Adaptive prediction model orchestration method and system based on load behavior profiling

Abstract

The application discloses a kind of based on the adaptive prediction model arrangement method and system of load behavior image, the method is carried out window construction to historical load time series data, extracts the load behavior fingerprint in sliding behavior window and carries out image processing, generates candidate training configuration set, carries out training verification to candidate configuration under unified verification protocol;With verification evaluation, tolerance threshold and configuration complexity determine target prediction model configuration, generate arrangement audit record;Through detecting behavior drift and combining audit record, re-arrangement control is carried out.The application can adaptively select the prediction model configuration matched with the current power consumption behavior of user, improve the stability, generalization performance and auditable of multi-user load prediction.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting and intelligent energy management technology, and in particular to an adaptive forecasting model arrangement method and system based on load behavior profiling. Background Technology

[0002] With the rapid development of smart grids and advanced metering systems, load forecasting on the residential user side plays an increasingly important role in demand response, energy efficiency management, and distributed energy optimization and scheduling. However, user-side load forecasting faces significant challenges due to user heterogeneity and time non-stationarity. Different users' electricity consumption behaviors may exhibit various patterns, such as stable periodicity, weak periodicity with high volatility, high zero-value peaks, and large-sample low zero-value patterns. Even the same user may experience shifts in their electricity consumption patterns due to seasonal changes, equipment replacement, changes in living conditions, or adjustments in operating conditions. If a uniform model structure, feature set, and post-processing strategy are applied to all users, it is easy to encounter situations where some users have good forecast accuracy while others experience continuous mismatch, making it difficult to meet the refined forecasting needs of multi-user scenarios.

[0003] Existing load forecasting technologies typically employ the following approaches: First, a unified model approach, where the same forecasting model is trained for all users. This approach is simple to implement but struggles to adapt to diverse user behavior patterns. Second, a manual rule-based grouping approach, which relies on expert experience to categorize users into several types and model them separately. This approach is highly subjective, has high maintenance costs, and is difficult to adapt to dynamically changing user groups over the long term. Third, a general clustering approach, which groups users into clusters based on the similarity of load curve shapes. However, this approach often fails to consider domain-specific features for power load forecasting, such as zero-value ratios, daily cycle autocorrelation, and peak tails, as routing criteria for model configuration, resulting in a lack of effective correlation between grouping results and forecasting model configuration. Fourth, an automated machine learning or grid search approach, which searches a large hyperparameter space to find the optimal configuration. However, in multi-user scenarios, this can lead to prohibitively high training computational costs.

[0004] Furthermore, existing technologies typically select the candidate configuration with the highest evaluation score on the validation set during the candidate model selection process. When the validation scores of multiple candidate configurations are close, this selection strategy is susceptible to validation window noise, leading to overfitting and poor generalization performance in practical applications. Simultaneously, existing technologies often lack a controlled re-orchestration mechanism for model configurations after user behavior shifts, making it difficult to balance configuration stability and prediction timeliness. Regarding the traceability of the model orchestration process, existing technologies generally lack audit records for the entire process of profile extraction, candidate generation, validation evaluation, and configuration selection, hindering online model orchestration operation and maintenance management and troubleshooting. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the following problems in the prior art: the unified model is difficult to adapt to the heterogeneous behavior of multiple users, resulting in insufficient prediction accuracy; the full space hyperparameter search has too high computational overhead in multi-user scenarios; when the verification scores of candidate configurations are close, direct selection is easily affected by verification noise, resulting in unstable selection; there is no controlled re-arrangement mechanism after user behavior drifts; and there is no audit traceability capability in the entire model arrangement process.

[0006] To address the aforementioned technical problems, this invention provides an adaptive prediction model orchestration method based on load behavior profiling, comprising: Obtain historical load time-series data of the target user, construct a time-series window from the historical load time-series data, and obtain a training window, a validation window, and a sliding behavior window; Load behavior features are extracted from the sliding behavior window to obtain a load behavior profile. Based on the load behavior profile results, a candidate training configuration set is generated, and the candidate configurations in the candidate training configuration set are trained and verified based on the training window and the verification window to obtain the verification evaluation value corresponding to each candidate configuration. Based on the verification evaluation value, tolerance screening is performed, and the target prediction model configuration is determined from the candidate training configuration set by combining the candidate configuration complexity and the load behavior profile results. Record the load behavior profile results, the candidate training configuration set, the validation evaluation value, and the target prediction model configuration to obtain the orchestration audit record; Obtain the updated load time-series data of the target user, and perform load behavior drift detection based on the updated load time-series data to obtain the load behavior drift result; Based on the load behavior drift results and the orchestration audit records, re-orchestration control is performed to obtain the current effective prediction model configuration.

[0007] In one embodiment of the present invention, the method for constructing a time-series window from the historical load time-series data to obtain a training window, a validation window, and a sliding behavior window is as follows: The historical load time series data is sorted by timestamp and the data validity is checked to obtain the effective load time series data; The effective load time-series data is segmented according to time sequence to obtain the initial training window and the initial validation window; Perform anti-leakage verification on the initial training window and the initial verification window to determine whether the initial training window and the initial verification window meet the anti-leakage constraints; When the conditions are met, the initial training window is determined as the training window, and the initial verification window is determined as the verification window; If the conditions are not met, at least one of the following is adjusted: the window boundary of the initial training window, the window boundary of the initial verification window, and the isolation interval between the initial training window and the initial verification window. The anti-leakage check is then performed again until a training window and a verification window that satisfy the anti-leakage constraint are obtained. Based on the current orchestration time and the preset behavior window length, the effective load time series data is slidably truncated to obtain the sliding behavior window.

