ;

[0214] ;

[0215] in, For the actual load, To predict load, This is the actual average load. This represents the number of predicted points for a single day.

[0216] In another embodiment, the present invention also provides a load forecasting system that takes into account data augmentation under extreme weather conditions, the system comprising:

[0217] The fixed sampling module is used to acquire meteorological data and historical industry load data of the target area at fixed sampling intervals, and generate a multivariate feature matrix after anomaly detection and interpolation repair.

[0218] The load type filtering module is used to calculate the correlation between temperature series and load series in the multivariate feature matrix, and to filter the set of temperature-sensitive target industry loads by combining the correlation strength and stability under extreme scenarios.

[0219] The weather recognition module is used to calculate the daily extreme temperature of the temperature sequence, identify different extreme weather days and the corresponding set of similar days based on the temperature threshold, and select representative test days from the set of similar days.

[0220] The model building module is used to expand and eliminate samples based on the set of similar days and representative test days, forming a basic training set and an enhanced training set; a baseline model is built through two-stage training, and the hyperparameters of the baseline model are optimized using the whale optimization algorithm to form a load prediction model;

[0221] The rolling forecasting module is used to perform daily rolling forecasts on representative test days using the optimized load forecasting model, and outputs the load forecast value of the corresponding target industry load in the target industry load set at the next time moment.

[0222] This invention validates the data-enhanced load prediction method under extreme weather scenarios using the MATLAB software simulation platform, demonstrating the effectiveness and feasibility of the proposed solution under extreme weather conditions such as cold waves and high temperatures. Data source: The invention uses full-year data from the publicly available dataset "Scientific Data:EWELD" for 2020. This data includes industry loads and meteorological data (temperature, humidity, etc.), which has been cleaned and interpolated for training. Data preprocessing: Min-Max normalization is performed to scale the feature values ​​to the interval [0,1], eliminating dimensional differences.

[0223] In terms of feature construction, this invention performs correlation analysis between temperature and various industry loads, preferably using the Pearson correlation coefficient and setting a threshold (e.g., |r|≥0.6) to screen temperature-sensitive load types; taking the financial services industry load as an example, it is identified as a temperature-sensitive load, thus serving as the target predicted load, and together with meteorological elements such as temperature and humidity, it forms a multivariate input feature for subsequent load prediction model training and inference.

[0224] In terms of model setup, to demonstrate the synergistic effect of "data augmentation and hyperparameter optimization" in this invention, LSTM and Bi-LSTM are selected as benchmark models for comparison, and a Bi-LSTM optimized by the whale optimization algorithm (WOA-Bi-LSTM) is constructed as the core prediction model to compare the prediction differences of different model structures and parameter optimization strategies under extreme weather scenarios.

[0225] Regarding the training strategy, this invention employs a two-stage training mechanism to ensure consistent evaluation criteria before and after enhancement. The first stage is the pre-enhancement training stage: after removing sample points corresponding to cold wave and high temperature test days, a basic training set is constructed using the remaining samples. The training and validation sets are then divided in an 8:2 ratio to complete the baseline training of LSTM and Bi-LSTM, as well as hyperparameter optimization and modeling of WOA-Bi-LSTM. The second stage is the post-enhancement fine-tuning stage: using samples from similar cold / high temperature days as the core, SMOTE is preferably used for sample synthesis and expansion (e.g., expansion by 20%). The synthesized samples are then merged with the basic training samples to form the enhanced training set. Under the same window length, feature set, and normalization strategy, the training and validation sets are again divided in an 8:2 ratio. LSTM, Bi-LSTM, and WOA-Bi-LSTM are then retrained or fine-tuned to achieve repeatability and consistency in the comparison before and after enhancement.

[0226] Regarding daily extreme weather forecasting, this invention performs rolling forecasts for both cold wave and high temperature test days, and outputs corresponding load forecast curves. The forecast results for the cold wave test day before enhancement are as follows: Figure 4 As shown, the enhanced cold wave test date prediction results are as follows: Figure 5 As shown; the predicted results for the high-temperature test day before enhancement are as follows: Figure 6 As shown, the predicted results for the enhanced high-temperature test day are as follows: Figure 7 As shown. By Figures 4 to 7 It is evident that without SMOTE enhancement, the model exhibits significant tracking biases in response to load abrupt changes, peak-valley variations, and uphill / downhill processes under extreme weather conditions. After introducing SMOTE enhancement with similar day constraints and completing the second-stage fine-tuning, the correlation between the predicted curve and the actual load curve is significantly improved, and the bias in the abrupt change interval is significantly suppressed.

