Partition load increase situation prediction method and system oriented to extreme weather

Abstract

The invention discloses a partition load growth situation prediction method and system oriented to extreme weather, and the method comprises the steps: firstly obtaining and preprocessing multi-source historical data, and building an extreme weather event recognition mechanism, so as to generate an extreme weather event tag sequence; then, the load response characteristics of all the partitions are quantified based on the extreme weather time period, and the partitions are divided into a plurality of load increase situation categories with different response modes by adopting a clustering algorithm; thirdly, independently constructing and training an exclusive load prediction model for each load growth situation category; and finally, during prediction execution, selecting the sentry subareas in each category, and dynamically calibrating original prediction results of other follower subareas in the same category by monitoring prediction residual errors of the sentry subareas in real time and calculating a prospective correction amount according to a deviation propagation model. According to the invention, combination of classified exclusive modeling and real-time dynamic correction is realized, and the precision and reliability of load prediction in extreme weather are improved.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting and intelligent control technology, and more specifically, to a method and system for predicting regional load growth trends under extreme weather conditions. Background Technology

[0002] With the increasing severity of global climate change, extreme weather events (such as persistent high temperatures, cold waves, strong winds, and torrential rains) are occurring more frequently, posing a severe challenge to the safe and stable operation of the power system. Under extreme weather conditions, the loads of residential air conditioning, heating, and electric heating equipment fluctuate dramatically, leading to a surge in electricity demand in some areas and even inducing problems such as grid congestion, frequency fluctuations, and energy consumption imbalances. Traditional load forecasting models are often unable to cope with the complex problem scenarios of sudden changes and high randomness in meteorological parameters under extreme weather conditions.

[0003] The regional load growth trend prediction system for extreme weather conditions is suitable for behavior modeling and trend analysis of regional power load during extreme weather events such as extreme high temperatures and cold waves. It can be applied to scenarios such as power grid dispatching, demand-side management, load balance control and urban energy security management.

[0004] For example, patent document CN117955083A discloses a method for predicting power load under extreme weather conditions. This method includes frequency domain decomposition of historical load data, dividing it into low-frequency load components and high-frequency components, then combining the decomposed components with extreme weather data to construct an input matrix, which is finally input into a stacked long short-term memory network for prediction, and using an improved particle swarm optimization algorithm to train the model. Another example is patent document CN118889379B, which provides a method and system for predicting load under extreme high-temperature weather based on tensor completion. It filters data by constructing a comprehensive high-temperature index including a heat index, uses a tensor completion algorithm to solve the data missing problem, and filters input features through correlation analysis. Finally, it uses a modified model combining a long short-term memory network and rough set theory for prediction.

[0005] However, these existing technologies all aim to build a general forecasting model and attempt to improve the overall forecasting accuracy of this model by improving algorithms or data processing methods. They fail to fully consider the significant differences in load response patterns across different geographical regions under extreme weather conditions. For example, regions dominated by industrial loads are significantly less sensitive to temperature changes than regions dominated by residential air conditioning loads. This results in a bottleneck in forecasting capabilities when facing specific regions with highly pronounced response characteristics. Second, existing forecasting systems are typically isolated forecasts operating independently, lacking a spatiotemporal linkage mechanism between regions. In fact, the movement of extreme weather events (such as cold fronts and typhoons) exhibits significant spatial transmission; forecasting deviations in upstream regions often foreshadow future trends in downstream regions. Existing technologies fail to utilize this deviation information from earlier regions to calibrate forecasts for later regions in real time, resulting in an inability to provide forward-looking accuracy compensation at the moment of abrupt weather changes.

[0006] Therefore, a technical system is needed that can integrate multi-source meteorological data, load response characteristics, regional clustering modeling, and deep neural network prediction to achieve accurate modeling and intelligent prediction of regional power load growth trends under extreme weather conditions, providing scientific support for subsequent regulation strategies. Summary of the Invention

[0007] To address the technical problems of decreased load forecasting accuracy and increased difficulty in regional load regulation under extreme weather conditions in existing technologies, this invention provides a method and system for predicting regional load growth trends under extreme weather conditions. First, multi-source historical data is acquired and preprocessed, and an adaptive extreme weather event identification mechanism is established to generate extreme weather event label sequences. Then, the load response characteristics of each region are quantified based on the extreme weather period, and a clustering algorithm is used to divide the regions into multiple load growth trend categories with different response patterns. Next, for each load growth trend category, a dedicated load forecasting model is independently constructed and trained. Finally, during forecasting execution, sentinel regions are selected within each category. By real-time monitoring of the forecast residuals of the sentinel regions and calculating forward-looking corrections based on the deviation propagation model, the original forecast results of other follower regions within the same category are dynamically calibrated. This invention combines category-specific modeling with real-time dynamic correction, improving the accuracy and reliability of load forecasting under extreme weather conditions.

[0008] The present invention adopts the following technical solution.

[0009] In a first aspect, this invention relates to a method for predicting regional load growth trends under extreme weather conditions, the method comprising the following steps: Step 1: Acquire and preprocess multi-source historical data for each partition within the target area; Step 2: Establish an extreme weather event identification mechanism, analyze the preprocessed multi-source historical data to identify and label the time periods of extreme weather events, and generate extreme weather event label sequences; Step 3: Based on the extreme weather event period, quantify the load response characteristics of each zone, and use a clustering algorithm to analyze the load response characteristics to divide each zone into at least two load growth trend categories. Step 4: Construct and implement a hierarchical prediction model to predict the load for future periods based on the load growth trend category, so as to generate load growth trend prediction results; Step 5: Select sentinel zones within each of the load growth trend categories, and designate the remaining zones as follower zones; monitor the load prediction residuals of the sentinel zones in real time, where the load prediction residuals are the difference between the measured load value of the sentinel zone and the load growth trend prediction result; when the load prediction residuals trigger preset calibration conditions, calculate the forward-looking correction amount of the follower zones based on the deviation propagation model constructed based on the spatiotemporal correlation between the sentinel zones and the follower zones; use the forward-looking correction amount to calibrate the load growth trend prediction result of the follower zones, and generate the final load growth trend prediction result.