[0008] In one embodiment of the present invention, the method for performing anti-leakage verification on the initial training window and the initial verification window, and determining whether the initial training window and the initial verification window satisfy the anti-leakage constraints, is as follows: Let the time set corresponding to the initial training window be . The time set corresponding to the initial verification window is The sampling time interval is The preset isolation steps are ; The leakage prevention constraint is deemed satisfied when both the initial training window and the initial verification window simultaneously meet the following conditions: , ; Otherwise, it is determined that the leakage prevention constraint is not met.

[0009] In one embodiment of the present invention, the method for extracting load behavior features from the sliding behavior window to obtain a load behavior profile is as follows: The load behavior characteristics of the sliding behavior window are calculated to obtain the load behavior fingerprint; The load behavior fingerprint is subjected to hierarchical profiling to obtain the load behavior profile result. The load behavior fingerprint includes the coefficient of variation, zero-value ratio, daily periodic autocorrelation, tail quantile, interquartile range, and tail ratio.

[0010] In one embodiment of the present invention, the method for performing hierarchical profiling processing on the load behavior fingerprint to obtain the load behavior profiling result is as follows: Based on the load behavior fingerprint, the primary behavior segment corresponding to the target user is determined; Determine whether the first-level behavior segment belongs to the behavior segment that needs to be refined to the second level; When it belongs to the category, determine the corresponding secondary behavior sub-segment of the primary behavior segment; generate the load behavior profile result based on the primary behavior segment and the secondary behavior sub-segment; When it does not belong to the category, the load behavior profile result is generated based on the first-level behavior segment.

[0011] In one embodiment of the present invention, the first-level behavior segment includes high zero-value spike type, weak periodic low zero-value type, large sample low zero-value type, stable periodic low zero-value type, high zero-value periodic type and default safe type; The primary behavior segments that require secondary refinement are the weak-period low-zero value type and the stable-period low-zero value type. When the first-level behavior segment is a weak-period low-zero value type, the first-level behavior segment is further refined based on the daily periodic autocorrelation, coefficient of variation and tail peak characteristics to obtain a second-level behavior sub-segment used to characterize the periodic strength, fluctuation degree and peak tail state. When the first-level behavior segment is a stable periodic low-zero value type, the first-level behavior segment is further refined based on the tail peak characteristics to obtain a second-level behavior sub-segment used to characterize the peak tail state in the stable periodic load.

[0012] In one embodiment of the present invention, the method for training and validating candidate configurations in the candidate training configuration set based on the training window and the verification window to obtain the verification evaluation value corresponding to each candidate configuration is as follows: Training is performed on each candidate configuration in the candidate training configuration set based on the training window; Based on the verification window, each candidate configuration after training is verified, and the verification evaluation value corresponding to each candidate configuration is calculated according to the preset evaluation rules. The training and verification processes for each candidate configuration use the same training window, the same verification window, and the same leakage prevention constraints. During training, the timing of training termination for each candidate configuration is controlled according to preset training stop conditions; When training resources allow, parallel training is performed on each candidate configuration. When training resources are limited, training is performed sequentially according to the priority of the candidate configurations, and basic fallback candidates are retained.

[0013] In one embodiment of the present invention, the method for determining the target prediction model configuration from the candidate training configuration set by performing tolerance screening based on the verification evaluation value and combining the candidate configuration complexity and the load behavior profile results is as follows: A reference evaluation value is determined based on the verification evaluation value of each candidate configuration, and a tolerance candidate set is constructed based on the reference evaluation value and a preset tolerance threshold. The tolerance candidate set includes candidate configurations whose verification evaluation value is not greater than the sum of the reference evaluation value and the preset tolerance threshold. In the tolerance candidate set, the target prediction model configuration is obtained by selecting from the complexity metric of each candidate configuration and the complexity preference corresponding to the load behavior profile result.

[0014] In one embodiment of the present invention, the method for obtaining updated load time-series data of a target user, and performing load behavior drift detection based on the updated load time-series data to obtain load behavior drift results is as follows: The sliding behavior window is updated based on the updated load time series data, and the load behavior fingerprint is re-extracted from the updated sliding behavior window to obtain the current behavior fingerprint; Obtain a reference behavior fingerprint, which is obtained from the orchestration audit record, and the reference behavior fingerprint is a load behavior fingerprint corresponding to the determined target prediction model configuration; The load behavior drift result is determined based on the degree of distribution difference between the current behavior fingerprint and the reference behavior fingerprint.

[0015] In one embodiment of the present invention, the method for re-arrangement control based on the load behavior drift result and the orchestration audit record is as follows: When the load behavior drift result indicates that the drift intensity reaches the preset entry condition and meets the minimum trigger interval constraint, the re-arrangement is triggered, and the steps of candidate training configuration set generation, training verification and target prediction model configuration determination are re-executed, the target prediction model configuration is re-determined, and the re-determined target prediction model configuration is used as the currently effective prediction model configuration. When the load behavior drift result indicates that the drift intensity reaches the preset entry condition but does not meet the minimum trigger interval constraint, re-arrangement is not triggered, and the determined target prediction model configuration is used as the currently effective prediction model configuration. When the load behavior drift result indicates that the drift intensity is lower than or equal to the preset exit condition, the re-arrangement is not triggered, the re-arrangement state is exited, and the determined target prediction model configuration is used as the currently effective prediction model configuration. When the load behavior drift result indicates that the drift intensity is between the preset entry condition and the preset exit condition, re-arrangement is not triggered, and the determined target prediction model configuration is used as the currently effective prediction model configuration. Wherein, the drift intensity threshold corresponding to the preset entry condition is higher than the drift intensity threshold corresponding to the preset exit condition; The minimum trigger interval constraint is used to limit the minimum time interval between two adjacent re-arrangements, and the minimum trigger interval constraint is determined based on the last re-arrangement time in the orchestration audit record.

