A Method for Constructing an Intelligent Agent for Power Load Forecasting Based on a Large Model

Abstract

This invention relates to the field of power systems, and more particularly to a method for constructing an intelligent agent for power load forecasting based on a large model. The method involves determining a fixed window and its corresponding fixed features, a sliding window and its corresponding sliding features, analyzing the probability of occurrence and the probability of identical features, calculating the sliding influence coefficient, classifying power influence tendencies, setting labels to be analyzed, and analyzing the power load curve and feature change curve for the sliding features with the assigned labels to determine the lag time. The method then calculates the offset influence feature value based on the sliding influence coefficient, sets corresponding labels, determines whether the sliding features need pre-analysis, obtains power load propagation keywords, calculates feature similarity, determines whether power reserves should be implemented, and completes the power load forecasting. This invention, by pre-analyzing power load data and specifically analyzing fixed and sliding features, provides a data benchmark for subsequent power load forecasting, improving forecast accuracy.

Description

Technical Field

[0001] This invention relates to the field of power systems, and in particular to a method for constructing an intelligent agent for power load forecasting based on a large model. Background Technology

[0002] Power load forecasting agents can accurately predict future electricity demand, playing a crucial role in ensuring power supply security, improving economic efficiency, and promoting the green energy transition. Their construction methods have evolved from traditional time series models and regression analysis that rely on mathematical assumptions, to classic machine learning that can capture nonlinear relationships, and then to deep learning models that are adept at uncovering long-term complex time series dependencies. Modern intelligent agents are no longer just single algorithms, but comprehensive AI systems that integrate big data processing, automated feature engineering, hyperparameter optimization, and model integration technologies. Their core development path is to continuously introduce more powerful models and automation technologies to achieve more accurate and efficient automated forecasting.

[0003] Chinese Patent Publication No. CN120031335A discloses a method for constructing an accurate power load prediction model using big data. The method includes the following steps: First, using IoT devices such as smart meters and sensors, multi-source data on power operation, user information, and weather are collected every 15 minutes. Abnormal data is cleaned using rule-based algorithms and encrypted before being transmitted to a distributed storage center. Next, information is retrieved from the stored data, and features strongly correlated with the load are selected using the Pearson correlation coefficient and normalized. Then, a deep learning model combining LSTM and CNN is constructed, hyperparameters are set, and the dataset is proportionally divided for training, ensuring the RMSE index on the test set is kept below 0.1, resulting in an accurate prediction model for subsequent load prediction. This invention can accurately predict power load, contributing to the stable operation of the power system, making power generation plans more rational, reducing energy waste, improving power supply reliability, and also reducing costs for enterprises and providing users with a better experience.

[0004] Chinese Patent Publication No. CN120320288A discloses a method for constructing and applying a power load probability prediction model, comprising: constructing a power load probability prediction model and training the power load probability prediction model using a training sample set; the power load probability prediction model includes: a Mamba encoder and a Gaussian process decoder; the Mamba encoder is used to capture the temporal dependencies of the input training samples to obtain the corresponding feature vectors; the Gaussian process decoder is used to model the interrelationships between the features in the feature vectors to obtain the corresponding covariance matrix and mean; the Gaussian distribution composed of the covariance matrix and the mean is used as the power load probability distribution for the next time step of the prediction; the model of this invention achieves high-precision, scalable and robust load prediction by efficiently processing long-sequence data, capturing complex temporal dependencies and providing interpretable uncertainty quantification, and is suitable for the complex needs of modern power grids.

[0005] It is evident that the existing technology still has the following problems:

[0006] Feature selection is a crucial step in the construction of intelligent agents, and its quality directly affects the performance and generalization ability of the model. If the feature selection is inappropriate or the selected data is abnormal, it will not only waste computing and storage resources, but also introduce model bias, leading to distorted prediction results and thus reducing prediction accuracy. Summary of the Invention

[0007] To address this issue, the present invention provides a method for constructing an intelligent agent for power load forecasting based on a large model. This method aims to overcome the problem that improper feature selection or abnormal data selection during the agent construction process not only wastes computing and storage resources but also introduces model bias, leading to distorted prediction results and reduced prediction accuracy.

[0008] To achieve the above objectives, this invention provides a method for constructing an intelligent agent for power load forecasting based on a large model, comprising:

[0009] Historical periodic data of power load is acquired, the historical periodic data is analyzed to obtain a fixed window and corresponding fixed features, a sliding window and corresponding sliding features are obtained, and the fixed features and the fixed keywords and sliding keywords corresponding to the sliding features are determined.

