Dynamic calibration method and system for electric load quantile prediction

Abstract

The application provides a dynamic calibration method and system for power load quantile prediction, and belongs to the technical field of power load prediction. The method comprises the following steps: interacting with a target scene, obtaining quantile prediction models of multiple target quantiles, and performing prediction performance evaluation on each quantile prediction model based on a preset dynamic calibration period; performing double-target calibration trigger discrimination based on the prediction performance evaluation result; determining multiple characteristic quantile prediction models, and obtaining a characteristic quantile prediction model set; obtaining a multi-source sample prediction data set; and performing progressive calibration on the double-target calibration trigger discrimination result according to the multi-source sample prediction data set and the characteristic quantile prediction model set in combination with a preset multi-modal calibration strategy. The performance of each quantile model is dynamically evaluated, and the cross-quantile prediction capability of adjacent quantile models is utilized for collaborative calibration, so that the overall precision and stability of quantile prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power load prediction, and particularly relates to a dynamic calibration method and system for power load quantile prediction. BACKGROUND

[0002] With the continuous expansion of the scale of the power system and the increasing complexity of the load composition, accurate prediction of power load has become one of the key technologies for safe and stable operation of the power system. As an important probability prediction method, quantile prediction can provide prediction intervals at different confidence levels and provide more comprehensive information support for power dispatching decisions.

[0003] The existing quantile prediction method usually trains independent prediction models for different quantile points, and each model only focuses on the prediction task of a specific quantile. Although this method can realize quantile prediction to some extent, it has the following technical defects: first, the existing quantile prediction model lacks a dynamic calibration mechanism. In actual application, due to the time-varying and complex nature of power load, the prediction performance of the model will change over time, and the existing method cannot dynamically identify the fluctuation of the model performance, nor can it calibrate the model in a timely manner according to the performance changes, resulting in difficulty in maintaining the stability of the prediction accuracy. Secondly, there is a lack of effective coordination mechanism among different quantile models. Adjacent quantiles often have similar variation rules and prediction characteristics, and their prediction experience is of great value to improving the performance of other quantile models. However, in the existing technology, each quantile model is trained and predicted independently, and cannot effectively utilize the prediction experience of adjacent quantiles for collaborative optimization. The above technical defects result in low prediction accuracy and stability of the existing quantile prediction, which is difficult to meet the actual demand of the power system for high-precision load prediction. SUMMARY

[0004] The present application provides a dynamic calibration method and system for power load quantile prediction to solve the technical problems of lack of dynamic calibration mechanism in the existing quantile prediction model and inability to effectively utilize the prediction experience of adjacent quantiles among different quantile models, resulting in low prediction accuracy and stability.

[0005] The technical solution of the present application to solve the above technical problems is as follows:

[0006] In a first aspect, the present application provides a dynamic calibration method for power load quantile prediction, comprising: interacting with a target scene, obtaining quantile prediction models of multiple target quantiles, and performing prediction performance evaluation on each quantile prediction model based on a preset dynamic calibration period; performing double-target calibration trigger discrimination based on the prediction performance evaluation result; combining the double-target calibration trigger discrimination result with a preset neighborhood constraint, traversing multiple quantile prediction models to perform performance matching-based model screening, determining multiple adaptive quantile prediction models, and obtaining an adaptive quantile prediction model set; using the double-target calibration trigger discrimination result and the adaptive quantile prediction model set as an index, corresponding sample prediction data is extracted, a multi-source sample prediction data set is obtained; and according to the multi-source sample prediction data set and the adaptive quantile prediction model set, the double-target calibration trigger discrimination result is progressively calibrated in combination with a preset multi-modal calibration strategy.

[0007] In a second aspect, the present application provides a dynamic calibration system for power load quantile prediction, comprising: a prediction performance evaluation module, configured to interact with a target scene, obtain quantile prediction models of multiple target quantiles, and perform prediction performance evaluation on each quantile prediction model based on a preset dynamic calibration period; a double-target trigger discrimination module, configured to perform double-target calibration trigger discrimination based on the prediction performance evaluation result; a performance matching screening module, configured to combine the double-target calibration trigger discrimination result with a preset neighborhood constraint, traverse multiple quantile prediction models to perform performance matching-based model screening, determine multiple adaptive quantile prediction models, and obtain an adaptive quantile prediction model set; a sample data extraction module, configured to use the double-target calibration trigger discrimination result and the adaptive quantile prediction model set as an index, corresponding sample prediction data is extracted, and a multi-source sample prediction data set is obtained; and a progressive calibration execution module, configured to progressively calibrate the double-target calibration trigger discrimination result in combination with a preset multi-modal calibration strategy according to the multi-source sample prediction data set and the adaptive quantile prediction model set.

[0008] The present application has the following beneficial effects:

[0009] The interactive target scene obtains quantile prediction models of multiple target quantiles, and performs prediction performance evaluation on each quantile prediction model based on a preset dynamic calibration period to provide performance data support for subsequent dynamic calibration. The double-target calibration trigger discrimination is performed based on the prediction performance evaluation result, so that the quantile model that needs to be calibrated is dynamically identified, unnecessary calibration operation is avoided, and the calibration efficiency is improved. The performance matching based model screening is performed on the multiple quantile prediction models combined with the double-target calibration trigger discrimination result and the preset neighborhood constraint to determine multiple matching quantile prediction models, obtain a matching quantile prediction model set, and select the adjacent quantile models that match the performance of the to-be-calibrated model and meet the neighborhood constraint, so as to ensure that the selected model has the ability to provide effective prediction experience. The sample prediction data is extracted corresponding to the double-target calibration trigger discrimination result and the matching quantile prediction model set, a multi-source sample prediction data set is obtained, and sample data from different matching quantile models is obtained to provide a multi-source data basis for subsequent collaborative calibration. The double-target calibration trigger discrimination result is progressively calibrated combined with the preset multi-modal calibration strategy based on the multi-source sample prediction data set and the matching quantile prediction model set, the prediction experience of the adjacent quantile model is utilized, and the to-be-calibrated model is collaboratively optimized by using multiple calibration strategies, so that the collaborative calibration between the quantile prediction models is realized.

[0010] Through the above technical solution, the dynamic calibration mechanism of the quantile prediction model is established, the effective utilization and collaborative optimization of the prediction experience between different quantile models are realized, the technical defects that the quantile prediction model lacks a dynamic calibration mechanism and cannot effectively utilize the prediction experience of the adjacent quantile model in the prior art are overcome, and thus the overall precision and stability of the quantile prediction are significantly improved, and the actual demand of the power system for high-precision load prediction is met. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 A flowchart of the dynamic calibration method of the power load quantile prediction provided by the present application is shown.

[0012] Figure 2 A structural diagram of the dynamic calibration system of the power load quantile prediction provided by the present application is shown.

[0013] In the drawings, the components represented by the numbers are as follows:

[0014] The prediction performance evaluation module 11, the double-target trigger discrimination module 12, the performance matching screening module 13, the sample data extraction module 14, and the progressive calibration execution module 15. DETAILED DESCRIPTION

[0015] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0016] In the description of the present application, the terms "first", "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0017] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or description". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given so that any person skilled in the art can implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can realize the present application without using these specific details. In other examples, well-known structures and processes will not be described in detail to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the shown embodiments, but is consistent with the broadest scope of the principles and characteristics disclosed.

[0018] Embodiment one, as shown in the present application, a dynamic calibration method for power load quantile prediction is provided, comprising: Figure 1

[0019] S1, interact with the target scene, obtain a plurality of target quantile quantile prediction model, and evaluate the prediction performance of each quantile prediction model based on the preset dynamic calibration period.

[0020] ​Specifically, first, interact with the target scene to obtain quantile prediction models for multiple target quantiles. Among them, the target scene refers to specific power load forecasting application environment and business requirements, including but not limited to specific regional power grid systems (such as a provincial power grid, a municipal distribution network), specific types of power load (such as industrial load, residential load, commercial load), specific time scale forecasting tasks (such as day-ahead forecasting, week-ahead forecasting, month-ahead forecasting), etc. Different target scenes correspond to different load characteristics, forecasting accuracy requirements and risk control needs, so a corresponding set of quantile prediction models needs to be constructed. Multiple target quantiles refer to a number of selected quantile values according to the actual needs of power system operation and dispatch. Quantile is a statistical concept used to describe the location characteristics of data distribution. For example, low quantiles (such as 0.05, 0.1, 0.2 quantiles): used to estimate the lower bound of load, helping to identify the risk of insufficient load, and providing a basis for standby power start-stop and demand response planning; Mid-quantile (such as 0.5 quantile, i.e. median): used to estimate the central tendency of load, as a benchmark value for regular dispatch planning; High quantile (such as 0.8, 0.9, 0.95 quantile): used to estimate the upper bound of load, helping to identify the risk of excessive load, and providing support for power grid safe operation and emergency planning. Assuming that the load forecasting result of a power grid at a certain time is: 0.1 quantile = 800 MW, 0.5 quantile = 1000 MW, 0.9 quantile = 1200 MW, which means: there is a 10% probability that the actual load will be lower than 800 MW, there is a 50% probability that the actual load will be lower than 1000 MW, and there is a 90% probability that the actual load will be lower than 1200 MW. Such interval prediction information is more informative and practical than a single point prediction value (such as predicting only 1000 MW).

[0021] Quantile prediction model refers to a machine learning model or statistical model that can output a specific quantile prediction value. Unlike traditional point prediction models (which only output one prediction value), quantile prediction models need to learn the probability distribution characteristics of load data, so they can output prediction results at different confidence levels. Each target quantile corresponds to an independent prediction model, for example: a 0.1 quantile prediction model is specifically used to predict a 0.1 quantile value, a 0.5 quantile prediction model is specifically used to predict a 0.5 quantile value, and so on. These models are usually built using deep learning networks (such as LSTM, GRU, Transformer, etc.), traditional machine learning methods (such as support vector regression, random forest, etc.) or statistical methods (such as ARIMA, exponential smoothing, etc.), and are trained and optimized using quantile loss functions (such as Pinball Loss).

[0022] Subsequently, the prediction performance of each quantile prediction model is evaluated based on a preset dynamic calibration period. The dynamic calibration period is a time interval preset according to the timeliness requirement of power load prediction and the variation law of model performance, and is a variable period for adaptive adjustment according to the variation of model performance. In each dynamic calibration period, all prediction results and corresponding actual observation values of each quantile prediction model in the period are collected, and the prediction performance of the model is quantitatively analyzed based on a preset evaluation index.