[0016] Based on the same inventive concept, this invention also provides an adaptive prediction model orchestration system based on load behavior profiling, comprising: The data acquisition and window construction module is used to acquire the historical load time series data of the target user, and construct a time series window on the historical load time series data to obtain a training window, a verification window and a sliding behavior window. The behavior profile extraction module is used to extract load behavior features from the sliding behavior window to obtain load behavior profile results. The candidate configuration generation and verification module is used to generate a set of candidate training configurations based on the load behavior profile results, and to train and verify the candidate configurations in the set of candidate training configurations based on the training window and the verification window to obtain the verification evaluation value corresponding to each candidate configuration. The target configuration determination module is used to perform tolerance screening based on the verification evaluation value, and determine the target prediction model configuration from the candidate training configuration set by combining the candidate configuration complexity and the load behavior profile results. The audit log module is used to record the load behavior profiling results, the candidate training configuration set, the verification evaluation value, and the target prediction model configuration to obtain the orchestration audit log; The drift detection module is used to acquire the updated load time-series data of the target user, perform load behavior drift detection based on the updated load time-series data, and obtain the load behavior drift result. The re-arrangement control module is used to perform re-arrangement control based on the load behavior drift results and the orchestration audit records to obtain the current effective prediction model configuration.

[0017] The technical solution of the present invention has the following advantages compared with the prior art: This invention constructs a training window, validation window, and sliding behavior window through a time-series window model, ensuring that the training and validation processes meet temporal causal constraints, reducing the risk of information leakage in time series prediction, and making the validation evaluation values ​​of candidate configurations more reliable. By extracting load behavior features from the sliding behavior window to obtain load behavior profiles, the model configuration selection can be adapted to the user's current actual electricity consumption status, avoiding the problem of a unified model being unable to adapt to the heterogeneity of multiple users. Based on the load behavior profile results, a candidate training configuration set is generated, converging the global hyperparameter search space into a small number of profile-driven candidate configurations, significantly reducing... The invention addresses several key challenges in multi-user load forecasting. It employs a multi-dimensional approach, including training computation budget, tolerance-based selection based on validation evaluation values, and determination of the target prediction model configuration by combining candidate configuration complexity and load behavior profiling results. This avoids overfitting due to validation noise and improves the generalization stability of the deployed configuration. Audit records are maintained for profiling results, candidate configuration sets, validation evaluation values, and target configurations, ensuring traceability of the orchestration process. Drift detection is performed based on updated load time-series data, and re-orchestration control is implemented after drift detection, incorporating audit records to enable adaptive updates to the model configuration based on changes in user electricity consumption behavior, guaranteeing predictive timeliness over long-term operation. This invention effectively solves the problems of poor uniform model adaptability, high search computation volume, unstable selection, lack of re-orchestration mechanisms, and audit traceability in existing technologies, providing an efficient, stable, and auditable adaptive configuration orchestration scheme for multi-user load forecasting. Attached Figure Description

[0018] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0019] Figure 1 This is a flowchart illustrating the adaptive prediction model orchestration method based on load behavior profiling in an embodiment of the present invention. Figure 2 This is a schematic diagram of the process for re-arrangement control based on load behavior drift results and orchestration audit records in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0021] Example 1: like Figure 1 As shown, this invention provides an adaptive prediction model orchestration method based on load behavior profiling, comprising: Obtain historical load time-series data of the target user, construct a time-series window from the historical load time-series data, and obtain a training window, a validation window, and a sliding behavior window; Load behavior features are extracted from the sliding behavior window to obtain a load behavior profile. Based on the load behavior profile results, a candidate training configuration set is generated, and the candidate configurations in the candidate training configuration set are trained and verified based on the training window and the verification window to obtain the verification evaluation value corresponding to each candidate configuration. Based on the verification evaluation value, tolerance screening is performed, and the target prediction model configuration is determined from the candidate training configuration set by combining the candidate configuration complexity and the load behavior profile results. Record the load behavior profile results, the candidate training configuration set, the validation evaluation value, and the target prediction model configuration to obtain the orchestration audit record; Obtain the updated load time-series data of the target user, and perform load behavior drift detection based on the updated load time-series data to obtain the load behavior drift result; Based on the load behavior drift results and the orchestration audit records, re-orchestration control is performed to obtain the current effective prediction model configuration.

[0022] This invention constructs a time-series window from the historical load time-series data of the target user to obtain a training window, a verification window, and a sliding behavior window reflecting the current electricity consumption state, satisfying time-series causal constraints, thus ensuring the fairness and reliability of the training and verification process. Load behavior features are extracted from the sliding behavior window to obtain a load behavior profile, enabling the generation of candidate configurations to accurately adapt to the user's current actual electricity consumption pattern. A candidate training configuration set is generated based on the load behavior profile results, and the training and verification windows are used to train and verify each candidate configuration to obtain verification evaluation values, converging the global search space into a small number of profile-driven candidate configurations. This invention significantly reduces the training computation budget in multi-user scenarios; it uses tolerance screening based on validation evaluation values ​​and combines candidate configuration complexity and load behavior profiling results to determine the target prediction model configuration, avoiding overfitting caused by validation window noise chasing small score differences and improving the generalization stability of the deployment configuration; it audits and records the profiling results, candidate configuration set, validation evaluation values, and target configuration to ensure complete traceability of the orchestration process; it detects behavior drift by acquiring updated load time-series data and performs re-orchestration control based on drift results and audit records, enabling the model configuration to adaptively update with changes in user electricity consumption behavior. This invention effectively solves the problems of poor uniform model adaptability, large computational load of full-space search, unstable selection, lack of re-orchestration mechanism, and audit traceability in existing technologies, providing an efficient, stable, and auditable adaptive configuration orchestration scheme for multi-user load forecasting.