[0010] Application labels are set for the fixed features. In response to the determination result of the sliding features, the occurrence probability and the same probability of the sliding features are analyzed, the sliding influence coefficient is calculated to classify the power influence tendency of the sliding features, and the labels to be analyzed are set for the sliding features.

[0011] For the sliding feature set as the label to be analyzed, the corresponding power load curve and feature change curve are extracted, the two curves are analyzed to determine the lag time, and the offset influence feature value of the power load is calculated in combination with the sliding influence coefficient. The corresponding label is set to determine whether the sliding feature needs to be pre-analyzed.

[0012] The power load propagation characteristics and corresponding power load propagation keywords are obtained, compared with the fixed keywords and sliding keywords, and the feature similarity is calculated. Combined with the fixed features and sliding features set as application tags, power reserves are determined to complete the prediction of power load.

[0013] Furthermore, the process of obtaining a fixed window and its corresponding fixed features, and obtaining a sliding window and its corresponding sliding features, includes:

[0014] Construct characteristic change curves in the power transmission process and place each characteristic change curve in the same coordinate system;

[0015] Identify characteristics that exhibit the same trend of change within the same time period;

[0016] The same time period is defined as a fixed window, and the feature is a fixed feature;

[0017] The features on the feature change curve are identified as sliding features;

[0018] The time window corresponding to the maximum sliding feature is determined as the sliding window.

[0019] Furthermore, the process of analyzing the probability of occurrence of the sliding feature and the same probability includes,

[0020] Determine the number of times the sliding feature appears on the same feature change curve;

[0021] The ratio of the number of occurrences to the total number of features is determined as the first occurrence probability;

[0022] The average of the sum of the first occurrence probabilities of each of the aforementioned feature change curves is determined as the occurrence probability;

[0023] Determine the number of feature change curves exhibiting the sliding characteristic;

[0024] The ratio of the number of curves to the total number of curves representing the characteristic change is determined to be the same probability.

[0025] Furthermore, the process of calculating the sliding influence coefficient includes,

[0026] The ratio of the occurrence probability to the baseline occurrence probability is determined as the first influencing factor;

[0027] The ratio of the same probability to the benchmark same probability is determined as the second influencing factor;

[0028] The weighted sum of the first influence factor and the second influence factor is determined to be the sliding influence coefficient.

[0029] Furthermore, the step of classifying the electric field influence tendency of the sliding feature involves assigning a label to be analyzed to the sliding feature, wherein...

[0030] If the sliding influence coefficient is greater than the sliding influence coefficient threshold, then the electric influence tendency of the sliding feature is determined to be a strong electric influence tendency, and a label to be analyzed is set for the sliding feature;

[0031] If the sliding influence coefficient is less than or equal to the sliding influence coefficient threshold, then the electric influence tendency of the sliding feature is determined to be a weak electric influence tendency, and no label is set for the sliding feature.

[0032] Furthermore, the process of analyzing the two curves to determine the lag duration includes,

[0033] Place the power load curve and the characteristic change curve in the same coordinate system;

[0034] Determine the sliding feature and the fixed feature corresponding to the label to be analyzed;

[0035] Determine the power start time point of the power load curve corresponding to each of the sliding features and the fixed features;

[0036] Determine the feature start time point of the feature change curve corresponding to each sliding feature and the fixed feature;

[0037] Determine several differences between the power start time point and the characteristic start time point;

[0038] The average value of each of the aforementioned differences is determined as the lag time.

[0039] Furthermore, the process of calculating the characteristic value of the offset effect of the power load includes,

[0040] The ratio of the lag duration to the baseline lag duration is determined as the lag effect factor;

[0041] The ratio of the sliding influence coefficient to the benchmark sliding influence coefficient is determined as the sliding influence factor;

[0042] The weighted sum of the lag effect factor and the sliding effect factor is determined to be the offset effect characteristic value.

[0043] Furthermore, the setting of corresponding labels is used to determine whether the sliding features need to be pre-analyzed, wherein,

[0044] If the offset influence feature value is greater than the offset influence feature value threshold, then an application label is set to determine that the sliding feature needs to be pre-analyzed.

[0045] If the offset influence feature value is less than or equal to the offset influence feature value threshold, then no label is set, and it is determined that the sliding feature does not need to be pre-analyzed.

[0046] Furthermore, the process of calculating feature similarity includes,

[0047] The similarity between the power load propagation keywords and the fixed keywords is calculated as the first similarity.

[0048] The similarity between the power load propagation keywords and the sliding keywords is calculated as the second similarity.