[0010] Preferably, the multi-source historical data includes: historical power load data, historical meteorological data corresponding to each zone, and supplementary information data for each zone; The historical meteorological data includes at least one of temperature, humidity, wind speed, and total solar radiation; The supplementary information data for the partition includes holiday markers and electricity consumption type labels.

[0011] Preferably, in step 2, the process of generating the extreme weather event label sequence includes: An adaptive event definition mechanism is adopted to calculate dynamic thresholds based on the historical meteorological data; The sliding window detection method is used to determine whether the historical meteorological data for N consecutive days exceeds or falls below the dynamic threshold in order to identify the period of the extreme weather event; The extreme weather event period is encoded using Boolean label vectors. All hourly time points corresponding to the period identified as an extreme weather event are encoded as 1, and the remaining time points are encoded as 0.

[0012] Preferably, in step 3, the load response characteristics of each zone are quantified based on the duration of the extreme weather event, including: Calculate the peak growth rate, which measures the increase in peak load during the extreme weather event period relative to the average daily load during normal periods; the average daily load during normal periods is calculated as the arithmetic mean of all load sampling points over M consecutive days prior to the start of the extreme weather event period. Calculate the peak-to-valley ratio, which measures the ratio between the maximum load and the minimum load during the period of the extreme weather event; Calculate the standard deviation fluctuation, which measures the standard deviation of all load sampling points during the period of the extreme weather event.

[0013] Preferably, in step 4, constructing and implementing a hierarchical prediction model specifically includes: Obtain weather forecast data and holiday markers for the target area for future periods; The meteorological forecast data for the future period and the holiday markers are input into a pre-trained situation classification model to obtain the predicted load growth situation category. Based on the predicted load growth trend category, select the corresponding dedicated load forecasting model from the dedicated load forecasting models corresponding to different load growth trend categories; The recent historical feature sequence is input into the selected dedicated load forecasting model to initialize the hidden state of the dedicated load forecasting model. The meteorological forecast data for the future period and the holiday markers are used as the driving input sequence to output the load growth trend forecast result.

[0014] Preferably, in step 5, selecting the sentry zones specifically includes: The response sensitivity metric is determined by performing a linear weighted combination of the peak growth rate, peak-to-valley ratio, and standard deviation fluctuation quantified in step 3. The data quality metric is determined by comprehensively evaluating the integrity rate and average transmission delay of the historical power load data and the historical meteorological data using a preset scoring function. Determine the spatiotemporal leading measure and adaptively adjust it according to the type of the extreme weather event period, including: If the extreme weather event period corresponds to a mobile weather event, it is determined according to the geographical order of the regions in the meteorological forecast path, with the region that is affected first receiving the highest score. If the extreme weather event period corresponds to a coverage weather event, it is determined based on the time delay calculated by cross-correlation analysis of the historical power load data, with the partition that responds most advanced receiving the highest score. The response sensitivity metric, the data quality metric, and the spatiotemporal leadership metric are weighted and summed to obtain the comprehensive qualification score. Based on the comprehensive qualification score, one or more partitions with the highest scores in each load growth trend category are identified as the sentinel partitions.

[0015] Preferably, step 5, the process of generating the final load growth trend prediction result, includes: The load forecast residuals of the sentinel partition are monitored. When the absolute value of the load forecast residuals is greater than a dynamic statistical threshold within a preset time window, the calibration condition is triggered. The dynamic statistical threshold is determined based on the standard deviation of the historical load forecast residuals of the dedicated load forecast model. When the calibration condition is triggered, the deviation propagation model is invoked; the deviation propagation model calculates and generates the forward correction amount for the follower partition based on the recent residual sequence of the sentinel partition and on pre-learned time delay and gain parameters. The forward-looking correction is applied additively to the load growth trend prediction results of the follower partition to generate the final load growth trend prediction results.

[0016] A second aspect of the present invention relates to a regional load growth trend prediction system for extreme weather conditions, implemented using the method described in the first aspect of the present invention; The data preprocessing module is used to acquire and preprocess multi-source historical data from each partition within the target area; An extreme weather event identification module is used to establish an extreme weather event identification mechanism, analyze the preprocessed multi-source historical data to identify and label extreme weather event periods, and generate extreme weather event label sequences. The load response pattern analysis module is used to quantify the load response characteristics of each zone based on the extreme weather event period, and to analyze the load response characteristics using a clustering algorithm to divide each zone into at least two load growth trend categories. A dedicated prediction model construction module is used to construct and implement a hierarchical prediction model, which is used to predict the load for future periods based on the load growth trend category, so as to generate load growth trend prediction results. The dynamic prediction calibration module is used to select sentinel zones within each of the load growth trend categories and designate the remaining zones as follower zones; monitor the load prediction residuals of the sentinel zones in real time, where the load prediction residuals are the difference between the measured load values ​​of the sentinel zones and the predicted values ​​generated by the dedicated load prediction model; when the load prediction residuals trigger preset calibration conditions, calculate the forward-looking correction amount for the follower zones based on the deviation propagation model constructed based on the spatiotemporal correlation between the sentinel zones and the follower zones; and use the forward-looking correction amount to calibrate the original prediction results of the follower zones to generate the final load growth trend prediction results.