[0023] In this embodiment of the invention, after obtaining the historical load time series data of the target user, the historical load time series data is sorted according to the timestamp so that each load record is arranged in chronological order.

[0024] Historical load time-series data can be represented as a sequence of load values ​​for target users at different time indices, where Indicates the target user in the time index The historical load values ​​are used. The sorted historical load time-series data undergoes a data validity check, which includes checking for duplicate timestamps, out-of-order timestamps, invalid load values, and prohibited future information fields. Missing or incomplete time points are identified and processed according to preset data validity rules, thus obtaining effective load time-series data. This effective load time-series data is used to construct subsequent training windows, validation windows, and sliding behavior windows, ensuring that subsequent candidate configuration training and validation, as well as load behavior profile extraction, are based on data that has undergone time-series organization and validity checks.

[0025] A training window, a validation window, and a sliding behavior window are constructed based on effective load time-series data. The training window is used to train the model on candidate configurations from the candidate training configuration set. The validation window is used to validate and evaluate the trained candidate configurations and participate in the selection of the target prediction model configuration. The sliding behavior window is used to extract load behavior characteristics of the target user near the current scheduling time to reflect the target user's current electricity consumption behavior status. The training window, validation window, and sliding behavior window are distinct in their purpose: the training window and validation window are used in the candidate configuration training and validation process, while the sliding behavior window is used in the load behavior profile generation process.

[0026] The payload time-series data is segmented chronologically to obtain an initial training window and an initial validation window. Purged and gap leak prevention checks are then performed on the initial training window and the initial validation window. Let the time set corresponding to the initial training window be... The time set corresponding to the initial verification window is The sampling time interval is The preset isolation steps are When the initial training window and the initial validation window simultaneously satisfy... as well as When the initial training window and the initial verification window meet the leakage prevention constraints, it is determined that they do not meet the leakage prevention constraints.

[0027] in, This is used to limit the maximum timestamp of the training window to be earlier than the minimum timestamp of the validation window, so that the training window and the validation window are distributed sequentially along the timeline. This is used to ensure that the time interval between the training window and the validation window is no less than the time length corresponding to a preset number of isolation steps, thus forming a purged interval or gap isolation region between the training window and the validation window. The purged interval or gap isolation region is used to reduce the risk of temporal overlap and nearest neighbor leakage between training samples and validation samples in rolling features, target prediction windows, label generation, or post-processing calibration, so that candidate training configurations are trained and validated under the same temporal causal constraints, thereby improving the comparability between validation evaluation values ​​of different candidate configurations.

[0028] When the initial training window and the initial verification window do not meet the anti-leakage constraints, at least one of the following is adjusted: the window boundary of the initial training window, the window boundary of the initial verification window, and the isolation interval between the initial training window and the initial verification window. This adjustment may include changing the end boundary of the training window, changing the start boundary of the verification window, or adjusting the length of the isolation interval between the training window and the verification window. After adjustment, the adjusted initial training window and initial verification window are re-verified according to the anti-leakage constraints until training windows and verification windows that meet the anti-leakage constraints are obtained. The training window that meets the anti-leakage constraints serves as the data window used for subsequent candidate configuration training, and the verification window that meets the anti-leakage constraints serves as the data window used for subsequent candidate configuration verification and evaluation.

[0029] After obtaining the training window and validation window, the payload time-series data is truncated using a sliding action window based on the current orchestration time and the preset action window length. Let the time index corresponding to the current orchestration time be denoted as . The preset behavior window length is Then the sliding behavior window can be represented as: , in, Indicates the expiration time index The sliding behavior window, Indicates the length of the behavior window. This represents the load values ​​arranged in chronological order within the sliding behavior window.

[0030] The sliding behavior window terminates no later than the current orchestration time, and only includes historical load data visible before the current orchestration time, excluding future target values ​​and test data after the validation window. Through these restrictions, the load behavior features extracted based on the sliding behavior window do not introduce future target information, maintaining consistency with the temporal causal constraints followed by the training and validation windows. Because the sliding behavior window updates continuously with the current orchestration time, the same target user can correspond to different sliding behavior windows at different times, allowing subsequent load behavior profiles to reflect the target user's current electricity consumption behavior, rather than relying solely on long-term static labels. Therefore, while ensuring leakage prevention constraints, the training window, validation window, and sliding behavior window respectively serve the functions of candidate configuration training, candidate configuration validation and evaluation, and load behavior profile extraction.

[0031] Furthermore, load behavior fingerprints are calculated within the sliding behavior window.

[0032] Let the sliding behavior window be The load value within the window is First calculate the window mean. and window standard deviation : , , in, Representation of behavior window Average load within, Representation of behavior window The load standard deviation within a window is used to characterize the degree of dispersion of the load value relative to the mean within the window.

[0033] Calculate the coefficient of variation based on the window mean and window standard deviation. : , in, It is a very small positive number, used to prevent the denominator from being zero. Coefficient of variation The coefficient of variation is used to characterize the overall load volatility and reflects the stability of the target user's electricity consumption behavior. The larger the coefficient of variation, the more drastic the load fluctuation and the more unstable the user's electricity consumption behavior; the smaller the coefficient of variation, the smoother the load fluctuation and the more stable the user's electricity consumption behavior.