[0023] The prediction performance evaluation is performed from two dimensions: one is prediction stability evaluation, which measures the stability performance of the model by calculating the fluctuation degree of the prediction results in the time sequence. Greater fluctuation indicates that the model is easily affected by data noise or external interference; the other is prediction accuracy evaluation, which measures the precision performance of the model by calculating the relative error or absolute error between the predicted value and the true value. Greater error indicates that the fitting ability or generalization ability of the model is insufficient. Through this two-dimensional performance evaluation mechanism, the running state of each quantile prediction model can be comprehensively reflected, providing a reliable data basis for subsequent calibration decisions.

[0024] S2, dual-target calibration trigger discrimination is performed based on the prediction performance evaluation result.

[0025] Specifically, the dual-target calibration trigger discrimination refers to a calibration trigger decision mechanism that considers both the prediction stability and the prediction accuracy, thereby comprehensively evaluating the comprehensive performance of the model.

[0026] Specifically, the dual-target calibration trigger discrimination includes the following two discrimination targets:

[0027] The first target is prediction stability discrimination. This target mainly evaluates the fluctuation of the prediction results of the quantile prediction model in the continuous time period. The prediction stability is quantified by a fluctuation index, and the prediction fluctuation entropy is used to calculate the fluctuation index. The prediction fluctuation entropy is an application of information entropy to the fluctuation analysis of prediction results. By calculating the information entropy value of the distribution probability of the prediction value in different intervals, the uncertainty degree of the prediction results is measured. When the prediction fluctuation entropy exceeds the preset first threshold, it indicates that the stability of the quantile prediction model is insufficient, and there is a large prediction fluctuation, which needs to be calibrated.

[0028] The second target is prediction accuracy discrimination. This target mainly evaluates the prediction accuracy performance of the quantile prediction model. The prediction accuracy is quantified by an error index, which is calculated using a prediction relative error rate. The prediction relative error rate refers to the relative deviation between the predicted value and the actual observed value, which can reflect the fitting accuracy of the model. In order to better reflect the influence of recent data on the performance of the model, the prediction relative error rate is calculated in combination with a time-weighted mechanism, so that the prediction results with closer time distance have higher weights. When the prediction relative error rate exceeds the preset second threshold, it indicates that the accuracy of the quantile prediction model is insufficient, the prediction accuracy is significantly decreased, and calibration is needed.

[0029] The core logic of the dual-target calibration trigger discrimination is that when the volatility index of a certain quantile prediction model exceeds the first threshold, or its error index exceeds the second threshold, it is determined that the model needs to be calibrated and optimized, and it is marked as a to-be-calibrated quantile prediction model. This logical relationship ensures that the calibration mechanism can be triggered in time as soon as the model performance deteriorates in either stability or accuracy, thereby ensuring the overall quality of power load prediction.

[0030] Through the dual-target calibration trigger discrimination, the current quantile prediction model with poor performance can be dynamically identified, and a dual-target calibration trigger discrimination result is output. The dual-target calibration trigger discrimination result contains the identification of all to-be-calibrated quantile prediction models that need to be calibrated, providing a clear target object for subsequent model calibration and optimization.

[0031] S3, in combination with the dual-target calibration trigger discrimination result and the preset neighborhood constraint, a plurality of quantile prediction models are traversed for performance-matched model screening to determine a plurality of matching quantile prediction models and obtain a matching quantile prediction model set.

[0032] Specifically, the neighborhood constraint refers to the range limitation of model screening set based on the principle of quantile value proximity. Since adjacent quantiles correspond to load distribution characteristics with certain similarity, the prediction models of adjacent quantiles may have complementarity and transferability in prediction tasks. The neighborhood constraint defines the neighborhood range by setting the upper and lower bounds of the quantile difference. For example, for a 0.5 quantile prediction model, its neighborhood range can be set to [0.3, 0.7], i.e. all quantile prediction models within a range of quantile difference not exceeding ±0.2. The setting of the neighborhood constraint ensures the correlation between models and avoids negative interference from models with too much difference.

[0033] The model screening based on performance matching is a process of evaluating the adaptability of quantile prediction models in the neighborhood range to the prediction task of the quantile prediction model to be calibrated. Specifically, for each quantile prediction model to be calibrated, its corresponding historical prediction data is extracted as a validation data set, and then other quantile prediction models in the neighborhood range are used to predict the validation data set, and the prediction performance of these neighborhood models on the prediction task of the quantile prediction model to be calibrated is calculated.

[0034] The performance matching process includes the following steps: first, obtaining the historical prediction data of the quantile prediction model to be calibrated as the first quantile validation data; second, re-predicting the validation data using each quantile prediction model in the neighborhood range, and calculating the performance indicators (such as mean absolute error, root mean square error, etc.) between the prediction results and the true observed values; finally, based on the size of the performance indicators, the adaptability of each neighborhood model to the prediction task of the quantile prediction model to be calibrated is evaluated.

[0035] The matching quantile prediction model refers to the neighborhood quantile prediction model that performs well in performance matching and has the potential to assist the quantile prediction model to be calibrated in calibration optimization. The matching quantile prediction model needs to meet the following conditions: first, it is located within the neighborhood constraint range of the quantile prediction model to be calibrated, ensuring the correlation between the models; second, it has good performance on the prediction task of the quantile prediction model to be calibrated, that is, its cross-quantile prediction ability meets the preset benchmark threshold requirements.

[0036] The model screening process adopts a progressive screening strategy: first, based on the performance matching results, the quantile prediction models in the neighborhood range are arranged in descending order to form a performance ranking sequence; second, a one-time screening based on the performance threshold is performed to eliminate models with performance lower than the preset performance threshold; finally, a second screening based on the number is performed to select the top several models with the best performance from the one-time screening results as the final matching quantile prediction models to control the computational complexity of the subsequent calibration process.

[0037] The matching quantile prediction model set refers to the set of matching quantile prediction models corresponding to all quantile prediction models to be calibrated. By traversing all quantile prediction models to be calibrated in the double-target calibration trigger judgment result, the corresponding matching quantile prediction model is determined for each quantile prediction model to be calibrated, and a one-to-many mapping relationship is established. Each quantile prediction model to be calibrated corresponds to at least one matching quantile prediction model to ensure that sufficient auxiliary information can be obtained in the subsequent calibration process. The matching quantile prediction model set provides high-quality candidate model resources for subsequent calibration strategies such as incremental learning and model fusion.

[0038] S4, taking the double-target calibration trigger judgment result and the set of goodness-of-fit quantile prediction models as indexes, corresponding sample prediction data is extracted, and a multi-source sample prediction data set is obtained.

[0039] Specifically, sample prediction data refers to various data samples used for model calibration training, including input feature data, prediction output data, real label data, and intermediate state information in the model prediction process. By extracting sample prediction data from different sources, more rich and diverse training information can be provided for the to-be-calibrated quantile prediction model, forming a multi-source sample prediction data set, thereby improving the calibration effect.

[0040] The multi-source sample prediction data set includes the following three types of data sources:

[0041] The first type is standard observation-sample prediction data. This type of data is obtained by calling the historical training data of the to-be-calibrated quantile prediction model, and contains historical load feature input (such as temperature, humidity, holiday information, historical load, etc.) and corresponding real load observation value. The standard observation-sample prediction data represents the standard supervision signal in the model training process, and provides reliable true value reference for the calibration process. This type of data usually contains time series features, weather features, economic indicators, and electricity behavior patterns, which can fully reflect the change law of power load.

[0042] The second type is to-be-calibrated model-sample prediction data. This type of data is obtained by collecting historical prediction data of the to-be-calibrated quantile prediction model, and contains the prediction output result and the corresponding prediction probability distribution information of the model at the historical moment. The prediction probability distribution is output by a softmax function, which reflects the confidence allocation of the model for different prediction results. The to-be-calibrated model-sample prediction data can reflect the current prediction preference and decision pattern of the model, and provide the experience information of the model itself for the calibration process. This type of data helps to identify the systematic bias and prediction blind area of the model, and provides a basis for targeted calibration.

[0043] The third type is goodness-of-fit model-sample prediction data. This type of data is obtained by calling the performance matching verification result of the goodness-of-fit quantile prediction model, and is processed by classification labeling. Classification labeling refers to labeling the prediction samples of the goodness-of-fit model as positive samples (good performance prediction results) and negative samples (poor performance prediction results) according to the performance of the goodness-of-fit model in the to-be-calibrated quantile prediction model verification task. The goodness-of-fit model-sample prediction data carries the prediction experience and knowledge of adjacent quantile models, and can provide cross-quantile learning reference for the to-be-calibrated quantile prediction model. Through classification labeling, the system can distinguish between effective knowledge and ineffective knowledge in the goodness-of-fit model, improving the accuracy and efficiency of knowledge transfer.

[0044] The construction process of the multi-source sample prediction data set is as follows: first, all to-be-calibrated quantile prediction models are determined according to the double-target calibration trigger judgment result; second, for each to-be-calibrated quantile prediction model, standard observation-sample prediction data is extracted from its historical training data, and to-be-calibrated model-sample prediction data is extracted from its historical prediction records; third, according to the set of matching quantile prediction models, prediction verification data of each matching model is extracted and marked as positive and negative samples to form matching model-sample prediction data; and finally, the three types of data sources are integrated to form a complete multi-source sample prediction data set.

[0045] By constructing the multi-source sample prediction data set, the historical experience of the to-be-calibrated model, the supervision information of the real observation data, and the cross-quantile knowledge of the matching model can be fully utilized to provide a comprehensive and rich data basis for subsequent progressive calibration, and the effect and efficiency of model calibration are improved.

[0046] S5, according to the multi-source sample prediction data set and the set of matching quantile prediction models, combined with a preset multi-modal calibration strategy, the double-target calibration trigger judgment result is progressively calibrated.

[0047] Specifically, the multi-modal calibration strategy refers to a comprehensive optimization strategy that combines multiple calibration methods for different types of model performance problems, for example, three calibration modes including incremental calibration, forward transfer calibration, and reverse transfer calibration, which can adapt to model performance degradation problems of different degrees and types. The core idea of the multi-modal calibration strategy is to optimize and improve the to-be-calibrated model from simple to complex and from local to global through a progressive calibration process.

[0048] Progressive calibration refers to a dynamic optimization process that attempts different calibration strategies in order of preset priority and decides whether to continue subsequent calibration steps according to the calibration effect. Progressive calibration adopts an early stopping mechanism, that is, when the effect of a calibration step meets the prediction performance requirement, the subsequent calibration process is terminated immediately, and the current calibration result is output, so as to improve the calibration efficiency and avoid overfitting.