[0017] A third aspect of the present invention relates to a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method described in the first aspect of the present invention.

[0018] A fourth aspect of the present invention relates to a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method of the first aspect of the present invention.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention extracts, vectorizes, models, and automatically classifies the load response characteristics of different regions under extreme weather conditions, categorizing regions into response types such as highly sensitive, slowly changing, and rigid. By constructing a dedicated load forecasting model for each region based on clustering results, it effectively overcomes the performance degradation problem of traditional unified forecasting models in dealing with regional differences, thereby achieving accurate modeling of load growth trends in different regions.

[0020] 2. This invention creatively introduces a "sentinel-follower" calibration framework. During extreme weather events, instead of relying on one-off static forecasts, it monitors the prediction residuals of dedicated models in real time through sentinel zones. Once a significant deviation is detected, a forward-looking correction is immediately calculated based on the transmission lag and intensity attenuation patterns of extreme weather in geographic space (i.e., spatiotemporal correlation), and the future predictions of follower zones of the same category are proactively calibrated. This adaptive correction capability effectively utilizes prior information from the first-mover region, enabling the prediction to possess real-time fault tolerance and self-evolution capabilities, thus improving the reliability of predictions under complex and ever-changing extreme weather conditions. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for predicting regional load growth trends under extreme weather conditions. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer and more accurate, the technical solutions of this invention are described in detail below through several specific embodiments. The embodiments used in this invention are merely illustrative and are not intended to limit the scope of this invention.

[0023] Embodiment 1 of the present invention, as follows Figure 1 As shown, this relates to a method for predicting regional load growth trends under extreme weather conditions. The method includes the following steps: Step 1: Acquire and preprocess multi-source historical data for each partition within the target area.

[0024] To ensure the model can fully learn extreme weather patterns, data collection spans the past three years, covering multiple typical extreme weather events (such as typhoons, prolonged heat waves, and cold waves), to ensure the model can fully learn and summarize load response characteristics under extreme patterns. The multi-source historical data includes at least: Historical power load data includes historical power load time series data for each zone within the target area. In order to capture fine-grained load changes, the preferred data sampling frequency is high-frequency sampling (15 min, 30 min, etc.). This embodiment uses 15 min as an example for illustration. Historical meteorological data, meteorological data corresponding to the geographical location of each region, with a data sampling frequency of hourly, and data dimensions including but not limited to temperature, humidity, wind speed, total solar radiation, etc. The system includes supplementary information data for each zone, such as holiday markers (e.g., weekdays, weekends, statutory holidays) and electricity consumption type labels based on the socio-economic attributes of each zone (e.g., the proportion of industrial, commercial, and residential land use), such as industrial-dominated, commercial-center, and residential-oriented. These electricity consumption type labels are numerically processed using one-hot encoding before being input into the model.

[0025] To eliminate noise and errors in the original data and standardize the data format, data preprocessing is performed. Data preprocessing operations include time alignment, missing value imputation, outlier handling, and deduplication. Specifically, this includes: All data is time-aligned based on hourly data. To unify the time granularity, downsampling is performed. Specifically, for power load data with a frequency higher than the hourly level (15 minutes), the arithmetic mean or maximum value of the four data points within each hour can be calculated. In this embodiment, the arithmetic mean is used as the load value for that hour. For data with a frequency lower than the hourly level (such as daily updated holiday markers or static zone electricity consumption type labels), a forward padding or broadcast strategy is adopted. Specifically, the values ​​of daily data (such as holiday markers) are assigned to all 24-hour intervals of that day; the values ​​of static data (such as electricity consumption type labels) are assigned to all their corresponding time series records. For other numerical data with irregular frequencies or requiring interpolation, linear interpolation or nearest neighbor interpolation methods are used to generate hourly data points.

[0026] To ensure data integrity, missing data points in the time series are filled using linear interpolation with 12 valid data points before and after the missing position (24 in total). If 12 valid data points cannot be obtained on one side of the missing position (such as the beginning or end of the sequence), all available valid data points on that side are used for interpolation; if there is no data on one side, the nearest neighbor value on the other side is used for filling.

[0027] Sort the data by timestamp. If duplicate records with the same timestamp are detected, keep only the first record and delete the rest.

[0028] An anomaly detection method based on K-nearest neighbor distance is used to identify abnormal data. Specifically, for each data point in the time series... Within a time window of size N (e.g., N=48), calculate Calculate the distance between the data points and all other data points within the window, and find its Kth nearest neighbor (e.g., K=5), and record this distance as . Set a distance threshold. (This threshold can be obtained by statistically analyzing historical normal data, such as by taking the 99th percentile of all distance values.) At that time, the judgment These are outliers. For identified outliers, linear interpolation is used for filling and correction, which involves replacing the outlier with 12 normal data points before and after it (a total of 24). In scenarios requiring higher accuracy, an ARIMA model trained on normal data segments can be used to predict and fill in outliers.

[0029] To eliminate the influence of different dimensions of data, the min-max method is used for data standardization.

[0030] Step 2: Establish an extreme weather event identification mechanism, analyze the preprocessed multi-source historical data to identify and label the time periods of extreme weather events, and generate extreme weather event label sequences.