[0034] Based on near-zero load threshold Calculate the zero-value ratio : , in, As an indicator function, when the sample Not higher than the near-zero load threshold The value is 1 if true, and 0 otherwise. Zero-value ratio. The zero-value ratio is used to characterize the proportion of low-load or near-zero-load samples within the sliding behavior window, reflecting the percentage of time that the target user's electrical equipment is in standby or off state for extended periods. A higher zero-value ratio indicates a large number of low-load periods in the user's electricity consumption behavior, and a greater number of zero-value or near-zero-value samples in the load sequence. This has a direct impact on the selection of the loss function and post-processing strategy for the prediction model.

[0035] For load time series data with a fixed sampling granularity, the number of sampling steps within a natural day is used as the basis. Calculate daily periodic autocorrelation : , in, The number of sampling steps within a calendar day is represented by the sampling time interval. according to Sure, This represents the Pearson correlation coefficient. Daily periodic autocorrelation. This is used to characterize the periodic correlation between the current window load and the load lagging one day, reflecting whether the target user's electricity consumption behavior has a stable daily periodic pattern. The higher the daily periodic autocorrelation, the more regular the user's daily electricity consumption pattern, and the more the prediction model can rely on periodic features; the lower the daily periodic autocorrelation, the greater the variation in the user's daily electricity consumption pattern, and the more the prediction model needs to rely on other features.

[0036] Calculate the 90th percentile of the window load 95th percentile and the 99th percentile As a tail quartile array, it is used to characterize the intensity of high load tails and peaks. The 90th percentile... Used to characterize higher load levels, 95th percentile Used to characterize the high-load tail section, 99th percentile Used to characterize extreme peak load levels. The combination of the above three quantiles can comprehensively depict the high-tail characteristics of the target user load distribution, especially providing a quantitative description of the intensity and frequency of load peaks.

[0037] Calculate interquartile range : , in, The third quartile of the window load, The first quartile of the window load. Interquartile range. Used to characterize robust diffusion, as an effective complement to standard deviation, it is not overly affected by extreme peak values ​​and can robustly characterize the intermediate dispersion of load data.

[0038] Calculate the tail ratio : , Tail ratio Used to characterize the exposure level of the tail end relative to the extreme spike. When A value significantly greater than 1 indicates the presence of a clear extreme peak load within the window, meaning the peak load level corresponding to the 99th percentile is much higher than the high load tail level corresponding to the 95th percentile, indicating a prominent peak event in user electricity consumption behavior; when When the value is close to 1, it indicates that the high-end tail of the load distribution is relatively smooth and the extreme peaks are not prominent.

[0039] This results in the formation of a coefficient of variation. Zero value ratio Daily cycle autocorrelation Tail digits , , Interquartile range and tail ratio Load behavior fingerprint : , The above load behavior fingerprint encompasses load volatility (coefficient of variation). Near-zero load ratio (zero value ratio) ), daily cycle patterns (daily cycle autocorrelation) Tail peak intensity (tail fraction) , , Robust diffusion (interquartile range) ) and extreme tail exposure (tail ratio) The six dimensions of load behavior characteristics can depict the electricity consumption behavior of the target user within the current sliding behavior window from multiple perspectives.

[0040] After obtaining the load behavior fingerprint, hierarchical profiling is performed on the load behavior fingerprint. First, the first-level behavior segment corresponding to the target user is determined according to the ordered threshold rule. The first-level behavior segments include high zero-value spike type, weak periodic low zero-value type, large sample low zero-value type, stable periodic low zero-value type, high zero-value periodic type, and default safe type. Each primary behavior segment is differentiated based on key indicators in the load behavior fingerprint. For example, the high zero-value spike type corresponds to users with a high zero-value ratio and a large coefficient of variation, indicating that their electricity consumption behavior has a large number of low-load periods accompanied by obvious spikes; the weak-period low zero-value type corresponds to users with weak daily periodic autocorrelation but a low zero-value ratio, indicating that their electricity consumption behavior lacks a stable daily periodic pattern and does not have a large number of zero values; the stable-period low zero-value type corresponds to users with strong daily periodic autocorrelation and a low zero-value ratio, indicating that their electricity consumption behavior has a stable daily periodic pattern; the large-sample low zero-value type corresponds to users with sufficient sample size and a low zero-value ratio, indicating that they have sufficient available training data and do not have a large number of zero values; the high zero-value periodic type corresponds to users with a high zero-value ratio and strong daily periodic autocorrelation, indicating that their electricity consumption behavior is mainly during low-load periods but still retains a certain periodic pattern; the default safety type serves as a fallback type when none of the above types match. The primary behavior segments are used to determine the training budget, feature complexity, and post-processing risk level of the model candidate configuration in a broad sense.

[0041] Determine whether the identified primary behavior segment belongs to the category requiring secondary refinement. Primary behavior segments requiring secondary refinement are the weak-cycle low-zero value type and the stable-cycle low-zero value type. These two types of behavior segments still exhibit significant internal differences, and secondary refinement can more accurately guide the generation of candidate configurations. When the identified primary behavior segment does not belong to the category requiring secondary refinement, the load behavior profile result is directly generated based on that primary behavior segment.