[0049] The average of the sum of the first similarity and the second similarity is determined as the feature similarity.

[0050] Furthermore, the determination to conduct power storage, wherein,

[0051] If the power load propagation characteristics meet the preset conditions, then it is determined to perform power storage;

[0052] If the power load propagation characteristics do not meet the preset conditions, then it is uncertain whether to perform power storage;

[0053] The preset condition is that the similarity is greater than the feature similarity threshold and the label of the power load propagation feature is the application label.

[0054] Compared with existing technologies, this invention determines a fixed window and its corresponding fixed features, a sliding window and its corresponding sliding features, analyzes the probability of occurrence and the probability of identical features, calculates the sliding influence coefficient, classifies power influence tendencies, sets labels to be analyzed, and analyzes the power load curve and feature change curve for the sliding features set as labels to be analyzed to determine the lag time. Combined with the sliding influence coefficient, the offset influence feature value is calculated, corresponding labels are set, it is determined whether the sliding features need to be pre-analyzed, power load propagation keywords are obtained, feature similarity is calculated, and power reserves are determined to complete the prediction of power load. This invention, by pre-analyzing power load data and specifically analyzing fixed and sliding features, provides a data benchmark for subsequent power load prediction, improving prediction accuracy.

[0055] In particular, by dividing the window into fixed and sliding windows, fixed and sliding features are determined, providing a theoretical basis for subsequent feature-specific analysis. In practice, feature extraction often heavily relies on experts' prior knowledge and manual interpretation, requiring significant manpower and time for data retrieval and pattern recognition, leading to low model development efficiency and resource utilization. Furthermore, it introduces subjective bias; due to differences in the experience domains and judgment standards of different experts, different interpretations of the same data pattern may occur, making it difficult to guarantee the repeatability and objectivity of the feature extraction process. Moreover, this subjectivity and uncertainty directly lead to representational errors between the extracted feature set and the objectively existing real patterns in the data. When these features, which fail to accurately reflect the actual power load driving mechanism, are input into the prediction model, it causes "distortion" in the model's cognition, resulting in prediction results deviating from reality. This significantly weakens the model's prediction accuracy, robustness, and generalization ability, making it difficult to effectively support precise power dispatch decisions. Therefore, this invention considers targeted feature analysis, determining fixed and sliding features to provide a theoretical data benchmark for subsequent calculation of the sliding influence coefficient, thereby improving prediction accuracy.

[0056] In particular, by quantifying the probability of occurrence of the sliding feature and calculating its sliding influence coefficient, it is found that in reality, power load sequences exhibit significant temporal heterogeneity. That is, each power consumption cycle is affected by unique internal and external factors, thus exhibiting unique load characteristics and driving mechanisms. If the predictive agent mechanically relies solely on the statistical regularity of historical features while ignoring potential specific disturbance events within a particular cycle, such as concerts, competitions, sudden temperature rises and falls, it will lead to serious misjudgments. When the load is underestimated, it will result in insufficient reserve capacity, threatening grid security. When the load is overestimated, it will lead to excessive reservation and idle power generation resources, increasing system operating costs and causing structural irrationality in resource allocation. In extreme cases, it may even directly manifest as a supply-demand imbalance. Based on this, the present invention considers analyzing the probability of occurrence of the sliding feature in historical cycles and its probability of occurrence in the current cycle to calculate the sliding influence coefficient, thereby achieving quantitative assessment and accurate response to load disturbance events and improving prediction accuracy.

[0057] In particular, by analyzing the lag duration and combining it with the sliding effect coefficient to calculate the characteristic value of the offset impact of power load, the potential impact magnitude and urgency of the sliding characteristics on power load are quantified. In actual power grid operation, there is a time lag effect between external disturbances and load response, that is, the change in power load and the change in induced characteristics are not synchronized. This dynamic asynchrony can lead to inaccurate predictions. For example, when the temperature drops sharply, the heating power load will not immediately reach its peak, but will lag for a period of time before it bursts forth. If this lag is not accurately captured by the model, it will affect the timing of dispatch instructions. Furthermore, if a disturbance event with a high sliding impact coefficient (such as a cold wave that is expected to cause huge load fluctuations) occurs simultaneously with a long lag duration that has not been accurately estimated, a dangerous superposition effect will occur. This coupling of "high intensity" and "temporal uncertainty" will drastically amplify the prediction deviation of power load. Based on this, this invention systematically considers both the accurate quantification of dynamic risks and the construction of a time-adaptive scheduling strategy, comprehensively analyzing the impact of both internal and external factors of the power grid on power load, in order to improve the power grid's ability to cope with uncertain disturbances and improve prediction accuracy.