[0031] Step 2, the process of generating the extreme weather event label sequence includes: An adaptive event definition mechanism is adopted to calculate dynamic thresholds based on the historical meteorological data; The sliding window detection method is used to determine whether the historical meteorological data for N consecutive days exceeds or falls below the dynamic threshold in order to identify the period of the extreme weather event; in this embodiment, the sliding window size N is set to 3 and the step size is 1.

[0032] The extreme weather event period is encoded using Boolean label vectors. All hourly time points corresponding to the period identified as an extreme weather event are encoded as 1, and the remaining time points are encoded as 0.

[0033] The classification model can also be trained in reverse based on the extreme responses of historical loads, and this model can be used as an event definer.

[0034] Specifically, based on the hourly meteorological data preprocessed in step 1, the meteorological data collected daily are merged, and the aggregation functions of the maximum and minimum values ​​are used to obtain the daily index sequence (e.g., daily maximum temperature, daily minimum temperature, and daily maximum wind speed).

[0035] An adaptive event definition mechanism is employed to overcome the limitations of traditional fixed threshold methods in adapting to regional and seasonal differences. This includes: High-temperature events: First, obtain the daily maximum temperature data of the corresponding season (e.g., summer, June-August for temperate regions in the Northern Hemisphere) for the historical W years (e.g., W=10) of the partition, and calculate its 95th percentile; when the daily maximum temperature exceeds the 95th percentile temperature value of the partition itself for three consecutive days covered by the sliding window, it is determined to be a high-temperature event. Cold wave event: First, obtain the daily minimum temperature data of the corresponding season (e.g., for temperate regions of the Northern Hemisphere, winter December-February) of the historical W years of the region, and calculate its 5th percentile; when the daily minimum temperature covered by the sliding window is lower than the 5th percentile temperature value of the region itself for three consecutive days, it is determined to be a cold wave event. Extreme gale event: First, obtain the daily maximum wind speed data of the region over the past W years and calculate its 95th percentile; when the daily maximum wind speed reaches or exceeds the 95th percentile wind speed value of the region itself for three consecutive days covered by the sliding window, it is determined to be an extreme gale event.

[0036] When any of the conditions are met, the date period covered by the sliding window (e.g., all three days) is collectively determined to be an extreme weather event.

[0037] As an alternative preferred implementation, the present invention also provides a machine learning-based event definition method, which does not require preset quantiles but is generated data-driven. Optionally, extreme weather events can be dynamically generated using a machine learning model based on load impact. This method first calculates daily load peaks by analyzing historical load data. and its daily peak increment .Will The top 5% of dates (d) are labeled as positive samples (Y=1, i.e., extreme response days), and the remaining dates are labeled as negative samples (Y=0). An input feature vector X is constructed using the aggregated daily meteorological data (including daily high temperature, daily low temperature, daily average humidity, daily maximum wind speed, etc.) of date d and the two days preceding it (a total of 3 days). A classification model (such as a decision tree, support vector machine, or random forest) is trained using X as the feature and Y as the label. The trained model itself constitutes an adaptive event definer. Its internal decision path (e.g., temperature > 34℃ and humidity > 75%) constitutes the adaptive recognition rule. In application, the aggregated meteorological data for the three consecutive days to be analyzed is input into the trained model. If the model determines Y=1, then the center or end date corresponding to that window is identified as the extreme weather occurrence day, and the window containing that date is labeled as an extreme weather event period.

[0038] Boolean label vectors are used to encode the regional time periods of extreme weather events in the sample sequence. All hourly time points corresponding to these date periods are encoded as 1, while all other time points are assigned a value of 0. This label sequence will serve as a key input feature for subsequent models.

[0039] Step 3: Based on the extreme weather event period, quantify the load response characteristics of each zone, and use a clustering algorithm to analyze the load response characteristics to divide each zone into at least two load growth trend categories.

[0040] In step 3, the load response characteristics of each zone are quantified based on the duration of the extreme weather event, including: Calculate peak growth rate The peak growth rate is used to measure the increase in peak load during the extreme weather event period relative to the average daily load during normal periods; the average daily load during normal periods is calculated as the arithmetic mean of all load sampling points over M consecutive days prior to the start of the extreme weather event period; the specific expression is:

[0041] in, This indicates the peak load during extreme weather events. This represents the average load over normal dates, specifically the arithmetic mean of all load sampling points over the seven consecutive days (168 hours) preceding the start of the extreme weather event. If other extreme weather records are included within this preceding seven-day period, the count is shifted backward until a normal period without extreme weather markers is selected.

[0042] Calculate the peak-to-valley ratio The peak-to-valley ratio is used to measure the ratio between the maximum and minimum load values ​​during the extreme weather event period. It can measure the degree of load fluctuation in a region during extreme weather events. The larger the value, the stronger the difference between day and night or between warm and cold periods. The specific expression is as follows:

[0043] in, This indicates the maximum load during the period of an extreme weather event. This indicates the minimum load during the period of an extreme weather event.

[0044] Calculate the standard deviation fluctuation, which measures the standard deviation of all load sampling points during the period of the extreme weather event. The specific expression is:

[0045] in, express Load value at any given time denoted as the average load, and n represents the total number of load sampling points during extreme weather events.

[0046] Next, the K-means clustering algorithm is used to cluster the regional response features to determine the cluster to which each region belongs. Subsequently, by analyzing the features of the cluster center vectors, semantic annotation is performed on each cluster, thus ultimately assigning each region a corresponding load growth trend category.