[0042] When a identified primary behavior segment requires secondary refinement, the corresponding secondary behavior sub-segments are further determined. When the primary behavior segment is a weak-cycle low-zero value type, it is further refined based on daily autocorrelation, coefficient of variation, and tail peak characteristics to obtain secondary behavior sub-segments characterizing cycle strength, volatility, and peak tail state. Specifically, within the weak-cycle low-zero value type, daily autocorrelation may exhibit different intensities, the coefficient of variation may reflect different degrees of volatility, and the tail ratio may indicate the presence of peak tails. These differences further classify users into different subtypes to adapt to different candidate configuration priorities. When the primary behavior segment is a stable-cycle low-zero value type, it is further refined based on tail peak characteristics to obtain secondary behavior sub-segments characterizing the peak tail state in stable-cycle loads. Specifically, within the stable cycle low zero-value type, although the daily cycle pattern is stable and the zero-value ratio is low, the tail ratio may indicate whether there are sharp tail events. Users with sharp tails and those without sharp tails need to be treated differently in the post-processing strategy of the prediction model.

[0043] Load behavior profiles are generated based on primary behavior segments and secondary behavior sub-segments. These profiles include primary behavior segment labels and secondary behavior sub-segment labels, which are used to generate subsequent candidate training configuration sets.

[0044] By performing hierarchical profiling of load behavior fingerprints into first-level behavior segments and second-level behavior sub-segments, continuous load behavior fingerprint indicators can be mapped into discrete profile labels with clear meanings in the power load domain. This makes the profiling results both interpretable and directly usable as input for policy routing. Compared to directly using general clustering results as the routing basis for model configuration, the hierarchical profiling method of this invention aligns profile labels with the domain knowledge of power load prediction, avoiding the problem of a lack of effective correlation between uninterpretable clustering results and prediction model configuration. Simultaneously, the hierarchical segmentation method can control the overall direction of candidate configurations at the first-level behavior segment level and refine the priority of candidate configurations at the second-level behavior sub-segment level, making the generation of candidate configurations more accurate, further narrowing the search range of candidate configurations, and reducing the training computation budget in multi-user scenarios.

[0045] Furthermore, the load behavior profiling results are input into the policy router, which generates a candidate training configuration set based on the first-level behavior segments and second-level behavior sub-segments. The candidate training configuration set is not obtained by indiscriminately searching the entire hyperparameter space, but rather by selecting a small number of candidate configurations from a pre-defined candidate configuration library that match the load behavior profiling results. Each candidate configuration is encoded by multiple interpretable configuration axes, including model width, loss function, feature set, post-processing strategy, and sampling strategy. Model width characterizes the capacity differences among candidate models; different widths correspond to different numbers of model parameters and expressive power. The loss function characterizes the differences in training objectives; different loss functions have different adaptability to different types of data distributions (such as load data with a large number of zero values ​​or spikes). The feature set characterizes the differences in input feature complexity; different feature sets contain different numbers and types of input features. The post-processing strategy characterizes whether prediction protection or spike protection is enabled, such as pruning outliers in the prediction results or specifically correcting peak periods. The sampling strategy characterizes whether balanced sampling or ordinary sampling is used, such as sampling zero-value periods and peak periods with different weights to improve model training performance. The combination of each configuration axis is pre-designed based on domain knowledge and historical experience, ensuring that the candidate configurations corresponding to different image segments have clear sources of difference and interpretability. The number of candidate configurations in the candidate training configuration set is usually small to reduce the training computation budget, while retaining basic backoff candidates to ensure that the orchestration process is not completely interrupted due to candidate training failures.

[0046] Different first-level and second-level behavior segments correspond to different candidate configuration priorities. For high-zero-value spike-type users, the policy router can prioritize generating candidate configurations that include higher model capacity, richer features, and spike protection post-processing. This is because high-zero-value spike-type users are characterized by a large number of low-load periods and prominent spike events, requiring the model to have strong nonlinear fitting capabilities, while spike protection post-processing is needed to prevent the predicted values ​​from being excessively affected by spikes. For weak-periodic low-zero-value users, the policy router adjusts the priority of candidate configurations according to the specific type of the second-level behavior segment. If the second-level segment represents a very weak period and large fluctuations, candidate configurations containing rich features can be prioritized to compensate for insufficient periodic information; if the second-level segment represents small fluctuations, lightweight configurations can be prioritized to reduce training costs. For stable-periodic low-zero-value users, the policy router adjusts the priority of candidate configurations according to whether there is a spike tail in the second-level behavior segment. If a spike tail exists, candidate configurations with post-processing protection can be prioritized; if there is no spike tail, lightweight basic configurations can be used first. For a large number of users with low zero-value patterns, the policy router can prioritize candidate configurations with richer feature sets because there is sufficient available training data to support more complex feature inputs. For users with high zero-value periods, the policy router can prioritize loss functions and sampling strategies with specialized mechanisms for handling zero values ​​to improve the model's fit over a large number of zero-value periods. For default-safe users, the policy router can output a preset default candidate configuration as a basic backoff candidate. In this way, the candidate training configuration set is limited to a range of candidates related to the load behavior profile, avoiding indiscriminate searching of the entire hyperparameter space, thereby significantly reducing the training computation budget in multi-user scenarios.

[0047] After obtaining the candidate training configuration set, each candidate configuration in the set is trained based on a training window, and each trained candidate configuration is validated based on a validation window. The validation evaluation value for each candidate configuration is calculated according to a preset evaluation rule. To ensure comparability of validation evaluation values ​​between different candidate configurations, the training and validation processes for each candidate configuration use the same training window, validation window, and anti-leakage constraints, along with a unified evaluation metric and a unified training termination condition. The training window provides training samples for each candidate configuration, the validation window provides validation samples, the anti-leakage constraints ensure that the training samples and validation samples meet the temporal causality requirement, the evaluation metric is used to uniformly measure the predictive performance of each candidate configuration, and the training termination condition controls the timing of training termination for each candidate configuration. The evaluation metric may include at least one of mean absolute error, root mean square error, or mean absolute percentage error, used to quantify the degree of prediction deviation of the candidate configuration on the validation window. The training termination condition can be determined based on the changes in the validation evaluation value in consecutive training rounds; for example, training is terminated when the validation evaluation value no longer decreases significantly in several consecutive rounds to avoid ineffective consumption of training resources.