[0049] The specific progressive calibration process includes the following three stages:

[0050] The first stage is incremental calibration. The standard observation-sample prediction data is used as a supervision signal to calibrate the to-be-calibrated quantile prediction model based on incremental learning. Incremental learning refers to a learning method that fine-tunes and optimizes model parameters based on new training data while maintaining the original knowledge structure of the model. During the incremental calibration process, the historical training data of the to-be-calibrated quantile prediction model is used to update the model parameters through optimization algorithms such as gradient descent, so that the model can better adapt to the current data distribution characteristics. After incremental calibration, the first prediction performance of the calibrated model is verified, including prediction stability and prediction accuracy indicators.

[0051] The second stage is forward transfer calibration. When the first prediction performance does not meet the preset prediction performance requirements, the forward transfer calibration stage is entered. This stage uses a supervised learning-based knowledge transfer method, and simultaneously uses the matching model-sample prediction data as a soft target and the standard observation-sample prediction data as a hard target for model optimization. The soft target refers to the prediction results from the matching quantile prediction model, which contains the experience knowledge of adjacent quantile models; the hard target refers to the true load observation value, which represents objective supervision information. Forward transfer calibration minimizes the prediction errors of the to-be-calibrated quantile prediction model for soft and hard targets by constructing a joint loss function, achieving effective transfer of cross-quantile knowledge. After calibration, the second prediction performance is verified.

[0052] The third stage is reverse transfer calibration. When the second prediction performance still does not meet the prediction performance requirements, the reverse transfer calibration stage is entered. This stage uses a reverse transfer approach, taking the matching quantile prediction model as the base model, using the historical prediction data of the to-be-calibrated quantile prediction model as a soft target and the standard observation data as a hard target to optimize the base model. The core idea of reverse transfer calibration is to use a better matching model as a new starting point, and by fusing the prediction experience of the calibration quantile prediction model and the true observation information, a new optimized model is constructed. This method is suitable for cases where the calibration quantile prediction model performance is severely degraded and needs to be deeply reconstructed. After calibration, the third prediction performance is verified.

[0053] The progressive calibration process has the following advantages: first, it uses a progressive strategy to select the appropriate calibration strength according to the degree of model performance degradation, avoiding unnecessary computational overhead; second, it fuses multi-source data information, fully utilizing historical observations, model experience and cross-quantile knowledge to improve calibration effectiveness; third, it uses an early stopping mechanism to terminate the calibration process when the performance requirements are met, ensuring calibration efficiency; fourth, it supports multiple calibration modalities, which can adapt to different types and degrees of model performance problems, and has good adaptability and robustness.

[0054] Through progressive calibration, the performance defects of the to-be-calibrated quantile prediction model can be effectively repaired, and the stability and accuracy of the to-be-calibrated quantile prediction model in the power load prediction task are improved, thereby providing reliable prediction support for the safe and stable operation of the power system.

[0055] Further, the interaction target scene is acquired, a plurality of target quantile prediction models are obtained, and the prediction performance of each quantile prediction model is evaluated based on a preset dynamic calibration period, including:

[0056] S11, based on a sliding window of a predetermined size, the prediction fluctuation entropy of each quantile prediction model is monitored and calculated in real time, and the output is a volatility index;

[0057] S12, combined with a time sequence weighting mechanism, the prediction relative error rate of each quantile prediction model is calculated, and the output is an error index, wherein the time sequence weighting weight decreases along the time distance direction;

[0058] S13, combined with the volatility index and the error index, the prediction performance evaluation result is formed.

[0059] In a preferred embodiment, firstly, the prediction volatility entropy of each quantile prediction model is monitored and calculated in real time based on a predetermined size of a sliding window, and the output is a volatility indicator. The sliding window refers to a data collection window that moves on a time series with a fixed size and a fixed step length. The size of the sliding window is set according to the time granularity of the power load prediction and the business needs. For example, for hourly load prediction, the sliding window size can be set to 24 hours, 48 hours or 72 hours, etc.; for daily load prediction, the sliding window size can be set to 7 days, 14 days or 30 days, etc. The sliding window moves forward at a fixed time interval (such as every hour or every day) to ensure continuous monitoring of the prediction performance changes of the model. The prediction volatility entropy is an index that applies information entropy theory to the volatility analysis of prediction results. Information entropy is an important index to measure the degree of information uncertainty, and the greater the value, the higher the uncertainty of the system. The calculation process of the prediction volatility entropy is as follows: firstly, in the current sliding window, all prediction results of the quantile prediction model are extracted to form a prediction value sequence; secondly, the value distribution characteristics of the prediction sequence are analyzed, the average level and the volatility range of the prediction value are calculated, and the prediction value range is divided into several interval segments according to the statistical law, and the number of intervals is usually set to 5 to 10; thirdly, the number of values falling into each interval segment in the prediction sequence is counted, and the proportion of each interval segment is calculated to form a probability distribution; then, the information entropy value is calculated based on the probability distribution of each interval segment as the prediction volatility entropy. When the prediction results are distributed in a few intervals, the probability distribution is uneven, the volatility entropy value is small, and the prediction is relatively stable; when the prediction results are scattered in multiple intervals, the probability distribution tends to be uniform, the volatility entropy value is large, and the prediction volatility is intense. The prediction volatility entropy can effectively quantify the stability performance of the quantile prediction model on the time series, and provide an objective quantitative basis for identifying the prediction stability problem of the model.

[0060] Subsequently, the prediction relative error rate of each quantile prediction model is calculated combined with the time sequence weighting mechanism, and the output is the error index. The time sequence weighting mechanism refers to a weighting strategy of assigning different weights to the prediction errors of different time points according to the distance of the data from the current time. The time sequence weighting mechanism is based on the principle that “recent data can better reflect the current performance state of the model”, so that the prediction results with closer time distance have higher weights, and the prediction results with farther time distance have lower weights. The weight gradually decreases with the increase of the time distance, which can adopt exponential decay, linear decay or piecewise decay attenuation mode. The calculation process of the prediction relative error rate is as follows: first, in the current sliding window, the prediction value sequence and the corresponding true observation value sequence of the quantile prediction model are extracted; second, the relative deviation degree between the prediction value and the true value at each time point is calculated, that is, the proportional relationship of the prediction error relative to the true value; third, according to the time sequence weighting mechanism, each time point relative error is assigned a weight, the error weight of the closer time point is larger, and the error weight of the farther time point is smaller; then, the weighted average value of all time points relative error is calculated to obtain the prediction relative error rate. By introducing the time sequence weighting mechanism, the prediction relative error rate can more sensitively reflect the recent prediction accuracy change of the model, highlight the influence of recent performance on the overall evaluation, and provide strong support for timely detection of model performance degradation.

[0061] Then, the prediction performance evaluation result is formed combined with the volatility index and the error index. The prediction performance evaluation result is a comprehensive performance evaluation result considering the prediction stability and the prediction accuracy of the model. The prediction performance evaluation result includes the volatility index value, the error index value corresponding to each quantile prediction model, and the comprehensive performance score calculated based on the two indexes. The comprehensive performance score is calculated by weighted fusion of the volatility index and the error index, and the weight distribution can be adjusted according to the needs of specific application scenarios, for example, the weight of the volatility index can be increased in application scenarios with high stability requirements, and the weight of the error index can be increased in application scenarios with high accuracy requirements. The prediction performance evaluation result provides a quantitative basis for subsequent double-target calibration trigger discrimination, can accurately identify quantile prediction models with insufficient prediction stability or poor prediction accuracy, and ensures that the power load prediction system can timely detect and solve performance problems and maintain overall prediction quality.

[0062] Further, double-target calibration trigger discrimination is performed based on the prediction performance evaluation result, including:

[0063] S21, obtaining the prediction performance requirement of the target scene, including the prediction stability requirement and the prediction accuracy requirement;

[0064] S22, calculating the first type threshold and the second type threshold according to the prediction stability requirement and the prediction accuracy requirement respectively, and obtaining the double-target calibration trigger discrimination limit;

[0065] S23, compare the prediction performance evaluation result with the double-target calibration trigger discrimination limit;

[0066] S24, if any of the double-target trigger conditions is met, it is determined that the corresponding quantile prediction model needs to be calibrated, and the corresponding label is a to-be-calibrated quantile prediction model, wherein the double-target trigger condition includes:

[0067] the volatility index exceeds the first type of threshold;

[0068] the error index exceeds the second type of threshold;

[0069] S25, output a plurality of the to-be-calibrated quantile prediction models as double-target calibration trigger discrimination results.

[0070] In a preferred embodiment, first, the prediction performance requirements of the target scene are obtained, including the prediction stability requirements and the prediction accuracy requirements. The prediction performance requirements of the target scene refer to the model performance standards and quality requirements formulated according to the business characteristics and technical requirements of specific power load prediction application scenes. Different power system application scenes have significant differences in prediction performance requirements, which need to be set individually in combination with actual business requirements. The prediction stability requirement refers to the requirement for the consistency and stability of the prediction results of the quantile prediction model in consecutive time periods. The prediction stability requirement mainly considers the following factors: the continuity requirement of power grid dispatching, i.e., the prediction results should not fluctuate sharply to avoid affecting the rationality of the dispatching plan; the credibility requirement of load prediction, i.e., the prediction results should have a certain predictability and regularity; the safety requirement of system operation, i.e., the mutation of the prediction results may cause power grid safety risks. The prediction stability requirement is quantified by an acceptable range of prediction fluctuation entropy, for example, for the load prediction of critical transmission lines, the stability requirement is high, and the corresponding fluctuation entropy threshold is set low; for the load prediction of general distribution networks, the stability requirement is relatively relaxed, and the corresponding fluctuation entropy threshold can be appropriately relaxed. The prediction accuracy requirement refers to the requirement for the prediction accuracy and error control of the quantile prediction model. The prediction accuracy requirement mainly considers the following factors: the economic requirement of power market transactions, i.e., the prediction error directly affects the power purchase cost and income; the safety margin requirement of power grid equipment, i.e., a large prediction error may cause equipment overload or capacity waste; the refinement requirement of load management, i.e., accurate load prediction helps to achieve refined demand-side management. The prediction accuracy requirement is quantified by an acceptable range of prediction relative error rate, for example, for the load prediction of important industrial users, the accuracy requirement is high, and the corresponding error rate threshold is set low; for the prediction of scattered residential loads, the accuracy requirement is relatively relaxed, and the corresponding error rate threshold can be appropriately relaxed.