[0047] Specifically, using the constructed response feature vectors of all N regions The range of i is from 1 to N. These are the peak growth rate, peak-to-trough ratio, and standard deviation fluctuation, respectively. This represents a three-dimensional real space. In this matrix, each row represents a sample (i.e., the response of a partition to a certain event), and each column represents a feature dimension.

[0048] Standardize each column of the matrix. Calculate the mean and standard deviation of each of the three columns, denoted as . , , , , , .in, The mean of the standard deviation volatility across all partitions. This represents the statistical standard deviation of the standard deviation volatility index for all partitions.

[0049] For each sample in the dataset Apply the following transformation to generate a new standardized vector. .

[0050]

[0051] Clustering of regional response features using the K-means clustering algorithm includes: The objective function is:

[0052] In the formula, The set of feature vectors in the cluster. Let be the cluster center vector. Number the clusters. This indicates the computation of an eigenvector. and the cluster center vector The Euclidean distance between them The function is minimized. For the cluster center vector... The selection of cluster centers is performed using the K-means++ method. That is, after randomly selecting the first cluster center, the sample point that is farthest from the existing cluster center is selected as the next cluster center, and so on until the initialization is completed. Then, the standard K-means algorithm is used for iterative calculation.

[0053] In this embodiment, the number of clusters K is set to 3. In practical applications, the value of K can be adaptively determined using the elbow rule or the profile coefficient.

[0054] After clustering is completed, the cluster center vectors are analyzed. The features are used to semantically label the clusters. For example, if the cluster exhibits a high peak growth rate and a high standard deviation, it is named "highly sensitive"; if the values ​​of the peak growth rate, peak-to-valley ratio, and standard deviation are generally low, it is named "rigid"; if the features are between the two cases, it is named "gradually changing".

[0055] Step 4: Construct and implement a hierarchical prediction model to predict the load for future periods based on the load growth trend category, so as to generate load growth trend prediction results.

[0056] Step 4 involves constructing and implementing a hierarchical prediction model, specifically including: Obtain weather forecast data and holiday markers for the target area for future time periods (e.g., the next 72 hours); The meteorological forecast data for the future period and the holiday markers are input into a pre-trained situation classification model to obtain the predicted load growth situation category; for example, it is determined to be "highly sensitive".

[0057] Based on the predicted load growth trend category, select the corresponding dedicated load forecasting model (e.g., highly sensitive LSTM) from the dedicated load forecasting models corresponding to different load growth trend categories. The recent historical feature sequence is input into the selected dedicated load forecasting model to initialize the hidden state of the dedicated load forecasting model. Weather forecast data for future periods corresponding to the target forecast period and the holiday markers are used as the driving input sequence to output the load growth trend forecast result, i.e., a power load forecast sequence containing both time and numerical dimensions (e.g., a MW-level load curve for the next 24 hours). It should be noted that the model's direct output is usually a normalized value, which needs to be denormalized to obtain the final load forecast value with physical units (MW), for example: [120MW, 125MW, 130MW, 145MW].

[0058] The recent historical feature sequence is a feature sequence of a preset length (e.g., 24 hours), including: discrete wavelet data obtained by decomposing historical load data, and historical meteorological data and historical holiday markers mentioned in step 1.

[0059] In this embodiment, based on a pre-trained hierarchical prediction model, the system includes a situation classification model and a dedicated load prediction model.

[0060] The situation classification model (Model A) is a pre-trained classifier (such as SVM or Random Forest). The model is obtained by using historical meteorological data (such as aggregated features of daily maximum and minimum temperatures in step 2) as training features and the situation category label determined in step 3 (such as "highly sensitive") as the training target.

[0061] The dedicated load forecasting model (Model B) is a set (e.g., three) of pre-trained time-series forecasting models (e.g., LSTM) employing an encoder-decoder architecture. This set of models divides historical data into subsets based on the load growth trend categories identified in step 3, and each subset is trained independently. During Model B training, to extract refined features, discrete wavelet transform (e.g., db4 wavelet, layer J=5) can be used to perform multi-scale decomposition of the historical load data for each subset, obtaining the decomposed low-frequency approximate components. and various high-frequency detail components The encoder receives the recent historical feature sequence ( , The historical weather and holiday features from step 1 are used as inputs to learn the historical dynamics of the load and generate a context vector. In this embodiment, these components can be input to the encoder as multi-channel features in parallel, or after dimensionality reduction through a feature fusion layer. The decoder receives weather forecast data for future periods, the holiday markers (and other known future features from steps 1 and 2) as the driving input sequence.

[0062] The LSTM network can be composed of an input layer, a single-layer LSTM hidden layer (containing 25 hidden units), a fully connected layer, and a regression layer, and is trained using the Adam optimizer and the MSE loss function.

[0063] Step 5: Select sentinel zones within each of the load growth trend categories, and designate the remaining zones as follower zones; monitor the load prediction residuals of the sentinel zones in real time, where the load prediction residuals are the difference between the measured load value of the sentinel zone and the load growth trend prediction result; when the load prediction residuals trigger preset calibration conditions, calculate the forward-looking correction amount of the follower zones based on the deviation propagation model constructed based on the spatiotemporal correlation between the sentinel zones and the follower zones; use the forward-looking correction amount to calibrate the load growth trend prediction result of the follower zones, and generate the final load growth trend prediction result.

[0064] In this context, the sentinel partition, within a specific load growth category, is where real-time load data is used as a benchmark to calibrate the real-time performance of the category-specific forecasting model. The follower partition's load forecast curve represents all other partitions except the sentinel partition, and it serves as the target for correction signals. The real-time deviation of the forecasting model is transformed into forward-looking corrections for partitions within the same category, thereby improving the accuracy and timeliness of the forecasting system in dynamic events.