[0048] When training resources permit, parallel training is performed on each candidate configuration in the candidate training configuration set, utilizing multiple processes or multiple computing devices to train multiple candidate configurations simultaneously, thereby shortening the overall training time. When training resources are limited, training is performed sequentially according to the priority of the candidate configurations. The priority of the candidate configurations is determined based on the load behavior profile results, the preset candidate order, or historical audit records. For example, when the load behavior profile result is a high zero-value spike, candidate configurations that include spike protection post-processing are trained first; when historical audit records exist, candidate configurations that have historically performed stably under the same profile segment are trained first. During sequential training, basic fallback candidates are retained to avoid complete interruption of the orchestration process when some candidate training fails. The basic fallback candidates are preset default candidate configurations in the candidate training configuration set. These default candidate configurations can be trained under any load behavior profile result, ensuring that the orchestration process has the minimum usable output. Through the above shared training and verification protocol, the verification evaluation values ​​of each candidate configuration are generated under the same data partitioning and evaluation rules, ensuring fair comparability and avoiding comparison bias caused by different candidate configurations using different data partitions. By combining parallel training and priority-based training in resource scheduling, efficiency is improved when training resources are sufficient, and the orchestration process can still be completed when resources are limited.

[0049] Furthermore, tolerance screening is performed based on the validation evaluation values. The predictive performance of each candidate configuration on the validation window is quantified by the validation evaluation values, which are loss-type evaluation values. The smaller the value, the smaller the prediction bias and the better the generalization performance of the candidate configuration on the validation window. Let the set of candidate training configurations be... Candidate configurations The corresponding verification loss is Determine the minimum verification loss based on the verification loss of each candidate configuration. : , in, Indicates candidate configuration The verification loss or overall evaluation value on the verification window.

[0050] Based on the minimum verification loss and preset tolerance threshold Constructing a tolerance candidate set : , in, This represents the tolerance selection threshold, used to determine whether the difference in validation loss between candidate configurations is significant. This represents the tolerance candidate set consisting of candidate configurations whose verification loss is no greater than the minimum verification loss plus the tolerance threshold.

[0051] When the verification loss of a candidate configuration is significantly better than that of all other candidate configurations, i.e., only one candidate configuration's verification loss falls into the range... Within the specified range, if the tolerance candidate set only contains this candidate configuration, then this candidate configuration is directly determined as the target prediction model configuration. When the validation loss of multiple candidate configurations falls within the specified range... Within a certain range, i.e., when the difference in validation loss among multiple candidate configurations is not significant, the tolerance candidate set contains multiple candidate configurations. In this case, further selection is needed by combining the complexity of the candidate configurations and the results of the load behavior profile. The above tolerance screening mechanism avoids the overfitting problem caused by chasing small score differences due to validation window noise or random fluctuations, making the determination of the target prediction model configuration more robust when the validation scores are close.

[0052] For cases where the tolerance candidate set contains multiple candidate configurations, the target prediction model configuration is further determined based on the complexity metric of each candidate configuration and the complexity preference corresponding to the load behavior profiling results. The complexity metric for the candidate configuration is segmented conditional complexity. This is expressed as a weighted sum of the complexities of multiple configuration axes: , in, This indicates a first-level line segment. This indicates a subfield representing a second-level behavior; , , , , These represent the complexity or risk level corresponding to the model width, loss function, feature set, post-processing strategy, and sampling strategy, respectively. , , , , These are corresponding weights used to adjust the contribution of each configuration axis's complexity to the overall complexity. The complexity level of the model width. The model capacity reflects the candidate configuration; the larger the capacity, the higher the complexity. The complexity level or risk level of the loss function. This reflects the sensitivity of different loss functions to data distribution; loss functions that require special handling for zero-valued or peaked distributions have different risk levels; the complexity level of the feature set. This reflects the quantity and type of input features; the richer the features, the higher the complexity. It also indicates the complexity or risk level of the post-processing strategy. This reflects whether predictive protection or spike protection is enabled, and the strength of the protection; the complexity level or risk level of the sampling strategy. This reflects whether balanced sampling is used and its complexity. The complexity or risk level of each configuration axis is pre-set based on domain knowledge and historical experience. The same configuration axis can have different complexity or risk levels under different first-level behavior segments and second-level behavior sub-segments. For example, the risk level of the spike protection post-processing strategy is lower under high-zero-value spike type, while the risk level of the spike protection post-processing strategy is higher under stable-cycle low-zero-value type spike-free tail sub-segment. Segment conditional complexity. This reflects the overall complexity of the candidate configurations under the current image segment.

[0053] Within the tolerance candidate set, selection is made based on the complexity preference corresponding to the current load behavior profile. Different first-level and second-level behavior sub-segments correspond to different complexity preferences. For example, for high-zero-value spike-type loads, candidate configurations with spike protection post-processing strategies are preferred, even if the model width and complexity of such a candidate configuration are large, as long as its validation loss falls within the tolerance set. For stable-cycle low-zero-value loads without spike tails, lightweight basic configurations are preferred, even if there are candidate configurations with higher complexity but slightly lower validation loss, the lower-complexity candidate configuration is preferred. For weak-cycle low-zero-value loads with extremely weak cycles and large fluctuations, candidate configurations with richer feature sets are preferred to compensate for insufficient cycle information. For candidate configurations with equal complexity, the prior priority under the current behavior segment is used. After performing a stable ranking, the target prediction model configuration is finally selected. : , in, This indicates the final selected target prediction model configuration; This indicates the prior priority under the current behavior segment, used for stable ranking when segment conditional complexity remains equal. For example, candidate configurations that perform more stably under the same profile segment in historical audit records receive higher prior priority. In this way, when the validation loss differences among multiple candidate configurations are not significant, instead of directly selecting the candidate configuration with the absolutely optimal validation score from the tolerance candidate set, it is based on segment conditional complexity. Priority of behavior segments By selecting more robust candidate configurations that better align with current profile segment preferences, the risk of overfitting due to validation window noise is reduced, thereby improving the generalization stability of the deployment configuration in actual predictions.