[0071] Subsequently, the first type of threshold and the second type of threshold are calculated according to the predicted stability requirement and the predicted accuracy requirement respectively, to obtain the dual-target calibration trigger discrimination limit. The first type of threshold is a volatility index critical value determined according to the predicted stability requirement. The calculation of the first type of threshold comprehensively considers factors such as historical performance baseline, business tolerance and industry standards. The specific calculation method includes: based on historical data, the volatility index distribution of each quantile prediction model is calculated, a reasonable quantile level (such as 95% quantile) is determined as the baseline value; combined with the stability requirement of the target scene, the first type of threshold is formed by appropriately adjusting the baseline value. The setting of the first type of threshold needs to ensure that the stability problem can be identified in time, and also needs to avoid being too sensitive to cause frequent calibration. The second type of threshold is an error index critical value determined according to the predicted accuracy requirement. The calculation method of the second type of threshold is similar to that of the first type of threshold, which is mainly based on the historical error distribution of the prediction model and the business accuracy requirement. The calculation process includes: the historical error index distribution of each quantile prediction model is calculated, the error trend and seasonal characteristics are analyzed; combined with the accuracy requirement of the target scene, an acceptable error level is determined; considering the influence factors such as prediction time scale, load type and external environment, the error threshold is corrected to form the second type of threshold. The dual-target calibration trigger discrimination limit is a composite discrimination standard composed of the first type of threshold and the second type of threshold, which provides clear discrimination basis for the subsequent calibration trigger decision.

[0072] Then, the prediction performance evaluation results are compared with the dual-target calibration trigger discrimination limit. The comparison process is a process of comparing and analyzing the current performance of each quantile prediction model with the preset performance requirement. The comparison process includes two dimensions of comparison: one is to compare the volatility index of each quantile prediction model with the first type of threshold to judge whether the prediction stability of the model meets the requirement; the other is to compare the error index of each quantile prediction model with the second type of threshold to judge whether the prediction accuracy of the model meets the requirement. The comparison result generates a two-dimensional performance evaluation conclusion for each quantile prediction model, providing an objective basis for the calibration trigger decision.

[0073] Subsequently, the quantile prediction model that needs to be calibrated is determined according to the dual-target trigger condition, and the dual-target calibration trigger discrimination result is output. The dual-target trigger condition adopts an "or" logical relationship, that is, when the performance of a quantile prediction model in either stability or accuracy dimension deteriorates, it is determined that the model needs to be calibrated and optimized. The specific trigger conditions include:

[0074] The first trigger condition: the volatility index exceeds the first type of threshold. When the prediction volatility entropy of a certain quantile prediction model exceeds the first type of threshold, it indicates that the prediction stability of the model is insufficient, the volatility degree of the prediction result exceeds the acceptable range of the business, and it may have an adverse effect on power grid dispatching and operation decision, so stability calibration is needed.

[0075] The second trigger condition: the error index exceeds the second threshold. When the prediction relative error rate of a certain quantile prediction model exceeds the second threshold, it indicates that the prediction accuracy of the model is insufficient, the prediction precision is lower than the business requirement, and it may lead to an increase in the operation cost of the power system or an increase in the security risk, and therefore accuracy calibration is required.

[0076] The to-be-calibrated quantile prediction model refers to a quantile prediction model that meets any of the double-target trigger conditions. All models that meet the trigger conditions are marked to form multiple to-be-calibrated quantile prediction models, forming a double-target calibration trigger discrimination result. The double-target calibration trigger discrimination result is a comprehensive result set containing information of all to-be-calibrated quantile prediction models, which provides a clear processing object for subsequent matching model screening, sample data extraction, and progressive calibration steps, ensuring the pertinence and effectiveness of the calibration process.

[0077] Further, in combination with the double-target calibration trigger discrimination result and the preset neighborhood constraint, the multiple quantile prediction models are traversed for performance matching-based model screening to determine multiple matching quantile prediction models and obtain a matching quantile prediction model set, including:

[0078] S31, a first to-be-calibrated quantile prediction model is randomly selected from the double-target calibration trigger discrimination result, and corresponding first quantile verification data is obtained;

[0079] S32, a neighborhood range is determined based on the neighborhood constraint, and performance matching verification is performed on multiple quantile prediction models in the neighborhood range based on the first quantile verification data;

[0080] S33, based on the performance matching verification result, the multiple quantile prediction models in the neighborhood range are serialized, and one-time screening based on a performance threshold and two-time screening based on a quantity are performed;

[0081] S34, a corresponding relationship between the first to-be-calibrated quantile prediction model and the screening result is established, the screening result is defined as the matching quantile prediction model, and the matching quantile prediction model is stored in the matching quantile prediction model set;

[0082] S35, multiple matching quantile prediction models corresponding to the to-be-calibrated quantile prediction models are determined by traversing the double-target calibration trigger discrimination result, and are added to the matching quantile prediction model set;

[0083] Each to-be-calibrated quantile prediction model corresponds to at least one matching quantile prediction model.

[0084] In a preferred embodiment, first, a first to-be-calibrated quantile prediction model is randomly selected from the double-target calibration trigger discrimination results, and corresponding first quantile verification data is obtained. Random selection refers to selecting one to-be-calibrated quantile prediction model as the current processing object, i.e., the first to-be-calibrated quantile prediction model, from all to-be-calibrated quantile prediction models contained in the double-target calibration trigger discrimination results by using a random sampling method. In actual implementation, other selection strategies can also be used, such as selection according to the performance degradation degree, selection according to the quantile value size order, etc. The first to-be-calibrated quantile prediction model refers to the to-be-calibrated quantile prediction model that is currently being screened, which is identified as a model whose performance does not meet the requirements in the double-target calibration trigger discrimination, and a suitable matching model needs to be found to assist in calibration optimization. The first quantile verification data refers to the historical prediction data of the first to-be-calibrated quantile prediction model, which is used to verify whether other models are suitable for the prediction task of the first to-be-calibrated quantile prediction model. The first quantile verification data contains the prediction input features, prediction output results, and corresponding true observation values of the first to-be-calibrated quantile prediction model in the past period of time. These data reflect the prediction task characteristics, data distribution characteristics, and performance level of the first to-be-calibrated quantile prediction model, and provide a standardized test benchmark for evaluating the adaptability of other models to the same task. The time span and data volume of the first quantile verification data are set according to the model type and prediction period, and usually the complete prediction records in the latest dynamic calibration period are selected.

[0085] Subsequently, a neighborhood range is determined based on neighborhood constraints, and the data based on the first quantile is verified, and a plurality of quantile prediction models in the neighborhood range are traversed to perform performance matching verification. The neighborhood range is a quantile prediction model search space determined according to the preset neighborhood constraints. The neighborhood constraints are based on the quantile similarity principle. Prediction models with similar quantile values have certain relevance and complementarity in prediction tasks and data characteristics. The neighborhood range is defined by setting the upper and lower bounds of the quantile difference. For example, for a 0.5 quantile model to be calibrated, if the neighborhood constraints are set to plus or minus 0.2, the neighborhood range is [0.3, 0.7], that is, it includes models with 0.3 quantile, 0.4 quantile, 0.6 quantile, 0.7 quantile, etc. Performance matching verification refers to verifying whether the quantile prediction models in the neighborhood range have certain performance on the prediction task of the first quantile model to be calibrated, and determining other models that have the potential to complete the task of the first quantile model to be calibrated. The performance matching verification process includes: using each quantile prediction model in the neighborhood range to re-predict the first quantile verification data, obtaining the prediction results of each neighborhood model on the calibration task; calculating the performance indicators between the prediction results of each neighborhood model and the true observed values in the verification data, such as mean absolute error, root mean square error, mean absolute percentage error, etc.; based on the performance indicators, evaluating the adaptability and prediction ability of each neighborhood model to the task of the first quantile model to be calibrated. Performance matching verification can identify neighborhood models that perform well on the same prediction task, providing high-quality candidate resources for subsequent knowledge transfer and model calibration.

[0086] Then, based on the performance matching verification result, the plurality of quantile prediction models in the serialized neighborhood range are serialized and subjected to a performance threshold-based first screening and a quantity-based second screening. Serialization refers to sorting the quantile prediction models in the neighborhood range according to their performance on the first to-be-calibrated quantile prediction model task based on the performance matching verification result, forming an ordered model sequence. The sorting is in descending order, i.e., the model with the best performance is placed at the front of the sequence, and the model with the poorer performance is placed at the back of the sequence. Serialization provides an ordered processing basis for subsequent hierarchical screening. The performance threshold-based first screening refers to setting a minimum requirement for performance, and removing neighborhood models with performance lower than the performance threshold. The performance threshold is set according to the historical performance level of the first to-be-calibrated quantile prediction model, the quality of service requirement, and the expected calibration effect. The first screening can filter out neighborhood models that are obviously not suitable as matching models, improving the efficiency and quality of the subsequent calibration process. The quantity-based second screening refers to selecting the top several models with the best performance from the first screening result as the final matching model candidates according to a quantity threshold. The quantity threshold is set according to the calculation resource limit, the calibration complexity, and the expected effect, and is usually set to 3 to 10 models. The second screening can control the number of matching models, avoid the calculation overhead and complexity problems caused by too many models, and at the same time ensure that there are enough candidate models to support the calibration process.

[0087] After that, a corresponding relationship between the first to-be-calibrated quantile prediction model and the screening result is established, the screening result is defined as a matching quantile prediction model, and is stored in a matching quantile prediction model set. The corresponding relationship refers to the mapping relationship between the first to-be-calibrated quantile prediction model and its corresponding matching quantile prediction model. This mapping relationship records which quantile prediction models can be used to assist the first to-be-calibrated quantile prediction model in calibration, as well as the performance of each quantile prediction model on the verification task and the sorting information. The corresponding relationship provides an index basis for subsequent sample data extraction and calibration strategy selection. The matching quantile prediction model refers to a neighborhood quantile prediction model that has the ability to assist the first to-be-calibrated quantile prediction model in calibration optimization after performance matching verification and double screening. The matching model has the following characteristics: located within the neighborhood constraint range of the first to-be-calibrated quantile prediction model, ensuring the correlation between the models; having good performance on the prediction task of the first to-be-calibrated quantile prediction model, meeting the preset performance threshold requirement; the number is controlled within a reasonable range, taking into account the calibration effect and calculation efficiency. The matching quantile prediction model set is a data structure that stores all matching model information, including matching model identification, corresponding to-be-calibrated model, performance verification result, screening sorting, etc.