[0065] First, before the occurrence of extreme weather events, each load growth trend category defined in step 4 is analyzed. Within (such as "highly sensitive" or "rigid"), a dynamic selection process for sentinel partitions is executed independently to objectively select the partition that best represents the real-time response characteristics of that category.

[0066] Step 5 involves selecting the sentry zones, specifically including: The response sensitivity metric is determined by linearly weighting the peak growth rate, peak-to-valley ratio, and standard deviation fluctuation quantified in step 3. The corresponding weights can be 0.3, 0.3, and 0.4, respectively.

[0067] The data quality metric is determined by comprehensively evaluating the integrity rate and average transmission delay of the historical power load data and historical meteorological data using a preset scoring function. For example, the scoring function can be a normalized weighted summation function, where the integrity rate (e.g., 99.5%) is directly used as the first score, and the average transmission delay (e.g., 2 seconds) is calculated as the second score using an inverse function (e.g., 1 / (1 + delay in seconds)). The first and second scores are then weighted and summed to obtain the final data quality metric. The corresponding weights can be 0.5 and 0.5, respectively.

[0068] Determine the spatiotemporal leading measure and adaptively adjust it according to the type of the extreme weather event period, including: First, the meteorological forecast information corresponding to the extreme weather events is analyzed. If the forecast contains keywords with clear movement paths, such as typhoons and fronts, it is determined to be a moving weather event; if the forecast contains broad and homogeneous keywords such as high-temperature warnings and cold waves, it is determined to be a covering weather event. It should be noted that, in addition to judgment based on text keywords, automatic judgment can also be made based on the spatiotemporal distribution characteristics of meteorological data: for example, the cross-correlation delay of meteorological data in each region is calculated; if the standard deviation of the delay exceeds a preset threshold, it is determined to be a moving event; otherwise, it is determined to be a covering event.

[0069] If the extreme weather event corresponds to a moving weather event, the score is determined based on the geographical order of the zones within the meteorological forecast path, with the zone most affected first receiving the highest score. The meteorological forecast path is defined as a sequence of geographic coordinates including timestamps. The minimum distance between the forecast path and the geographic location of each zone is calculated, and the forecast timestamp corresponding to this minimum distance is identified, i.e., the first impact time. The metric is calculated as the reciprocal of the sum of this first impact time and a preset minimum positive number. This preset minimum positive number ensures that the denominator is not zero, thus guaranteeing that the zone most affected first receives the highest score.

[0070] If the extreme weather event period corresponds to a widespread weather event, the score is determined based on the time delay calculated by cross-correlation analysis of the historical power load data, with the zone that responds most strongly receiving the highest score. Specifically, first, the average load curve of all zones within the load growth trend category is calculated; then, the cross-correlation function between the load curve of each zone and the average load curve is calculated, and the time delay at which the correlation is highest is identified. The metric is calculated as the negative of this time delay, thus ensuring that the zone that responds most strongly receives the highest score.

[0071] The response sensitivity metric, the data quality metric, and the spatiotemporal leadership metric are weighted and summed to obtain the comprehensive qualification score, which is expressed as follows:

[0072] in, This refers to any partition within the load growth trend category c. For partitioning In terms of response sensitivity, data quality, and spatiotemporal leadership metrics The weighting coefficients for each dimension can be 0.3, 0.3, and 0.4, respectively. This is the normalization function.

[0073] Based on the comprehensive qualification score, one or more partitions with the highest scores in each load growth trend category are identified as sentinel partitions. This ensures that the selected sentinel partitions are optimally representative in terms of data quality and response leadership, thereby providing the most reliable source of deviation signals for subsequent calibration.

[0074] Secondly, during extreme weather events, a real-time monitoring and calibration phase is initiated.

[0075] Step 5, the process of generating the final load growth trend prediction result, includes: The load forecast residuals of the sentinel partition are monitored. When the absolute value of the load forecast residuals is greater than a dynamic statistical threshold within a preset time window, the calibration condition is triggered. The dynamic statistical threshold is determined based on the standard deviation of the historical load forecast residuals of the dedicated load forecast model. When the calibration condition is triggered, the deviation propagation model is invoked; the deviation propagation model calculates and generates the forward correction amount for the follower partition based on the recent residual sequence of the sentinel partition and on pre-learned time delay and gain parameters. The forward-looking correction is applied additively to the load growth trend prediction results of the follower partition to generate the final load growth trend prediction results.

[0076] Specifically, all sentry zones will be divided at high frequency. Real-time load data , and the corresponding category in step 4 Exclusive Model Generated Predictions By comparing the results, the load forecast residual sequence can be calculated. :

[0077] To determine whether the residual constitutes a systematic bias requiring calibration, a dynamic statistical threshold is set. (For example, based on twice the standard deviation of historical residuals). When the absolute value of the residuals of any sentinel partition... Duration exceeding the preset time window (For example =15 minutes) still greater than When this happens, it will automatically trigger an action targeting its category. The collaborative calibration procedure.