[0227] To further illustrate the enhancement effect from the perspective of error distribution, this invention statistically analyzed the prediction error distribution before and after enhancement in high-temperature and cold-wave scenarios. The comparison results of the high-temperature error distribution are as follows: Figure 8 As shown in the figure, the comparison results of the cold wave error distribution are as follows: Figure 9 As shown. By Figure 8 and Figure 9 As can be seen, the enhanced error distribution is more concentrated and the outlier error is reduced, indicating that the SMOTE enhancement method of this invention can improve the robustness and stability of the model under extreme weather conditions.

[0228] Regarding evaluation metrics, this invention uses root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) to quantitatively evaluate the predictive performance of each model on cold wave and high temperature test days. The evaluation results before data augmentation are shown in Table 1, and the evaluation results after data augmentation are shown in Table 2.

[0229] Table 1. Statistics of Daily Extreme Weather Forecast Errors Before Data Augmentation

[0230]

[0231] As shown in Table 1, in the pre-enhancement stage, WOA-Bi-LSTM exhibits lower RMSE and MAPE and higher R² compared to LSTM and Bi-LSTM, indicating that hyperparameter optimization based on WOA can improve the fitting ability of Bi-LSTM to load time series characteristics.

[0232] Table 2. Statistics of Daily Forecast Errors for Extreme Weather After Data Augmentation

[0233]

[0234] As shown in Table 2, after introducing SMOTE enhancement and completing the second stage of training, the RMSE and MAPE of the three models were further reduced and the R² was further improved on the cold wave and high temperature test days. Among them, WOA-Bi-LSTM+SMOTE had the best overall effect, indicating that "WOA optimization and SMOTE enhancement" can synergistically improve the prediction accuracy in extreme weather scenarios.

[0235] As can be seen from the comparison results in Tables 1 and 2, the WOA-Bi-LSTM model has achieved better prediction accuracy than the traditional LSTM and Bi-LSTM without data augmentation. After introducing SMOTE data augmentation with similar day constraints and completing the second stage of fine-tuning, the RMSE and MAPE of the three models are significantly reduced in cold wave and high temperature scenarios. Among them, the WOA-Bi-LSTM model has the best overall prediction performance, which verifies the effectiveness and feasibility of the "hyperparameter optimization and extreme sample augmentation synergistic mechanism" proposed in this invention.

[0236] From the perspective of sample expansion effect, taking the load of the financial services industry as an example, the number of samples for cold wave-like days was 96, and after expansion by 20%, the number of synthetic samples was 19; the number of samples for high temperature-like days was 4224, and after expansion, the number of synthetic samples was 845; after the fusion of synthetic samples, the total number of training samples increased from 34944 to 35808, thereby alleviating the problem of insufficient model training caused by the scarcity of extreme weather samples and improving the model's generalization ability in extreme weather scenarios.

[0237] In summary, the proposed combination of "temperature-sensitive load screening, extreme weather identification, similar-day constrained SMOTE enhancement, and WOA-Bi-LSTM prediction" effectively alleviates the problem of scarce extreme weather samples and improves the model's prediction accuracy and generalization ability under extreme weather scenarios such as cold waves and high temperatures. Based on this technical solution, this invention can achieve higher accuracy and stability in both conventional and extreme weather load forecasting tasks, and has broad application prospects, particularly in areas such as refined load forecasting by industry, extreme weather load early warning, and power system dispatch auxiliary decision-making.

[0238] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

Claims

1. A load forecasting method considering data augmentation under extreme weather conditions, characterized in that, include: Meteorological data and historical industry load data for the target area are acquired at fixed sampling intervals, and after anomaly detection and interpolation repair, a multivariate feature matrix is ​​generated. Calculate the correlation between temperature and load sequences in the multivariate feature matrix, and combine the correlation strength and stability under extreme scenarios to screen the target industry load set that is sensitive to temperature. The daily extreme temperature is calculated on a daily basis according to the temperature sequence. Different extreme weather days and the corresponding set of similar days are identified based on the temperature threshold. Representative test days are selected from the set of similar days. Based on the set of similar days and representative test days, the samples are expanded and eliminated to form a basic training set and an enhanced training set. A baseline model is built through two-stage training, and the hyperparameters of the baseline model are optimized using the whale optimization algorithm to form a load prediction model. The optimized load forecasting model is used to make daily rolling forecasts for representative test days, and the load forecast values ​​of the corresponding target industry loads in the target industry load set at the next time point are output. The process of calculating the correlation between temperature and load sequences in the multivariate feature matrix, and combining the correlation strength and stability under extreme scenarios, to screen the set of temperature-sensitive target industry loads includes: Obtain temperature series and industry load series of any type, and perform sliding windowing on temperature series and industry load series using a preset sliding window length; select temperature subsequence and load subsequence within any window, and calculate the Pearson correlation coefficient of the window; Based on the pre-defined extreme scenario annotation results, the window set is divided into an extreme window set and a normal window set, and the correlation strength and stability of the temperature subsequence and the load subsequence within the extreme window are statistically analyzed according to the Pearson correlation coefficient. Temperature sensitivity is determined based on preset dynamic correlation threshold and stability threshold. When the correlation strength is greater than or equal to the dynamic correlation threshold and the stability is less than or equal to the stability threshold, the current type of industry load sequence is determined to be temperature sensitive and included in the target industry load set. The formula for the correlation strength is: ； In the formula, This indicates the correlation strength between the temperature subsequence and the load subsequence within the extreme window; Represents an extreme set of windows; This represents the Pearson correlation coefficient within the window; Represents a window; The formula for stability is: ； In the formula, This indicates the stability of the temperature subsequence and the load subsequence within the extreme window.