[0058] In particular, by comparing electricity load propagation keywords with fixed and sliding keywords, and combining this with the application tag settings, the method determines whether to implement power reserves. In reality, new electricity consumption cycles often exhibit significant load characteristic continuity with historical cycles. Traditional methods, lacking effective feature identification and quantitative evaluation mechanisms, cannot reliably identify this continuity, leading to decision-making delays. Therefore, this invention considers using keyword matching and tag filtering mechanisms to accurately identify load patterns in the current cycle that are not only highly similar to historical high-impact characteristics but have also been verified as requiring close monitoring. When such characteristics are identified, power reserves can be determined before the actual arrival of peak load, improving the accuracy and timeliness of resource allocation, ensuring sufficient power available for dispatch and allocation during critical load periods, and enhancing forecast accuracy. Attached Figure Description

[0059] Figure 1 This is a schematic diagram illustrating the steps of the method for constructing a power load forecasting agent based on a large model, as described in an embodiment of the invention.

[0060] Figure 2 A logic block diagram illustrating the tendency of the sliding feature to influence electrical effects, as described in an embodiment of the invention.

[0061] Figure 3 This is a logic block diagram illustrating how to determine whether the sliding feature needs to be pre-analyzed according to an embodiment of the invention.

[0062] Figure 4 A logic block diagram for determining power storage in an embodiment of the invention. Detailed Implementation

[0063] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0064] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0065] Please see Figure 1 , Figure 1 This is a schematic diagram illustrating the steps of a method for constructing a power load forecasting agent based on a large model, as described in an embodiment of the invention. The method for constructing a power load forecasting agent based on a large model of the present invention includes:

[0066] Historical periodic data of power load is acquired, the historical periodic data is analyzed to obtain a fixed window and corresponding fixed features, a sliding window and corresponding sliding features are obtained, and the fixed features and the fixed keywords and sliding keywords corresponding to the sliding features are determined.

[0067] Application labels are set for the fixed features. In response to the determination result of the sliding features, the occurrence probability and the same probability of the sliding features are analyzed, the sliding influence coefficient is calculated to classify the power influence tendency of the sliding features, and the labels to be analyzed are set for the sliding features.

[0068] For the sliding feature set as the label to be analyzed, the corresponding power load curve and feature change curve are extracted, the two curves are analyzed to determine the lag time, and the offset influence feature value of the power load is calculated in combination with the sliding influence coefficient. The corresponding label is set to determine whether the sliding feature needs to be pre-analyzed.

[0069] The power load propagation characteristics and corresponding power load propagation keywords are obtained, compared with the fixed keywords and sliding keywords, and the feature similarity is calculated. Combined with the fixed features and sliding features set as application tags, power reserves are determined to complete the prediction of power load.

[0070] Specifically, there are no restrictions on how historical periodic data is obtained. For example, it can be publicly available open-source official data, or internal data submitted after obtaining explicit authorization. It is understood that the use of all data must be ensured to be compliant and authorized, which will not be elaborated further.

[0071] Specifically, a fixed window is a period of time that occurs regularly and is predictable within a cycle (such as daily peak hours). In practice, the window corresponding to a probability of the characteristic occurring within a cycle being greater than 90% is designated as a fixed window. A sliding window is a period of time during which a non-periodic, sudden event lasts (such as a cold wave).

[0072] Specifically, a fixed feature is a load pattern or event that occurs stably within a fixed window, such as "the highest temperature of the day is between 2 pm and 4 pm every day". A sliding window is a sudden and highly variable load pattern or event that occurs within a sliding window, such as "a sudden event such as "a sudden drop in temperature or short-term heavy rainfall during a rainfall process".

[0073] It is understandable that the events corresponding to fixed and sliding characteristics can be obtained from historical periodic data. For example, in an electricity consumption cycle, a regular voltage dip or fluctuation pattern occurs every working day from 8:30 to 9:00. By obtaining data at the corresponding time, it is confirmed that the voltage fluctuation is highly consistent with the concentrated start-up period of large industrial users in the jurisdiction, and it is confirmed that the event corresponding to the fixed characteristic is the concentrated start-up of large industrial users in the jurisdiction.

[0074] Specifically, there are no restrictions on the method of determining keywords through features. For example, each feature can be converted into a standardized text description, and the existing open-source large model can be used to perform vector transformation on the text. The transformed vector is then used as the keyword. For example, in the sentence "During the evening rush hour on weekdays, the air conditioning load surged due to a sudden temperature rise of more than 5°C", the keywords corresponding to example 1 are "weekdays, evening rush hour, sudden temperature rise, 5 degrees Celsius, air conditioning load".