[0078] Once the calibration procedure is triggered, the pre-trained bias propagation model for that category is immediately invoked. By performing cross-correlation analysis on the residual sequences of historical sentinel and follower partitions of the same category, and identifying the time offset corresponding to the peak of the correlation coefficient, the time lag can be determined. The residual sequence of the follower partition is used as the dependent variable, and the residual sequence of the sentinel partition after time lag adjustment is used as the independent variable. The gain is determined by linear regression analysis or nonlinear fitting (such as multinomial regression or kernel ridge regression). In this embodiment, to balance computational efficiency, linear regression coefficients are preferably used as the gain G. The bias propagation model is based on the recent residual sequence of the current sentinel partition. ,in This represents the current and past time step indices. The pre-learned time delay is then applied. With gain Calculate and generate partitions for all other followers within that category. Forward-looking correction curve :

[0079] Subsequently, this forward-looking correction curve is applied additively in real time to the original prediction curve of the corresponding follower partition to generate and output the final calibrated prediction result. It can be represented by the following formula:

[0080] in, Index the time step for future predictions (e.g., for hourly predictions, h=1,2,...,24 represents the 1st, 2nd,...24th hour in the future, respectively). It is the category in step 4 Dedicated prediction model The generated original predicted values. This represents the correction amount for follower partition j at time t+h.

[0081] To ensure calibration accuracy, the entire process strictly adheres to the category isolation principle: calibration operations triggered by sentinels of a specific category only affect the follower partitions within that category and will not interfere with the prediction results of other categories. This method upgrades static prediction capabilities into a dynamically adaptive intelligent prediction system. The calibration mechanism in this step gives the method online fault tolerance; even if the dedicated model in step 4 experiences prediction deviations due to unknown unforeseen factors, it can quickly capture and proactively correct these deviations, improving the reliability and robustness of predictions in real-world, extreme weather scenarios.

[0082] The forecast results are output as heat maps and trend charts for dispatching assistance. The forecast results can be integrated into the control center, load management platform, or power supply station operation and maintenance terminal to proactively formulate measures to cope with high-temperature load surges. Regional early warnings or control suggestions are automatically generated based on load growth trends. An early warning is triggered when the predicted load growth rate exceeds a set safety threshold. This safety threshold can be empirically set by power grid operation experts based on factors such as power grid redundancy and historical peak load in each region. After an early warning is triggered, the system automatically generates corresponding control suggestions based on a preset expert rule base. This expert rule base contains multiple rules summarized from the experience of power grid operation experts. For example, two warning levels can be set: Level 1 (orange) and Level 2 (red), each corresponding to different load growth rate thresholds. Examples include: "When the predicted load growth rate exceeds 15% in the next two hours, it is recommended to dispatch backup power"; "When a Level 2 warning is triggered and the regional peak-shaving capacity is less than 20%, it is recommended to provide cross-regional support," etc. The system automatically matches the rules according to the triggering conditions and outputs corresponding control measures.

[0083] Example 2: A regional load growth trend prediction system for extreme weather conditions, implemented using the method described in Embodiment 1 of this invention; comprising: The data preprocessing module is used to acquire and preprocess multi-source historical data from each partition within the target area; An extreme weather event identification module is used to establish an extreme weather event identification mechanism, analyze the preprocessed multi-source historical data to identify and label extreme weather event periods, and generate extreme weather event label sequences. The load response pattern analysis module is used to quantify the load response characteristics of each zone based on the extreme weather event period, and to analyze the load response characteristics using a clustering algorithm to divide each zone into at least two load growth trend categories. A dedicated prediction model construction module is used to construct and implement a hierarchical prediction model, which is used to predict the load for future periods based on the load growth trend category, so as to generate load growth trend prediction results. The dynamic prediction calibration module is used to select sentinel zones within each of the load growth trend categories and designate the remaining zones as follower zones; monitor the load prediction residuals of the sentinel zones in real time, where the load prediction residuals are the difference between the measured load values ​​of the sentinel zones and the predicted values ​​generated by the dedicated load prediction model; when the load prediction residuals trigger preset calibration conditions, calculate the forward-looking correction amount for the follower zones based on the deviation propagation model constructed based on the spatiotemporal correlation between the sentinel zones and the follower zones; and use the forward-looking correction amount to calibrate the original prediction results of the follower zones to generate the final load growth trend prediction results.

[0084] Example 3: A terminal includes a processor and a storage medium; the storage medium is used to store instructions; the processor is used to perform operations according to the instructions to execute the steps of the method described in Embodiment 1 of the present invention.

[0085] Example 4: A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method described in Embodiment 1 of the present invention.

[0086] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0087] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0088] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0089] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for predicting regional load growth trends under extreme weather conditions, characterized in that, The method includes the following steps: Step 1: Acquire and preprocess multi-source historical data for each partition within the target area; Step 2: Establish an extreme weather event identification mechanism, analyze the preprocessed multi-source historical data to identify and label the time periods of extreme weather events, and generate extreme weather event label sequences; Step 3: Based on the extreme weather event period, quantify the load response characteristics of each zone, and use a clustering algorithm to analyze the load response characteristics to divide each zone into at least two load growth trend categories. Step 4: Construct and implement a hierarchical prediction model to predict the load for future periods based on the load growth trend category, so as to generate load growth trend prediction results; Step 5: Select sentinel zones within each of the load growth trend categories, and designate the remaining zones as follower zones; monitor the load prediction residuals of the sentinel zones in real time, where the load prediction residuals are the difference between the measured load value of the sentinel zone and the load growth trend prediction result; when the load prediction residuals trigger preset calibration conditions, calculate the forward-looking correction amount of the follower zones based on the deviation propagation model constructed based on the spatiotemporal correlation between the sentinel zones and the follower zones; use the forward-looking correction amount to calibrate the load growth trend prediction result of the follower zones, and generate the final load growth trend prediction result.