[0088] The dual-target calibration trigger discriminates each to-be-calibrated quantile prediction model to repeat the performance model screening process of steps S31 to S34, and determines the corresponding performance model for all to-be-calibrated quantile prediction models. The traversal process ensures that each to-be-calibrated quantile prediction model can obtain a suitable set of performance quantile prediction models for support, and provides sufficient resource guarantee for comprehensive model calibration. The constraint condition that each to-be-calibrated quantile prediction model corresponds to at least one performance quantile prediction model guarantees the feasibility and effectiveness of the calibration process. When a to-be-calibrated quantile prediction model cannot find a performance model that meets the performance requirement within its neighborhood range, the constraint condition can be met by relaxing the neighborhood constraint, reducing the performance threshold, or using other alternative strategies.

[0089] Through the above steps, a complete performance model support system can be established for all to-be-calibrated quantile prediction models, providing high-quality model resources and knowledge sources for subsequent progressive calibration.

[0090] Further, the dual-target calibration trigger discriminates result and the performance quantile prediction model set are indexed to correspondingly extract sample prediction data, and a multi-source sample prediction data set is obtained, including:

[0091] S41, the historical training data of the to-be-calibrated quantile prediction model is called to form standard observation-sample prediction data;

[0092] S42, the historical prediction data of the to-be-calibrated quantile prediction model is collected to form to-be-calibrated model-sample prediction data, wherein the historical prediction data is associated with a prediction probability distribution;

[0093] S43, the performance matching verification result of the performance quantile prediction model is called and classified to construct performance model-sample prediction data;

[0094] S44, the standard observation-sample prediction data, the performance model-sample prediction data, and the to-be-calibrated model-sample prediction data are output as the multi-source sample prediction data set.

[0095] In a preferred embodiment, first, the historical training data of the to-be-calibrated quantile prediction model is called to form standard observation-sample prediction data. The historical training data refers to the original data set used by the to-be-calibrated quantile prediction model in the initial training stage, which contains input feature data and corresponding real label data for model learning. The standard observation-sample prediction data is a standardized sample data set constructed on the basis of the historical training data, which contains the following components: input feature data, including historical load values, meteorological data, time features, holiday identifiers, economic indicators, and other multi-dimensional feature information; real label data, i.e., actual load observation values at the corresponding time, as the objective standard for model learning; data preprocessing information, including data cleaning, normalization, feature engineering, and other processing records. The standard observation-sample prediction data provides benchmark supervision information for the calibration process, ensuring that the calibrated model can maintain effective learning ability for the basic prediction task and avoiding knowledge forgetting or performance degradation problems in the calibration process.

[0096] Subsequently, the historical prediction data of the to-be-calibrated quantile prediction model is collected to form to-be-calibrated model-sample prediction data. The historical prediction data refers to the actual prediction output records of the to-be-calibrated quantile prediction model in the past period of time, which contains the prediction results of the model for each time load, prediction timestamps, input features, and other information. The historical prediction data reflects the current prediction behavior pattern, decision preference, and performance level of the model, providing an important basis for understanding the current situation of the model and identifying the root cause of the problem. Among them, the historical prediction data is associated with a prediction probability distribution. The prediction probability distribution refers to the probability distribution information of the prediction results output by the model, which is usually calculated and represented by a softmax function. The softmax function converts the original output of the model into a probability distribution form, so that the sum of the probabilities of each possible prediction value is equal to 1, facilitating the quantification of the confidence of the model for different prediction results. The prediction probability distribution contains the uncertainty information of the model prediction, which can reflect the confidence level of the model for the prediction results. When the model is very confident about a prediction result, the probability value of that result will be higher, and the probability values of other results will be lower; when the model has great uncertainty about the prediction result, the probability values of each possible result will be closer. The to-be-calibrated model-sample prediction data is a sample data set constructed around the historical prediction data, which contains prediction input features, prediction output results, prediction probability distributions, prediction time information, and other elements. This data set can comprehensively reflect the prediction behavior characteristics and performance of the to-be-calibrated model, providing the model's own experience information and knowledge accumulation for the calibration process.

[0097] Then, the performance matching verification result of the matching sex quantile prediction model is called and classified to build matching sex model-sample prediction data. The performance matching verification result refers to the prediction performance record of the matching sex quantile prediction model on the verification task of the to-be-calibrated model, including the prediction output of the matching sex model on the verification data, the performance index calculation result, the sorting information and the like. The performance matching verification result reflects the adaptability and knowledge transfer potential of the matching sex model on the cross-quantile prediction task. The classification marking refers to the classification process of marking the prediction samples of the matching sex model as positive samples and negative samples according to the performance quality of the matching sex model on the verification task. The specific implementation of the classification marking includes: setting a performance evaluation threshold, and classifying the prediction results of the matching sex model according to the deviation degree of the prediction results from the true values; the samples with a prediction error less than the threshold are marked as positive samples, indicating that the samples represent the excellent prediction experience of the matching sex model and have positive knowledge transfer value; the samples with a prediction error greater than the threshold are marked as negative samples, indicating that the samples represent the prediction deviation of the matching sex model and may have a negative impact on the to-be-calibrated model. Through the classification marking, the effective knowledge and ineffective knowledge in the matching sex model can be distinguished, and the accuracy and calibration effect of knowledge transfer can be improved. The matching sex model-sample prediction data is a sample data set constructed based on the performance matching verification result and the classification marking, including the prediction input features, the prediction output results, the positive and negative sample markings, the performance evaluation information and the like of the matching sex model. The data set carries the prediction experience of the adjacent quantile model and the cross-quantile knowledge, and provides external learning reference and knowledge supplement for the to-be-calibrated model.

[0098] After that, the standard observation-sample prediction data, the matching sex model-sample prediction data and the to-be-calibrated model-sample prediction data are output as a multi-source sample prediction data set. The multi-source sample prediction data set provides comprehensive and rich training resources for subsequent progressive calibration, and can support the implementation of various calibration strategies such as incremental calibration, positive transfer calibration and reverse transfer calibration. The standard observation-sample prediction data provides a reliable supervision benchmark to ensure that the calibration process does not deviate from the basic prediction target; the to-be-calibrated model-sample prediction data provides model status information to help identify and correct prediction deviation; and the matching sex model-sample prediction data provides external knowledge supplement to promote cross-quantile knowledge transfer and model fusion.

[0099] By constructing the multi-source sample prediction data set, the data information and knowledge resources of different sources can be fully utilized to provide multi-dimensional calibration support for the to-be-calibrated quantile prediction model, and the comprehensiveness and effectiveness of the calibration effect can be significantly improved.

[0100] Further, the multi-modal calibration strategy at least includes incremental calibration, positive transfer calibration and reverse transfer calibration, and the double-target calibration trigger discrimination result is subjected to progressive calibration, including:

[0101] S51, performing incremental calibration based on incremental learning on the to-be-calibrated quantile prediction model with the standard observation-sample prediction data as supervision, and verifying a first prediction performance after the incremental calibration;

[0102] S52, if the first prediction performance does not meet the prediction performance requirement, performing forward transfer calibration based on supervised learning on the to-be-calibrated quantile prediction model with the standard observation-sample prediction data as a hard target and the to-be-calibrated model-sample prediction data as a soft target, and verifying a second prediction performance after the forward transfer calibration;

[0103] S53, if the second prediction performance does not meet the prediction performance requirement, performing backward transfer calibration based on supervised learning on the to-be-calibrated quantile prediction model with the standard observation-sample prediction data as a hard target and the to-be-calibrated model-sample prediction data as a soft target, and verifying a third prediction performance after the backward transfer calibration;

[0104] S54, if the third prediction performance meets the prediction performance requirement, outputting the to-be-calibrated quantile prediction model after the backward transfer calibration for power load quantile prediction.

[0105] In a preferred embodiment, the multi-modal calibration strategy at least includes the incremental calibration, the forward transfer calibration and the backward transfer calibration.

[0106] When performing incremental calibration, the standard observation-sample prediction data is used as supervision to perform incremental calibration of the to-be-calibrated quantile prediction model based on incremental learning, and the first prediction performance after incremental calibration is verified. Incremental calibration is a model optimization method based on incremental learning theory, aiming to fine-tune and optimize the model parameters using new training data while maintaining the original knowledge structure of the model. The core idea of incremental calibration is to improve the prediction performance of the model by adjusting the local parameters without retraining the entire model, which has the characteristics of high computational efficiency and good knowledge retention. The incremental calibration process based on incremental learning includes the following steps: first, load the current parameter state of the to-be-calibrated quantile prediction model as the initial point of incremental learning; second, use the standard observation-sample prediction data as the supervision signal, and use gradient descent and other optimization algorithms to fine-tune and update the model parameters; third, use a small learning rate and a limited number of training rounds to ensure that the parameter update is moderate and avoid excessive disturbance to the original knowledge; finally, monitor the loss function changes during training, and stop training when the loss converges or reaches the preset number of rounds. The technical advantages of incremental calibration are: maintaining the stability of the model structure, avoiding significant model reconstruction; using the supervision information of historical training data to ensure the correctness of the calibration direction; the computational complexity is relatively low, suitable for real-time or quasi-real-time model maintenance scenarios. The first prediction performance refers to the performance of the to-be-calibrated quantile prediction model on the standard test data set after incremental calibration, including the recalculation results of volatility indicators and error indicators. The first prediction performance verification process uses the same evaluation method as step S1 to ensure the consistency and comparability of the evaluation standard.

[0107] If the first prediction performance does not meet the prediction performance requirement, a forward transfer calibration phase is entered. The forward transfer calibration is a model optimization method based on the knowledge transfer theory, and the to-be-calibrated model is optimized by simultaneously using external knowledge sources and internal supervision signals. The forward transfer calibration adopts a double-target learning strategy, which not only ensures the fitting ability of the model to the true label, but also promotes the effective absorption of the model to the external knowledge. Specifically, the soft target of the matching model-sample prediction data and the hard target of the standard observation-sample prediction data are used, and the forward transfer calibration based on supervised learning is performed by minimizing the loss of the to-be-calibrated quantile prediction model for the soft target and the hard target, and the second prediction performance after the forward transfer calibration is verified. The soft target refers to the prediction result derived from the matching model-sample prediction data, which represents the prediction experience and knowledge of the adjacent quantile model. The soft target contains the solution of the matching model to the same or similar prediction task, which can provide the prediction idea and decision reference for the to-be-calibrated model. The hard target refers to the true load observation value in the standard observation-sample prediction data, which represents the objective supervision information and absolute learning standard. The forward transfer calibration process realizes double-target optimization by constructing a joint loss function, which includes two components: the hard target loss, which measures the deviation between the model prediction result and the true observation value; and the soft target loss, which measures the difference between the model prediction result and the matching model prediction result. The forward transfer calibration balances the attention degree of the model to accuracy and knowledge transfer by adjusting the weight ratio of the two loss terms, and realizes supervised knowledge fusion learning. The second prediction performance refers to the performance evaluation result of the to-be-calibrated quantile prediction model after the forward transfer calibration is completed.