[0075] It is understandable that labels are not entity labels, but rather an abstract concept within the field of data science, referring to the category identifiers of variables or samples in a model.

[0076] Specifically, the power load curve represents the actual power operation values, with time on the horizontal axis and active power on the vertical axis. The characteristic change curve represents the values ​​after feature extraction, with time on the horizontal axis and active power on the vertical axis.

[0077] Specifically, the process of obtaining a fixed window and its corresponding fixed features, and obtaining a sliding window and its corresponding sliding features, includes:

[0078] Construct characteristic change curves in the power transmission process and place each characteristic change curve in the same coordinate system;

[0079] Identify characteristics that exhibit the same trend of change within the same time period;

[0080] The same time period is defined as a fixed window, and the feature is a fixed feature;

[0081] The features on the feature change curve are identified as sliding features;

[0082] The time window corresponding to the maximum sliding feature is determined as the sliding window.

[0083] Specifically, there is no limitation on the specific method for constructing the feature change curve. For example, it can be based on the moving average method of time-domain smoothing, which smooths the original curve by calculating the average value of adjacent data points, thereby constructing the curve. Of course, any method that can obtain the feature change curve can also be used, which will not be elaborated here.

[0084] It is understandable that the data source for constructing the characteristic change curve is historical periodic data.

[0085] It is understandable that the same trend of change indicates that the difference in slope between the two curves does not exceed the preset slope difference. In practice, the preset slope difference is 10%. Those skilled in the art can also determine the preset slope difference according to the actual situation. As long as it is reasonable, it will not be elaborated here.

[0086] Specifically, by dividing the window into fixed and sliding windows, fixed and sliding features are determined, providing a theoretical basis for subsequent feature-specific analysis. In practice, feature extraction often heavily relies on expert prior knowledge and manual interpretation, requiring significant manpower and time for data retrieval and pattern recognition, leading to low model development efficiency and resource utilization. Furthermore, it introduces subjective bias; due to differences in the experience domains and judgment standards of different experts, the same data pattern may be interpreted differently, making it difficult to guarantee the repeatability and objectivity of the feature extraction process. Moreover, this subjectivity and uncertainty directly lead to representational errors between the extracted feature set and the objectively existing real patterns in the data. When these features, which fail to accurately reflect the actual power load driving mechanism, are input into the prediction model, it causes distortion in the model's cognition, resulting in prediction results deviating from reality. This significantly weakens the model's prediction accuracy, robustness, and generalization ability, making it difficult to effectively support precise power dispatch decisions. Therefore, this invention considers targeted feature analysis, determining fixed and sliding features to provide a theoretical data benchmark for subsequent calculation of the sliding influence coefficient, thereby improving prediction accuracy.

[0087] Specifically, the process of analyzing the probability of occurrence of the sliding feature and the process of having the same probability include,

[0088] Determine the number of times the sliding feature appears on the same feature change curve;

[0089] The ratio of the number of occurrences to the total number of features is determined as the first occurrence probability;

[0090] The average of the sum of the first occurrence probabilities of each of the aforementioned feature change curves is determined as the occurrence probability;

[0091] Determine the number of feature change curves exhibiting the sliding characteristic;

[0092] The ratio of the number of curves to the total number of curves representing the characteristic change is determined to be the same probability.

[0093] Specifically, the total number of features is the total number of features that occur during the operation of a power load.

[0094] Specifically, the process of calculating the sliding influence coefficient includes,

[0095] The ratio of the occurrence probability to the baseline occurrence probability is determined as the first influencing factor;

[0096] The ratio of the same probability to the benchmark same probability is determined as the second influencing factor;

[0097] The weighted sum of the first influence factor and the second influence factor is determined to be the sliding influence coefficient.

[0098] Specifically, the baseline probability of occurrence is calculated in advance by obtaining the historical probability of occurrence of this feature during the operation of several power loads, and determining the average of the historical probability of occurrence as the baseline probability of occurrence.

[0099] Specifically, the baseline same probability is calculated in advance by obtaining the historical same probability of this feature during the operation of several power loads, and the mean of each historical same probability is determined as the baseline same probability.

[0100] Specifically, the sum of the weight coefficients of the first and second influencing factors is 1. When adjusting the weight coefficients, considering that the probability of the occurrence of the sliding characteristic leads to a more drastic change in the power load transmission, the weight coefficient of the first influencing factor is set to 0.6 and the weight coefficient of the second influencing factor is set to 0.4.