2. The method for predicting regional load growth trends under extreme weather conditions according to claim 1, characterized in that: In step 1, the multi-source historical data includes: historical power load data, historical meteorological data corresponding to each zone, and supplementary information data for each zone; The historical meteorological data includes at least one of temperature, humidity, wind speed, and total solar radiation; The supplementary information data for the partition includes holiday markers and electricity consumption type labels.

3. The method for predicting regional load growth trends under extreme weather conditions according to claim 2, characterized in that: Step 2, the process of generating the extreme weather event label sequence includes: An adaptive event definition mechanism is adopted to calculate dynamic thresholds based on the historical meteorological data; The sliding window detection method is used to determine whether the historical meteorological data for N consecutive days exceeds or falls below the dynamic threshold in order to identify the period of the extreme weather event; The extreme weather event period is encoded using Boolean label vectors. All hourly time points corresponding to the period identified as an extreme weather event are encoded as 1, and the remaining time points are encoded as 0.

4. The method for predicting regional load growth trends under extreme weather conditions according to claim 3, characterized in that: In step 3, the load response characteristics of each zone are quantified based on the duration of the extreme weather event, including: Calculate the peak growth rate, which measures the increase in peak load during the extreme weather event period relative to the average daily load during normal periods; the average daily load during normal periods is calculated as the arithmetic mean of all load sampling points over M consecutive days prior to the start of the extreme weather event period. Calculate the peak-to-valley ratio, which measures the ratio between the maximum load and the minimum load during the period of the extreme weather event; Calculate the standard deviation fluctuation, which measures the standard deviation of all load sampling points during the period of the extreme weather event.

5. The method for predicting regional load growth trends under extreme weather conditions according to claim 4, characterized in that: Step 4 involves constructing and implementing a hierarchical prediction model, specifically including: Obtain weather forecast data and holiday markers for the target area for future periods; The meteorological forecast data for the future period and the holiday markers are input into a pre-trained situation classification model to obtain the predicted load growth situation category. Based on the predicted load growth trend category, select the corresponding dedicated load prediction model; The recent historical feature sequence is input into the selected dedicated load forecasting model to initialize the hidden state of the dedicated load forecasting model. The meteorological forecast data for the future period and the holiday markers are used as the driving input sequence to obtain the load growth trend forecast result.

6. The method for predicting regional load growth trends under extreme weather conditions according to claim 5, characterized in that: Step 5 involves selecting the sentry zones, specifically including: The response sensitivity metric is determined by performing a linear weighted combination of the peak growth rate, peak-to-valley ratio, and standard deviation fluctuation quantified in step 3. The data quality metric is determined by comprehensively evaluating the integrity rate and average transmission delay of the historical power load data and the historical meteorological data using a preset scoring function. Determine the spatiotemporal leading measure and adaptively adjust it according to the type of the extreme weather event period, including: If the extreme weather event period corresponds to a mobile weather event, it is determined according to the geographical order of the regions in the meteorological forecast path, with the region that is affected first receiving the highest score. If the extreme weather event period corresponds to a coverage weather event, it is determined based on the time delay calculated by cross-correlation analysis of the historical power load data, with the partition that responds most advanced receiving the highest score. The response sensitivity metric, the data quality metric, and the spatiotemporal leadership metric are weighted and summed to obtain the comprehensive qualification score. Based on the comprehensive qualification score, one or more partitions with the highest scores in each load growth trend category are identified as the sentinel partitions.

7. The method for predicting regional load growth trends under extreme weather conditions according to claim 6, characterized in that: Step 5, the process of generating the final load growth trend prediction result, includes: The load forecast residuals of the sentinel partition are monitored. When the absolute value of the load forecast residuals is greater than a dynamic statistical threshold within a preset time window, the calibration condition is triggered. The dynamic statistical threshold is determined based on the standard deviation of the historical load forecast residuals of the dedicated load forecast model. When the calibration condition is triggered, the deviation propagation model is invoked; the deviation propagation model calculates and generates the forward correction amount for the follower partition based on the recent residual sequence of the sentinel partition and on pre-learned time delay and gain parameters. The forward-looking correction is applied additively to the load growth trend prediction results of the follower partition to generate the final load growth trend prediction results.

8. A regional load growth trend prediction system for extreme weather conditions, comprising a regional load growth trend prediction method for extreme weather conditions as described in any one of claims 1-7, characterized in that: The data preprocessing module is used to acquire and preprocess multi-source historical data from each partition within the target area; An extreme weather event identification module is used to establish an extreme weather event identification mechanism, analyze the preprocessed multi-source historical data to identify and label extreme weather event periods, and generate extreme weather event label sequences. The load response pattern analysis module is used to quantify the load response characteristics of each zone based on the extreme weather event period, and to analyze the load response characteristics using a clustering algorithm to divide each zone into at least two load growth trend categories. A dedicated prediction model construction module is used to construct and implement a hierarchical prediction model, which is used to predict the load for future periods based on the load growth trend category, so as to generate load growth trend prediction results. The dynamic prediction calibration module is used to select sentinel partitions within each of the load growth trend categories and determine the remaining partitions as follower partitions; and to monitor the load prediction residuals of the sentinel partitions in real time, wherein the load prediction residuals are the difference between the measured load value of the sentinel partition and the predicted value generated by the dedicated load prediction model. When the load prediction residual triggers the preset calibration condition, the forward correction amount of the follower partition is calculated based on the deviation propagation model constructed based on the spatiotemporal correlation between the sentinel partition and the follower partition. By using forward-looking corrections, the original forecast results of the follower partition are calibrated to generate the final load growth trend forecast results.

9. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-7.

Images