[0108] If the second prediction performance does not meet the prediction performance requirement, a reverse migration calibration phase is entered. The reverse migration calibration is a deep calibration method using the reverse knowledge migration idea, and is suitable for the case that the performance of the model to be calibrated is seriously degraded and the effect of the traditional calibration method is limited. The core idea of the reverse migration calibration is to take the performance better matching quantile prediction model as the basic framework, to fuse the prediction experience and real observation information of the model to be calibrated, and to construct a new optimized model. Specifically, taking the matching quantile prediction model as the base model, taking the standard observation-sample prediction data as the hard target, taking the model to be calibrated-sample prediction data as the soft target, and taking the simultaneous minimization of the loss of the base model for the soft target and the hard target as the goal, the reverse migration calibration based on supervised learning is carried out, and the third prediction performance after the reverse migration calibration is verified. Among them, the base model refers to the matching model with the best performance selected from the matching quantile prediction model set. The model performs well in the cross-quantile prediction task and has the potential to undertake new prediction tasks. The base model provides a stable starting point and a reliable model architecture for the reverse migration calibration. The reverse migration calibration process includes: taking the matching quantile prediction model as the initial model, retaining its network structure and most of the parameters; taking the standard observation-sample prediction data as the main supervision signal to ensure the adaptability of the new model to the target prediction task; taking the model to be calibrated-sample prediction data as the auxiliary soft target to fuse the effective prediction experience of the original model to be calibrated; and through joint loss function optimization, the effective fusion of multi-source knowledge and the significant improvement of model performance are realized. The technical advantages of the reverse migration calibration are: taking the high-performance model as the starting point, avoiding the optimization difficulty starting from the low-performance state; fully utilizing the cross-quantile knowledge, expanding the learning resources of the model; and using multi-source information fusion, improving the comprehensiveness and robustness of the calibration effect. The third prediction performance refers to the performance evaluation result of the base model on the target prediction task after the reverse migration calibration is completed. If the third prediction performance meets the prediction performance requirement, the base model after the reverse migration calibration is output for power load quantile prediction.

[0109] The progressive calibration strategy has the following technical features: progressive calibration strength, from mild incremental calibration to moderate forward migration calibration, and then to deep reverse migration calibration, gradually increasing the calibration strength; progressive calculation complexity, preferentially trying the calibration method with less calculation overhead, and only using the complex calibration strategy when necessary; progressive knowledge utilization, from single historical data supervision to multi-source knowledge fusion, fully mining and utilizing various available information; performance guarantee mechanism, through step-by-step verification and early stopping strategy, ensuring the effectiveness and efficiency of the calibration process.

[0110] Through the progressive calibration strategy, the appropriate calibration method and calibration strength can be adaptively selected according to the degree and type of model performance degradation, thereby ensuring the reliability of the calibration effect and taking into account the calculation efficiency and resource consumption, and providing an efficient, flexible and robust model maintenance solution for power load quantile prediction.

[0111] Further, the embodiments of the application also include:

[0112] S61, obtaining historical prediction performance evaluation results to form a prediction performance evaluation result sequence;

[0113] S62, performing time series analysis on the prediction performance evaluation result sequence and the volatility index to calculate a real-time volatility trend;

[0114] S63, performing time series analysis on the prediction performance evaluation result sequence and the error index to calculate a real-time error trend;

[0115] S64, weighting the calibration severity coefficients of the plurality of quantile prediction models according to the volatility index, the error index, the real-time volatility trend and the real-time error trend;

[0116] S65, differentially updating the dynamic calibration period according to the plurality of calibration severity coefficients.

[0117] In a preferred embodiment, first, historical prediction performance evaluation results are obtained to form a prediction performance evaluation result sequence. The historical prediction performance evaluation results refer to the performance evaluation records of each quantile prediction model in the past plurality of calibration periods, including the corresponding volatility index value, error index value, comprehensive performance score and calibration trigger condition of each calibration period and other information. The historical prediction performance evaluation results reflect the change law of model performance over time, providing a data basis for analyzing model performance trends and predicting future performance trends. The prediction performance evaluation result sequence is a time series data structure formed by arranging the historical prediction performance evaluation results in chronological order. The sequence contains performance evaluation snapshots at multiple time points, which can clearly show the historical evolution trajectory of model performance. The time span of the sequence is set according to business needs and data availability, usually covering complete records of the last several calibration periods to ensure the reliability and representativeness of trend analysis. The prediction performance evaluation result sequence provides structured data input for subsequent time series analysis and trend calculation.

[0118] Then, the prediction performance evaluation result sequence and the volatility index are combined for time series analysis to calculate the real-time volatility trend. Time series analysis refers to the process of statistically analyzing and mining rules from the prediction performance evaluation result sequence using time series analysis methods. Time series analysis methods include trend analysis, periodicity analysis, smoothing processing, difference operation, and other technical means, which can extract valuable change rules and trend information from time series data. The real-time volatility trend refers to the change direction and degree of the prediction stability of the quantile prediction model in the time dimension. The real-time volatility trend is calculated by the following steps: first, extract the volatility index value at each time point from the prediction performance evaluation result sequence to form a volatility index time series; second, smooth the volatility index time series to eliminate the influence of short-term fluctuations and random noise; third, calculate the first-order difference or trend slope of the time series to quantify the change rate and direction of volatility; and finally, combine the latest volatility index value to comprehensively evaluate the current volatility trend. The real-time volatility trend can reflect the development trend of the model prediction stability, including stability improvement trend, stability deterioration trend, or stability stable trend, and other different states. By analyzing the volatility trend, the development direction of the model stability problem can be identified in advance to provide forward-looking guidance for preventive calibration and calibration cycle adjustment.

[0119] Subsequently, the prediction performance evaluation result sequence and the error index are combined for time series analysis to calculate the real-time error trend. The real-time error trend refers to the change direction and degree of the prediction accuracy of the quantile prediction model in the time dimension. The calculation process of the real-time error trend is similar to that of the real-time volatility trend: extract the error index time series from the prediction performance evaluation result sequence; smooth and analyze the trend of the time series; calculate the change rate and direction of the error index; and combine the latest error index value to evaluate the current error trend. The real-time error trend can reflect the evolution of the model prediction accuracy, including accuracy improvement trend, accuracy decline trend, or accuracy stable trend, and other different modes. Error trend analysis helps to identify the development rules of model precision problems to provide a basis for accurate calibration decisions and resource allocation.

[0120] Then, the calibration severity coefficients of the plurality of quantile prediction models are evaluated according to the volatility index, the error index, the real-time volatility trend and the real-time error trend. The calibration severity coefficient is a comprehensive index quantifying the urgency of calibration demand and the strength requirement of calibration for each quantile prediction model. The calibration severity coefficient comprehensively considers the current performance state and future performance development trend of the model, and provides a quantitative basis for formulating differentiated calibration strategies. The weighted evaluation process includes the following steps: first, the volatility index, the error index, the real-time volatility trend and the real-time error trend are normalized to eliminate the influence of different index dimensions and numerical ranges; second, the weight coefficients are set according to the influence of each index on the calibration demand, usually the weight of the current performance index is higher, and the weight of the trend index is relatively lower, but the specific weight allocation can be adjusted according to the application scene and business demand; third, the weighted comprehensive score of each quantile prediction model is calculated, which reflects the comprehensive calibration demand degree of the model; finally, the calibration severity coefficient is determined based on the weighted comprehensive score, and the higher the score, the more urgent the calibration demand, and the greater the corresponding calibration severity coefficient. The calibration severity coefficient provides a quantitative standard for subsequent calibration cycle adjustment, so that the system can implement differentiated calibration management strategies according to the actual demand status of different models.

[0121] Then, the dynamic calibration cycle is updated differently according to the plurality of calibration severity coefficients. Differentiated updating means that different calibration cycles are set for different models according to the calibration severity coefficients of the plurality of quantile prediction models, to realize personalized calibration time management. The core idea of differentiated updating is the optimal allocation of calibration resources, which focuses the limited computing resources and time resources on the models with the most urgent calibration demand. The dynamic calibration cycle updating strategy includes: for the models with higher calibration severity coefficients, the calibration cycle is shortened, the calibration frequency is increased, intensive monitoring and timely calibration are realized; for the models with lower calibration severity coefficients, the calibration cycle is extended, the calibration frequency is reduced, unnecessary computing overhead is avoided; for the models with medium calibration severity coefficients, the current calibration cycle is maintained or adjusted slightly. The adjustment range of the dynamic calibration cycle is usually limited by upper and lower bounds to avoid excessive system load caused by too short calibration cycle or performance monitoring lag caused by too long calibration cycle. The adjustment range can be realized by fixed step, proportional adjustment or adaptive adjustment.

[0122] Through dynamic calibration cycle optimization based on historical performance change trend, more intelligent and refined model maintenance management can be realized, not only improving the pertinence and effectiveness of calibration effect, but also optimizing the utilization efficiency of computing resources, providing sustainable technical support for long-term stable operation of the power load quantile prediction system.

[0123] Further, if any one of the first prediction performance and the second prediction performance meets the prediction performance requirement, the progressive calibration is interrupted immediately, and a corresponding one of the incremental calibration result and the forward migration calibration result is output for power load quantile prediction.

[0124] In a preferred embodiment, the progressive calibration strategy interruption mechanism specifically includes: if any one of the first prediction performance and the second prediction performance meets the prediction performance requirement, the progressive calibration is interrupted immediately, and a corresponding one of the incremental calibration result and the forward migration calibration result is output for power load quantile prediction. Specifically, in the progressive calibration process, when the performance of a certain calibration stage meets the preset requirement, the subsequent calibration steps are terminated immediately, and the optimization strategy of the current calibration result is output, thereby avoiding unnecessary over-optimization and waste of computing resources, and reducing the risk of overfitting and model complexity.

[0125] Specifically, after the incremental calibration of step S51 is completed, the first prediction performance is verified. If both the volatility index and the error index in the first prediction performance meet the prediction performance requirement set in step S21, that is, the volatility index does not exceed the first threshold and the error index does not exceed the second threshold, it is determined that the incremental calibration effect is sufficient, and subsequent forward migration calibration and reverse migration calibration are not needed. At this time, the progressive calibration process is interrupted immediately, and the incremental calibration result is directly output as the final calibration model.