[0101] Specifically, by quantifying the probability of occurrence of sliding characteristics and calculating their sliding impact coefficient, this invention addresses the significant temporal heterogeneity of power load sequences in reality. Each power consumption cycle is influenced by unique internal and external factors, resulting in distinctive load characteristics and driving mechanisms. If the predictive agent mechanically relies solely on the statistical regularities of historical characteristics while ignoring potential specific disturbances within a particular cycle, such as concerts, sporting events, sudden temperature rises, and sudden temperature drops, it will lead to serious misjudgments. Underestimating the load can result in insufficient reserve capacity, threatening grid security; overestimating the load can lead to excessive reservation and idle power generation resources, increasing system operating costs and causing structural irrationality in resource allocation. In extreme cases, this can even manifest as a supply-demand imbalance. Therefore, this invention analyzes the probability of occurrence of sliding characteristics in historical cycles and their probability of occurrence in the current cycle to calculate the sliding impact coefficient, thereby achieving quantitative assessment and accurate response to load disturbance events and improving prediction accuracy.

[0102] Please see Figure 2 , Figure 2 This is a logic block diagram illustrating the tendency of the sliding feature to be classified according to an embodiment of the invention. Specifically, the tendency of the sliding feature to be classified is determined by assigning a label to be analyzed to the sliding feature, wherein...

[0103] If the sliding influence coefficient is greater than the sliding influence coefficient threshold, then the electric influence tendency of the sliding feature is determined to be a strong electric influence tendency, and a label to be analyzed is set for the sliding feature;

[0104] If the sliding influence coefficient is less than or equal to the sliding influence coefficient threshold, then the electric influence tendency of the sliding feature is determined to be a weak electric influence tendency, and no label is set for the sliding feature.

[0105] Specifically, the sliding influence coefficient threshold characterizes a boundary by which sliding characteristics affect power transmission. It is calculated in advance by obtaining several historical sliding influence coefficients during the operation of several power loads and determining the average value of each historical sliding influence coefficient as the sliding influence coefficient threshold.

[0106] Specifically, the process of analyzing the two curves to determine the lag duration includes,

[0107] Place the power load curve and the characteristic change curve in the same coordinate system;

[0108] Determine the sliding feature and the fixed feature corresponding to the label to be analyzed;

[0109] Determine the power start time point of the power load curve corresponding to each of the sliding features and the fixed features;

[0110] Determine the feature start time point of the feature change curve corresponding to each sliding feature and the fixed feature;

[0111] Determine several differences between the power start time point and the characteristic start time point;

[0112] The average value of each of the aforementioned differences is determined as the lag time.

[0113] Specifically, the process of calculating the characteristic value of the impact of power load offset includes,

[0114] The ratio of the lag duration to the baseline lag duration is determined as the lag effect factor;

[0115] The ratio of the sliding influence coefficient to the benchmark sliding influence coefficient is determined as the sliding influence factor;

[0116] The weighted sum of the lag effect factor and the sliding effect factor is determined to be the offset effect characteristic value.

[0117] Specifically, the benchmark lag time is calculated in advance. The historical lag times during the operation of several power loads are obtained in advance, and the average of each historical lag time is determined as the benchmark lag time.

[0118] Specifically, the benchmark sliding influence coefficient is calculated in advance. Several historical sliding influence coefficients during the operation of power loads are obtained in advance, and the average value of each historical sliding influence coefficient is determined as the benchmark sliding influence coefficient.

[0119] Specifically, the sum of the weighting coefficients of the lag effect factor and the sliding effect factor is 1. When configuring the weighting coefficients, considering that the impact caused by data offset is affected by both factors, the weighting coefficients of the lag effect factor and the sliding effect factor are both set to 0.5.

[0120] Specifically, by analyzing the lag time and combining it with the sliding effect coefficient, the characteristic value of the offset impact of power load is calculated to quantify the potential impact and urgency of the sliding characteristics on power load. In actual power grid operation, there is a time lag effect between external disturbances and load response, meaning that changes in power load and changes in induced characteristics are not synchronized. This dynamic asynchrony can lead to inaccurate predictions. For example, when the temperature drops sharply, heating power load will not immediately reach its peak, but will lag for a period of time before it surges. If this lag is not accurately captured by the model, it will lead to the timing of dispatch instructions being off. Furthermore, if a disturbance event with a high sliding impact coefficient (such as a cold wave that is expected to cause huge load fluctuations) occurs simultaneously with a long lag duration that has not been accurately estimated, a dangerous superposition effect will occur. This coupling of "high intensity" and "temporal uncertainty" will drastically amplify the prediction deviation of power load. Based on this, this invention systematically considers both the accurate quantification of dynamic risks and the construction of a time-adaptive scheduling strategy, comprehensively analyzing the impact of both internal and external factors of the power grid on power load, in order to improve the power grid's ability to cope with uncertain disturbances and improve prediction accuracy.