2. The load forecasting method considering data enhancement under extreme weather conditions according to claim 1, characterized in that, The process of acquiring meteorological data and historical industry load data for the target area at fixed sampling intervals, and generating a multivariate feature matrix after anomaly detection and interpolation repair, includes: Acquire meteorological data and corresponding historical industry load data for the target region, and align them to the same fixed sampling interval according to the timestamp to form the original multivariate time series; The original multivariate time series is sampled for consistency verification. When duplicates, jump points or abnormal intervals are found, deduplication and alignment are performed, and the unaligned segments are marked as intervals to be repaired. The Lagrange interpolation method is used to interpolate and repair the points to be repaired in the interval to be repaired. After repair, the continuity and boundary rationality checks are performed to obtain a continuous multivariate time series. Various industry load sequences are extracted from the multivariate time series, and the total load sequence is obtained by summing the multiple types of industry load sequences according to time. The industry load sequence, total load sequence, temperature sequence and humidity sequence are then spliced ​​together to form a multivariate feature matrix.

3. The load forecasting method considering data enhancement under extreme weather conditions according to claim 1, characterized in that, The process of calculating daily extreme temperatures from the temperature sequence, identifying different extreme weather days and their corresponding sets of similar days based on temperature thresholds, and selecting representative test days from the set of similar days includes: The temperature sequence is divided into days to obtain a daily temperature sequence containing a fixed number of sampling points, and the daily extreme temperature, including the daily minimum temperature and the daily maximum temperature, is calculated. The system sets conditions for determining cold wave days and high temperature days. When a target date meets the cold wave day condition, the corresponding date is marked as a cold wave day; when a target date meets the high temperature day condition, the corresponding date is marked as a high temperature day. The extreme weather days include both cold wave days and high temperature days. The cold wave day condition indicates that the temperature drop of the target date within a preset time range reaches a preset temperature drop threshold, and the lowest temperature of the target date is lower than a preset low temperature threshold. The high temperature day condition indicates that the highest temperature of the target date is higher than a preset high temperature threshold. Dates not marked as cold wave days or high temperature days are labeled as regular weather days, and clustering is performed using the range of daily minimum temperature and daily maximum temperature as features to classify regular weather days into multiple weather types. Select dates marked as cold wave days or high temperature days, and build sets of similar days for cold waves or high temperature days respectively. Select representative test days that meet the preset test conditions to construct a test set. The representative test days include cold wave test days and high temperature test days.

4. A load forecasting method considering data enhancement under extreme weather conditions according to claim 3, characterized in that, The process of selecting dates marked as cold wave days or high temperature days, constructing sets of similar cold wave days or high temperature days respectively, and selecting representative test days that meet preset conditions to construct a test set includes: Dates marked as cold wave days are selected as the set of similar cold wave days; Dates marked as high-temperature days are selected as the high-temperature similarity day set. When the high-temperature similarity day set is empty, dates that meet the condition that the daily maximum temperature is greater than the high-temperature threshold are selected to form the high-temperature similarity day set. When the high-temperature similarity day set is still empty, the dates are sorted from high to low according to the daily maximum temperature and a preset number of dates are selected as the high-temperature similarity day set. Select cold wave test days that meet the preset conditions from the cold wave similar day set, and select the day with the highest daily maximum temperature from the high temperature similar day set as the high temperature test day; remove all sampling points corresponding to the cold wave test day and the high temperature test day from the training set and use them as the test set.