[0126] When the first prediction performance does not meet the prediction performance requirement, but the second prediction performance meets the prediction performance requirement after the forward migration calibration of step S52 is completed, the immediate interruption operation is also performed. At this time, the forward migration calibration result is output as the final calibration model, and the reverse migration calibration with the highest computational complexity does not need to be performed.

[0127] Through the immediate interruption mechanism, efficient, accurate, and adaptive model calibration is achieved, which provides a calibration solution for the power load quantile prediction system that guarantees quality and efficiency, and can provide the optimal calibration strategy selection under different performance degradation scenarios.

[0128] Further, based on the performance matching verification result, the plurality of quantile prediction models in the neighborhood range are serialized, and one-time filtering based on a performance threshold and secondary filtering based on a quantity are performed, including:

[0129] S331, based on the performance matching verification result, the plurality of quantile prediction models are arranged in descending order to obtain a model sequence;

[0130] S332, based on the performance threshold, the model sequence is traversed for one-time filtering, and if the number of models of the one-time filtering result is less than a quantity threshold, the one-time filtering result is directly fed back as the filtering result.

[0131] S333, if the number of models in the first screening result is greater than the number threshold, extracting the first number threshold outputs in the first screening result as the screening result.

[0132] In a preferred embodiment, first, based on the performance matching verification result, the plurality of quantile prediction models are ranked in descending order to obtain a model sequence. The descending order ranking means that according to the performance of each neighborhood quantile prediction model in the performance matching verification, the models are sorted in order from high to low according to the performance index. The performance index usually adopts a comprehensive evaluation score or a single key indicator (such as the reciprocal of the mean absolute error, the prediction accuracy, etc.), to ensure that the better the performance of the model, the higher the ranking. The model sequence is an ordered model list formed after performance sorting, which clearly reflects the relationship between the adaptability of each neighborhood model to the calibration task. The model sequence provides a clear priority order for subsequent hierarchical screening, so that the candidate model with the best performance can be selected first. Each model record in the sequence contains model identification, performance evaluation score, ranking position, etc. information, which is convenient for subsequent screening operation and result tracing.

[0133] Then, based on the performance threshold, the model sequence is traversed for a first screening. If the number of models in the first screening result is less than the number threshold, the first screening result is directly fed back as the screening result. The performance threshold is a preset minimum performance requirement standard, which is used to filter out neighborhood models with obviously insufficient performance. The performance threshold is set by considering factors such as the historical performance level of the calibration model to be calibrated, the quality of service requirements, and the expected calibration effect. The threshold setting needs to ensure that the selected matching models have basic calibration assistance capabilities, and also needs to avoid too strict standards leading to too few available models. The first screening means that according to the sorting order of the model sequence, the performance evaluation score of each model is checked one by one to see if it meets the performance threshold requirement. The screening process starts from the top of the sequence (the model with the best performance), and then traverses down, retaining the models with performance scores higher than or equal to the performance threshold, and removing the models with performance scores lower than the performance threshold. The first screening can effectively filter out candidates that are obviously not suitable as matching models, improving the quality and efficiency of the subsequent calibration process.

[0134] The number threshold is a preset upper limit of the number of matching models, which is used to control the number of models in the final screening result, to avoid the problem of high computational complexity caused by too many matching models. The number threshold needs to balance the calibration effect and the computational cost, and is usually set to a range of 3 to 10 models. When the number of models in the first screening result is less than the number threshold, it indicates that there are relatively few candidate models that meet the performance requirements, and in this case, the entire first screening result is directly used as the final screening result, without the need for a second screening. This situation usually occurs in scenarios where the overall performance of the neighborhood models is not high, or the performance threshold is set relatively strictly.

[0135] If the number of the first screening result is greater than the number threshold, the first screening result is extracted to obtain the screening result. The second screening is a fine screening process for further reducing the number of the matching model when the number of the first screening result exceeds the number limit. The second screening selects a number of models with the best performance from the candidate models meeting the performance requirements based on the principle of "selecting the best from the best". The operation process of extracting the first number threshold model is as follows: since the first screening result has been sorted according to the performance, the system directly selects the first number threshold model from the top of the sorting result as the final matching quantile prediction model. This selection method ensures that the final screening result contains the candidate models with the best performance, providing high-quality model resources for the subsequent calibration process. The technical advantages of the second screening are: accurate number control, strictly limiting the number of matching models according to the preset number threshold; reliable quality guarantee, preferentially selecting candidate models with the best performance; controllable calculation complexity, effectively controlling the calculation overhead of the subsequent calibration process by limiting the number of matching models.

[0136] Through the combination strategy of the two screenings, fine screening management of the matching quantile prediction model is realized, which not only guarantees the quality requirements of the screening result, but also meets the needs of number control and calculation efficiency. The screening mechanism can adapt to different quality distribution of the neighborhood model, and can be strictly screened when the candidate model is sufficient, and can be moderately relaxed when the candidate model is scarce, which embodies good adaptability and flexibility. The screening result provides a set of matching models that have been verified in quality and optimized in quantity for subsequent sample data extraction and progressive calibration, ensuring that the calibration process can fully utilize high-quality external knowledge resources, while avoiding the problem of calculation efficiency caused by too many models, laying a foundation for improving the overall calibration effect.

[0137] Embodiment two, as shown in Figure 2 based on the same inventive concept of the power load quantile prediction dynamic calibration method provided in embodiment one, the embodiment of the present application also provides a power load quantile prediction dynamic calibration system, comprising:

[0138] The prediction performance evaluation module 11 is used for interacting with the target scene, obtaining quantile prediction models of a plurality of target quantiles, and performing prediction performance evaluation on each quantile prediction model based on a preset dynamic calibration period;

[0139] The double-target trigger discrimination module 12 is used for double-target calibration trigger discrimination based on the prediction performance evaluation result;

[0140] The performance matching screening module 13 is configured to screen the plurality of quantile prediction models based on performance matching by combining the double-target calibration trigger discrimination result and the preset neighborhood constraint.

[0141] The sample data extraction module 14 is configured to extract sample prediction data corresponding to the double-target calibration trigger discrimination result and the set of quantile prediction models to obtain a plurality of source sample prediction data sets.

[0142] The progressive calibration execution module 15 is configured to perform progressive calibration on the double-target calibration trigger discrimination result based on the plurality of source sample prediction data sets and the set of quantile prediction models in combination with a preset multi-modal calibration strategy.

[0143] Further, the execution steps of the prediction performance evaluation module 11 include:

[0144] Based on a predetermined size of a sliding window, the prediction volatility entropy of each quantile prediction model is monitored and calculated in real time, and the output is a volatility index.

[0145] In combination with a time sequence weighting mechanism, the prediction relative error rate of each quantile prediction model is calculated, and the output is an error index, wherein the time sequence weighting weight decreases along the time distance direction.

[0146] In combination with the volatility index and the error index, the prediction performance evaluation result is formed.

[0147] Further, the execution steps of the double-target trigger discrimination module 12 include:

[0148] The prediction performance requirements of the target scene are obtained, including prediction stability requirements and prediction accuracy requirements.

[0149] The first threshold and the second threshold are calculated according to the prediction stability requirements and the prediction accuracy requirements, respectively, to obtain the double-target calibration trigger discrimination limit.

[0150] The prediction performance evaluation result is compared with the double-target calibration trigger discrimination limit.

[0151] If any of the double-target trigger conditions is met, it is determined that the corresponding quantile prediction model needs to be calibrated, and the corresponding label is marked as a to-be-calibrated quantile prediction model, wherein the double-target trigger conditions include:

[0152] The volatility index exceeds the first threshold;

[0153] The error index exceeds the second threshold;

[0154] Output a plurality of the to-be-calibrated quantile prediction models as a double-target calibration trigger discrimination result.

[0155] Further, the execution steps of the performance matching screening module 13 include:

[0156] In the double-target calibration trigger discrimination result, randomly select a first to-be-calibrated quantile prediction model and correspondingly obtain first quantile verification data;

[0157] Determine a neighborhood range based on the neighborhood constraint, and based on the first quantile verification data, traverse a plurality of quantile prediction models in the neighborhood range to perform performance matching verification;

[0158] Based on the performance matching verification result, serialize a plurality of quantile prediction models in the neighborhood range, and perform one screening based on a performance threshold and two screenings based on a quantity;

[0159] Establish a correspondence between the first to-be-calibrated quantile prediction model and the screening result, define the screening result as the adaptive quantile prediction model, and store it in the adaptive quantile prediction model set;

[0160] Traverse the double-target calibration trigger discrimination result to determine a plurality of adaptive quantile prediction models corresponding to the to-be-calibrated quantile prediction model, and add them to the adaptive quantile prediction model set;

[0161] Wherein, each to-be-calibrated quantile prediction model corresponds to at least one adaptive quantile prediction model.

[0162] Further, the execution steps of the sample data extraction module 14 include:

[0163] Call the historical training data of the to-be-calibrated quantile prediction model to form standard observation-sample prediction data;

[0164] Collect historical prediction data of the to-be-calibrated quantile prediction model to form to-be-calibrated model-sample prediction data, wherein the historical prediction data is associated with a prediction probability distribution;

[0165] Call the performance matching verification result of the adaptive quantile prediction model and perform classification labeling to construct adaptive model-sample prediction data;

[0166] Output the standard observation-sample prediction data, the adaptive model-sample prediction data, and the to-be-calibrated model-sample prediction data as the multi-source sample prediction data set.

[0167] Further, the multi-modal calibration strategy at least includes incremental calibration, forward transfer calibration, and reverse transfer calibration, and the execution steps of the progressive calibration execution module 15 include:

[0168] performing incremental calibration based on incremental learning on the to-be-calibrated quantile prediction model with the standard observation-sample prediction data as supervision, and verifying a first prediction performance after the incremental calibration;

[0169] If the first prediction performance does not meet the prediction performance requirement, performing forward transfer calibration based on supervised learning with the matching model-sample prediction data as a soft target, the standard observation-sample prediction data as a hard target, and a target of simultaneously minimizing losses of the to-be-calibrated quantile prediction model for the soft target and the hard target, and verifying a second prediction performance after the forward transfer calibration;

[0170] If the second prediction performance does not meet the prediction performance requirement, performing backward transfer calibration based on supervised learning with the matching quantile prediction model as a base model, the standard observation-sample prediction data as a hard target, and the to-be-calibrated model-sample prediction data as a soft target, and a target of simultaneously minimizing losses of the base model for the soft target and the hard target, and verifying a third prediction performance after the backward transfer calibration;

[0171] If the third prediction performance meets the prediction performance requirement, outputting the base model after the backward transfer calibration for power load quantile prediction.