[0121] Please see Figure 3 , Figure 3 This is a logic block diagram illustrating how to determine whether the sliding feature needs to be pre-analyzed, according to an embodiment of the invention. Specifically, corresponding labels are set to determine whether the sliding feature needs to be pre-analyzed, wherein...

[0122] If the offset influence feature value is greater than the offset influence feature value threshold, then an application label is set to determine that the sliding feature needs to be pre-analyzed.

[0123] If the offset influence feature value is less than or equal to the offset influence feature value threshold, then no label is set, and it is determined that the sliding feature does not need to be pre-analyzed.

[0124] Specifically, the offset impact characteristic value threshold characterizes a boundary that the power load offset will affect the grid islanding. It is calculated in advance. Several historical offset impact characteristic values ​​corresponding to the occurrence of grid islanding are obtained in advance. The product of each historical offset impact characteristic value and the offset coefficient is determined as the offset impact characteristic value threshold. The offset coefficient is obtained in the interval [0.8, 1.0]. In order to improve the judgment accuracy, the offset coefficient is determined to be 0.9 in the implementation.

[0125] Specifically, the process of calculating feature similarity includes,

[0126] The similarity between the power load propagation keywords and the fixed keywords is calculated as the first similarity.

[0127] The similarity between the power load propagation keywords and the sliding keywords is calculated as the second similarity.

[0128] The average of the sum of the first similarity and the second similarity is determined as the feature similarity.

[0129] Specifically, there are no restrictions on the methods for calculating keyword similarity. For example, cosine similarity can be used, or other methods can be used, as long as keyword similarity can be obtained. This will not be elaborated further.

[0130] Please see Figure 4 , Figure 4 This is a logic block diagram illustrating the determination of power storage according to an embodiment of the invention. Specifically, determining to perform power storage, wherein...

[0131] If the power load propagation characteristics meet the preset conditions, then it is determined to perform power storage;

[0132] If the power load propagation characteristics do not meet the preset conditions, then it is uncertain whether to perform power storage;

[0133] The preset condition is that the similarity is greater than the feature similarity threshold and the label of the power load propagation feature is the application label.

[0134] Understandably, characteristics of electricity load propagation can come from real-time meteorological data, event announcements, or social media information.

[0135] Specifically, the feature similarity threshold represents a boundary that needs to be reserved in advance. It is calculated in advance, and several historical feature similarities of grid islanding are obtained in advance. The mean of each historical feature similarity is determined as the feature similarity threshold.

[0136] Specifically, by comparing electricity load propagation keywords with fixed and sliding keywords, and combining this with the application tag settings, the method determines whether to implement electricity reserves. In reality, new electricity consumption cycles often exhibit significant load characteristic continuity with historical cycles. Traditional methods, lacking effective feature identification and quantitative evaluation mechanisms, cannot reliably identify this continuity, leading to decision-making delays. Therefore, this invention considers using keyword matching and tag filtering mechanisms to accurately identify load patterns in the current cycle that are not only highly similar to historical high-impact characteristics but have also been verified as requiring close monitoring. When such characteristics are identified, electricity reserves can be determined before the actual arrival of peak load, improving the accuracy and timeliness of resource allocation, ensuring sufficient power available for dispatch and allocation during critical load periods, and enhancing forecast accuracy.