5. A load forecasting method considering data enhancement under extreme weather conditions according to claim 1, characterized in that, The sample is expanded and eliminated based on the set of similar days and representative test days to form a basic training set and an enhanced training set. A baseline model is built through two-stage training, and the hyperparameters of the baseline model are optimized using the whale optimization algorithm to form a load prediction model, including: Based on the cold wave test days and high temperature test days, a mapping relationship between the daily index and the sample index is established to locate all sampling points corresponding to the two types of representative test days; the sampling points are removed from the original sample set, and the remaining samples after removing the representative test days are used to construct the basic training set; The samples in the basic training set that correspond to the sets of cold wave-similar days and the sets of high temperature-similar days are labeled as extremely similar samples, and the remaining samples are labeled as non-extreme samples. Load period data corresponding to cold waves and high temperatures are selected from multivariate time series to construct an extreme weather sample set; and by expanding, enhancing and standardizing the extreme weather sample set, the predicted load sequence of the target industry is output. Set the sliding window length, and perform sliding windowing on the normalized time series in chronological order to generate supervised learning sample pairs; The first stage of training of the baseline models of long short-term memory neural network and bidirectional long short-term memory neural network was carried out using the basic training set, and the key hyperparameters of the bidirectional long short-term memory neural network baseline model were optimized using the whale optimization algorithm to form a load prediction model. Using extremely similar samples as the core, an enhanced training set is constructed by expanding and merging, and the enhanced training set is used to perform the second stage of training, thereby realizing the retraining and fine-tuning of the load prediction model.

6. A load forecasting method considering data enhancement under extreme weather conditions according to claim 5, characterized in that, The process of performing the first stage training of the baseline models of the Long Short-Term Memory Neural Network and the Bidirectional Long Short-Term Memory Neural Network using the basic training set, and optimizing the key hyperparameters of the Bidirectional Long Short-Term Memory Neural Network baseline model using the whale optimization algorithm to form the load prediction model includes: The supervised learning sample pairs are divided into a training subset and a validation subset according to a preset ratio, which serve as the first-stage training set and the first-stage validation set, respectively. Using a multivariate sequence fragment composed of feature dimensions and window length as input, and the load prediction value of the target industry load at the next time moment as output, a long short-term memory neural network benchmark model and a bidirectional long short-term memory neural network benchmark model are constructed and trained. Key hyperparameters affecting the predictive performance of the bidirectional long short-term memory neural network benchmark model are selected to form a parameter vector to be optimized, and the hyperparameter search range is set. The fitness function and fitness index of the whale optimization algorithm are set, and the key hyperparameters are optimized through iterative optimization to generate the optimal parameters and obtain the optimized load prediction model.

7. A load forecasting method considering data enhancement under extreme weather conditions according to claim 6, characterized in that, The process involves setting the fitness function and fitness index of the whale optimization algorithm, and iteratively optimizing key hyperparameters to generate optimal parameters, resulting in an optimized load prediction model. Feature matrices containing multiple types of component load sequences, total load sequences, temperature sequences, and humidity sequences are extracted from the original multivariate time series and normalized to eliminate the influence of dimensions. The prediction error of the first-stage validation set is used as the fitness function of the whale optimization algorithm, and the root mean square error is used as the fitness index. The key hyperparameters of the bidirectional long short-term memory neural network benchmark model are globally optimized using the whale optimization algorithm. Several hyperparameter combinations are initialized to form a whale population. The fitness value of the bidirectional long short-term memory neural network benchmark model corresponding to each hyperparameter group is calculated in the first stage validation set. The population position is iteratively updated by simulating the behavior of humpback whales, and the hyperparameter combinations are continuously optimized. The optimal parameters of the bidirectional long short-term memory neural network baseline model are output after the preset number of iterations or the convergence condition is met, and a load prediction model is constructed based on the optimal parameters.

8. A load forecasting method considering data augmentation under extreme weather conditions according to claim 6, characterized in that, The process of using extremely similar samples as the core, expanding and merging them to construct an enhanced training set, and then using this enhanced training set to perform the second stage of training to achieve retraining and fine-tuning of the load prediction model includes: Using extremely similar samples as the core, synthetic minority class oversampling augmentation is performed on cold wave similar day samples and high temperature similar day samples respectively to generate cold wave synthetic samples and high temperature synthetic samples. These are then combined with the basic training set after removing representative test days to form an enhanced training set. Normalization is performed on the augmented training set and parameters are recorded. Supervision sample pairs are constructed using the same window length as in the first stage and divided into a second-stage training set and a second-stage validation set according to a preset ratio. The long short-term memory neural network benchmark model, the bidirectional long short-term memory neural network benchmark model, and the load prediction model are retrained or fine-tuned on the augmented training set.

Images