[0172] Further, the embodiments of the present application further include a dynamic period adjustment module, and the execution steps of the dynamic period adjustment module include:

[0173] obtaining a historical prediction performance evaluation result to form a prediction performance evaluation result sequence;

[0174] performing time series analysis on the prediction performance evaluation result sequence and the volatility index to calculate a real-time volatility trend;

[0175] performing time series analysis on the prediction performance evaluation result sequence and the error index to calculate a real-time error trend;

[0176] weighting evaluation of calibration severity coefficients of a plurality of the quantile prediction models according to the volatility index, the error index, the real-time volatility trend and the real-time error trend;

[0177] differentially updating the dynamic calibration period according to a plurality of the calibration severity coefficients.

[0178] Further, if either of the first prediction performance and the second prediction performance meets the prediction performance requirement, interrupting the progressive calibration in real time and outputting a corresponding one of the incremental calibration result and the forward transfer calibration result for power load quantile prediction.

[0179] Further, the execution steps of the performance matching screening module 13 further include:

[0180] Based on the performance matching verification result, the plurality of quantile prediction models are arranged in descending order to obtain a model sequence;

[0181] Based on the performance threshold, one screening is performed on the model sequence, if the number of models in the one screening result is less than the number threshold, the one screening result is directly fed back as the screening result;

[0182] If the number of models in the one screening result is greater than the number threshold, the first number threshold outputs in the one screening result are extracted as the screening result.

[0183] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0184] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0185] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks. Figure 1 The functions specified in one or more flows and / or blocks.

[0186] These computer program instructions can also be stored in a computer-readable memory capable of guiding a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including instruction devices, which implement the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks. Figure 1 The functions specified in one or more flows and / or blocks.

[0187] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operations steps are performed on the computer or other programmable devices to generate computer-implemented processes, thus the instructions executed on the computer or other programmable devices provide the function of realizing the processes specified in the flowcharts Figure 1 one flowchart or multiple flowcharts and / or blocks Figure 1 one block or multiple blocks to realize the function specified in the flowcharts

[0188] Although the preferred embodiments of the application have been described, those skilled in the art will be able to make additional modifications and variations to these embodiments without departing from the spirit and scope of the application.

[0189] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and its equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A method of dynamic calibration of electric power load quantile forecasts, characterized in that, The method comprises the following steps: An interactive target scene is obtained, a plurality of target quantile prediction models are obtained, and the prediction performance of each quantile prediction model is evaluated based on a preset dynamic calibration cycle; Based on the prediction performance evaluation result, a double-target calibration trigger discrimination is performed; Combined with the double-target calibration trigger discrimination result and a preset neighborhood constraint, a performance matching-based model screening is performed on the plurality of quantile prediction models to determine a plurality of adaptive quantile prediction models, and an adaptive quantile prediction model set is obtained; Using the double-target calibration trigger discrimination result and the adaptive quantile prediction model set as an index, sample prediction data is extracted correspondingly, and a multi-source sample prediction data set is obtained; According to the multi-source sample prediction data set and the adaptive quantile prediction model set, the double-target calibration trigger discrimination result is progressively calibrated based on a preset multi-modal calibration strategy; Wherein, obtaining the adaptive quantile prediction model set comprises: In the double-target calibration trigger discrimination result, a first to-be-calibrated quantile prediction model is randomly selected, and first quantile verification data is correspondingly obtained; Based on the neighborhood constraint, a neighborhood range is determined, and performance matching verification is performed on a plurality of quantile prediction models in the neighborhood range based on the first quantile verification data; Based on the performance matching verification result, the plurality of quantile prediction models in the neighborhood range are serialized, and a performance threshold-based primary screening and a quantity-based secondary screening are performed; A corresponding relationship between the first to-be-calibrated quantile prediction model and the screening result is established, the screening result is defined as the adaptive quantile prediction model, and the adaptive quantile prediction model set is stored; The double-target calibration trigger discrimination result is traversed, and the adaptive quantile prediction model corresponding to each to-be-calibrated quantile prediction model is determined and added to the adaptive quantile prediction model set; Wherein, each to-be-calibrated quantile prediction model corresponds to at least one adaptive quantile prediction model.

2. The method of dynamic calibration of electrical load quantile forecasts of claim 1, wherein, An interactive target scene is obtained, a plurality of target quantile prediction models are obtained, and the prediction performance of each quantile prediction model is evaluated based on a preset dynamic calibration cycle, which comprises: Based on a sliding window of a predetermined size, the prediction fluctuation entropy of each quantile prediction model is monitored and calculated in real time, and the output is a volatility index; Combined with a time sequence weighting mechanism, the prediction relative error rate of each quantile prediction model is calculated, and the output is an error index, wherein the time sequence weighting weight decreases along the time distance direction; Combined with the volatility index and the error index, the prediction performance evaluation result is formed.

3. The method of dynamic calibration of electrical load quantile forecasts of claim 2, wherein, Based on the prediction performance evaluation result, a double-target calibration trigger discrimination is performed, which comprises: The prediction performance requirements of the target scene are obtained, including prediction stability requirements and prediction accuracy requirements; According to the prediction stability requirements and the prediction accuracy requirements, a first threshold and a second threshold are calculated respectively, and a double-target calibration trigger discrimination limit is obtained; The prediction performance evaluation result is compared with the double-target calibration trigger discrimination limit; If any one of the double-target trigger conditions is met, it is determined that the corresponding quantile prediction model needs to be calibrated, and the corresponding label is a to-be-calibrated quantile prediction model, wherein the double-target trigger conditions include: The volatility index exceeds the first type of threshold value; The error index exceeds the second type of threshold value; Output the to-be-calibrated quantile prediction model as a double-target calibration trigger discrimination result.

4. The method of dynamic calibration of electrical load quantile forecasts of claim 3, wherein, With the double-target calibration trigger discrimination result and the set of matching quantile prediction models as indexes, corresponding sample prediction data is extracted to obtain a multi-source sample prediction data set, including: Call the historical training data of the to-be-calibrated quantile prediction model to form standard observation-sample prediction data; Collect historical prediction data of the to-be-calibrated quantile prediction model to form to-be-calibrated model-sample prediction data, wherein the historical prediction data is associated with a prediction probability distribution; Call the performance matching verification result of the matching quantile prediction model and perform classification labeling to construct a matching model-sample prediction data; Output the standard observation-sample prediction data, the matching model-sample prediction data, and the to-be-calibrated model-sample prediction data as the multi-source sample prediction data set.

5. The method of dynamic calibration of electrical load quantile forecasts of claim 4, wherein, The multi-modal calibration strategy at least includes incremental calibration, forward transfer calibration, and reverse transfer calibration, and the double-target calibration trigger discrimination result is progressively calibrated, including: With the standard observation-sample prediction data as supervision, the to-be-calibrated quantile prediction model is incrementally calibrated based on incremental learning, and the first prediction performance after incremental calibration is verified; If the first prediction performance does not meet the prediction performance requirement, the matching model-sample prediction data is used as a soft target, the standard observation-sample prediction data is used as a hard target, and the to-be-calibrated quantile prediction model is forward transferred based on supervised learning while minimizing the loss of the soft target and the hard target, and the second prediction performance after forward transfer calibration is verified; If the second prediction performance does not meet the prediction performance requirement, the matching quantile prediction model is used as a base model, the standard observation-sample prediction data is used as a hard target, and the to-be-calibrated model-sample prediction data is used as a soft target, and the base model is backward transferred based on supervised learning while minimizing the loss of the soft target and the hard target, and the third prediction performance after reverse transfer calibration is verified; If the third prediction performance meets the prediction performance requirement, the base model after reverse transfer calibration is output for power load quantile prediction.

6. The method of dynamic calibration of electrical load quantile forecasts of claim 2, wherein, Further including: Obtain a historical prediction performance evaluation result to form a prediction performance evaluation result sequence; Combine the prediction performance evaluation result sequence and the volatility index to perform time series analysis to calculate a real-time volatility trend; Combine the prediction performance evaluation result sequence and the error index to perform time series analysis to calculate a real-time error trend; According to the volatility index, the error index, the real-time volatility trend, and the real-time error trend, the calibration severity coefficient of the quantile prediction model is weighted and evaluated. According to the plurality of calibration weight coefficients, the dynamic calibration period is updated differentially.

7. The method of dynamic calibration of electrical load quantile forecasts of claim 5, wherein, If any one of the first prediction performance and the second prediction performance meets the prediction performance requirement, the progressive calibration is interrupted in real time, and a corresponding one of an incremental calibration result and a positive migration calibration result is output for power load quantile prediction.

8. The method of dynamic calibration of electrical load quantile forecasts of claim 1, wherein, Based on the performance matching verification result, the plurality of quantile prediction models within the neighborhood range are serialized, and a performance threshold-based primary screening and a quantity-based secondary screening are performed, including: Based on the performance matching verification result, the plurality of quantile prediction models are arranged in descending order to obtain a model sequence; Based on the performance threshold, the model sequence is traversed for primary screening, and if the number of models of the primary screening result is less than a quantity threshold, the primary screening result is directly fed back as the screening result; If the number of models of the primary screening result is greater than the quantity threshold, the first quantity threshold of the primary screening result is extracted as the screening result.

9. A dynamic calibration system for electric load quantile forecasts, characterized by A dynamic calibration method for implementing power load quantile prediction according to any one of claims 1 to 8, comprising: a prediction performance evaluation module configured to interact with a target scene, obtain quantile prediction models of a plurality of target quantiles, and perform prediction performance evaluation on each quantile prediction model based on a preset dynamic calibration period; a double-target trigger discrimination module configured to perform double-target calibration trigger discrimination based on the prediction performance evaluation result; a performance matching screening module configured to combine the double-target calibration trigger discrimination result and a preset neighborhood constraint, traverse the plurality of quantile prediction models, and perform model screening based on performance matching to determine a plurality of adaptive quantile prediction models and obtain an adaptive quantile prediction model set; a sample data extraction module configured to extract sample prediction data corresponding to the double-target calibration trigger discrimination result and the adaptive quantile prediction model set as an index to obtain a multi-source sample prediction data set; a progressive calibration execution module configured to perform progressive calibration on the double-target calibration trigger discrimination result based on the multi-source sample prediction data set and the adaptive quantile prediction model set in combination with a preset multi-modal calibration strategy.

Images