[0137] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0138] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing an intelligent agent for power load forecasting based on a large model, characterized in that, include: Historical periodic data of power load is acquired, the historical periodic data is analyzed to obtain a fixed window and corresponding fixed features, a sliding window and corresponding sliding features are obtained, and the fixed features and the fixed keywords and sliding keywords corresponding to the sliding features are determined. Application labels are set for the fixed features. In response to the determination result of the sliding features, the occurrence probability and the same probability of the sliding features are analyzed, the sliding influence coefficient is calculated to classify the power influence tendency of the sliding features, and the labels to be analyzed are set for the sliding features. For the sliding feature set as the label to be analyzed, the corresponding power load curve and feature change curve are extracted, the two curves are analyzed to determine the lag time, and the offset influence feature value of the power load is calculated in combination with the sliding influence coefficient. The corresponding label is set to determine whether the sliding feature needs to be pre-analyzed. The characteristics of power load propagation and the corresponding keywords of power load propagation are obtained, compared with the fixed keywords and the sliding keywords, the feature similarity is calculated, and the fixed features and the sliding features set as application tags are combined to determine the power reserve and complete the prediction of power load. The process of analyzing the probability of occurrence of the sliding feature and the same probability includes, Determine the number of times the sliding feature appears on the same feature change curve; The ratio of the number of occurrences to the total number of features is determined as the first occurrence probability; The average of the sum of the first occurrence probabilities of each of the aforementioned feature change curves is determined as the occurrence probability; Determine the number of feature change curves exhibiting the sliding characteristic; The probability is that the ratio of the number of curves to the total number of curves representing the characteristic change is the same. The process of calculating the sliding influence coefficient includes: The ratio of the occurrence probability to the baseline occurrence probability is determined as the first influencing factor; The ratio of the same probability to the benchmark same probability is determined as the second influencing factor; The weighted sum of the first influence factor and the second influence factor is determined to be the sliding influence coefficient.

2. The method for constructing an intelligent agent for power load forecasting based on a large model according to claim 1, characterized in that, The process of obtaining a fixed window and its corresponding fixed features, and obtaining a sliding window and its corresponding sliding features, includes: Construct characteristic change curves in the power transmission process and place each characteristic change curve in the same coordinate system; Identify characteristics that exhibit the same trend of change within the same time period; The same time period is defined as a fixed window, and the feature is a fixed feature; The features on the feature change curve are identified as sliding features; The time window corresponding to the maximum sliding feature is determined as the sliding window.

3. The method for constructing an intelligent agent for power load forecasting based on a large model according to claim 1, characterized in that, The method involves classifying the electric field influence tendency of the sliding feature and assigning a label to be analyzed to the sliding feature. If the sliding influence coefficient is greater than the sliding influence coefficient threshold, then the electric influence tendency of the sliding feature is determined to be a strong electric influence tendency, and a label to be analyzed is set for the sliding feature; If the sliding influence coefficient is less than or equal to the sliding influence coefficient threshold, then the electric influence tendency of the sliding feature is determined to be a weak electric influence tendency, and no label is set for the sliding feature.

4. The method for constructing an intelligent agent for power load forecasting based on a large model according to claim 1, characterized in that, The process of analyzing the two curves to determine the lag time includes, Place the power load curve and the characteristic change curve in the same coordinate system; Determine the sliding feature and the fixed feature corresponding to the label to be analyzed; Determine the power start time point of the power load curve corresponding to each of the sliding features and the fixed features; Determine the feature start time point of the feature change curve corresponding to each sliding feature and the fixed feature; Determine several differences between the power start time point and the characteristic start time point; The average value of each of the aforementioned differences is determined as the lag time.

5. The method for constructing an intelligent agent for power load forecasting based on a large model according to claim 1, characterized in that, The process of calculating the characteristic value of the offset effect of power load includes, The ratio of the lag duration to the baseline lag duration is determined as the lag effect factor; The ratio of the sliding influence coefficient to the benchmark sliding influence coefficient is determined as the sliding influence factor; The weighted sum of the lag effect factor and the sliding effect factor is determined to be the offset effect characteristic value.

6. The method for constructing an intelligent agent for power load forecasting based on a large model according to claim 1, characterized in that, The setting of corresponding labels is used to determine whether the sliding features need to be pre-analyzed. If the offset influence feature value is greater than the offset influence feature value threshold, then an application label is set to determine that the sliding feature needs to be pre-analyzed. If the offset influence feature value is less than or equal to the offset influence feature value threshold, then no label is set, and it is determined that the sliding feature does not need to be pre-analyzed.

7. The method for constructing an intelligent agent for power load forecasting based on a large model according to claim 1, characterized in that, The process of calculating feature similarity includes: The similarity between the power load propagation keywords and the fixed keywords is calculated as the first similarity. The similarity between the power load propagation keywords and the sliding keywords is calculated as the second similarity. The average of the sum of the first similarity and the second similarity is determined as the feature similarity.

8. The method for constructing an intelligent agent for power load forecasting based on a large model according to claim 1, characterized in that, The determination to conduct power reserves, wherein... If the power load propagation characteristics meet the preset conditions, then it is determined to perform power storage; If the power load propagation characteristics do not meet the preset conditions, then it is uncertain whether to perform power storage; The preset conditions are that the feature similarity is greater than the feature similarity threshold, and at least one associated feature is set as an application label.

Images