[0054] Furthermore, the load behavior profiling results, candidate training configuration sets, verification evaluation values ​​corresponding to each candidate configuration, and target prediction model configuration are recorded to obtain orchestration audit records. The orchestration audit record is a structured information set generated during each orchestration process, used to record the complete decision-making chain from load behavior profiling extraction to target prediction model configuration determination. The orchestration audit record preserves the original data status, intermediate processing results, and final selection conclusions upon which this orchestration was based, making the orchestration process reproducible, traceable, and auditable afterward. The orchestration audit record provides a complete description of the currently effective prediction model configuration and its background, and also serves as the data foundation for subsequent drift detection and re-orchestration control.

[0055] The orchestration audit log includes the target user identifier, data time range, sliding behavior window length, sampling granularity, load behavior fingerprint, first-level behavior segment, second-level behavior sub-segment, candidate training configuration set, verification evaluation value corresponding to each candidate configuration, priority and complexity corresponding to each candidate configuration, the finally selected target prediction model configuration, and the currently effective prediction model configuration. The orchestration audit log also includes the current orchestration time and the last re-orchestration time, used for determining the minimum trigger interval constraint in re-orchestration control.

[0056] When the orchestration process is initiated due to behavioral drift, difficult user triggering, candidate expansion triggering, or re-orchestration triggering, the orchestration audit log further records the corresponding triggering reason, including the triggering type and triggering time. Behavioral drift triggering refers to initiating re-orchestration due to the detection of a significant drift in user load behavior; difficult user triggering refers to initiating candidate expansion or re-orchestration because a target user has been marked as a difficult user; candidate expansion triggering refers to initiating the expansion process because the size of the candidate configuration set needs to be expanded; and re-orchestration triggering refers to initiating re-orchestration because the re-orchestration conditions are met. By recording the above triggering reasons, the background and decision-making basis for each orchestration process can be traced.

[0057] The orchestration audit log also saves the load behavior fingerprint corresponding to the determined target prediction model configuration as a reference behavior fingerprint. The reference behavior fingerprint is used in subsequent drift detection to compare the distribution difference with the current behavior fingerprint, quantifying the degree of change in user electricity consumption behavior relative to the reference state. The orchestration audit log also saves the last re-orchestration time, used to determine whether the time interval between the current time and the last re-orchestration time meets the minimum trigger interval constraint, preventing multiple re-orchestrations triggered by continuous drift detection within a short period.

[0058] By orchestrating audit logs, it is possible to trace the complete basis for assigning a specific predictive model configuration to a target user at a specific point in time. This includes the load behavior status at that time, the profile segmentation results, the candidate configuration list, the verification performance of each candidate configuration, the selected configuration and its complexity information, and the reason for triggering re-orchestration. Versioned management of orchestration audit logs allows for the retrospective analysis of each orchestration result. When abnormal online prediction performance occurs, the problem can be quickly located by reviewing historical audit logs, improving the maintainability and interpretability of the model orchestration process.

[0059] Furthermore, after the target prediction model is configured and put into use, the updated load time-series data of the target users is periodically acquired or according to preset trigger conditions, and the sliding behavior window is updated based on the updated load time-series data. The method for updating the sliding behavior window is as follows: taking the current time as the new end time, continuous historical load data is truncated backward according to the preset behavior window length to form the updated sliding behavior window. On the updated sliding behavior window, the coefficient of variation, zero-value ratio, daily periodic autocorrelation, tail quantiles, interquartile range, and tail ratio are recalculated according to the same load behavior fingerprint extraction method described above to obtain the current behavior fingerprint. The reference behavior fingerprint is obtained from the orchestration audit record. The reference behavior fingerprint is the load behavior fingerprint corresponding to the determined target prediction model configuration, that is, the behavior fingerprint on which the load behavior profile result was formed in the previous orchestration process. It is stored in the orchestration audit record for comparison benchmarks for subsequent drift detection. The load behavior drift result is determined according to the degree of distribution difference between the current behavior fingerprint and the reference behavior fingerprint.

[0060] In a preferred embodiment, the degree of distributional variability is determined by a population stability index. , The population stability index is described by metrics such as distance and residual tail variation. Let the population stability index between the current behavioral fingerprint and the reference behavioral fingerprint be... , Distance is The tail of the residual changes as follows The corresponding normalized reference value is , , The above indicators are normalized to obtain the normalized drift intensity. : , in, Used to measure the degree of deviation between the current behavioral fingerprint and the reference behavioral fingerprint across the entire distribution. Used to measure the maximum difference between the two in their cumulative distribution. Used to measure the degree of change in the 90th percentile of the residual distribution, reflecting the drift in tail behavior. Normalized drift strength. The larger the value, the greater the deviation of the target user's current behavioral state from the reference behavioral state. When When the preset entry conditions are exceeded, it indicates that the target user's electricity consumption behavior has changed significantly, requiring a reconfiguration and orchestration of the model; when When the threshold for exiting the program is lower than the preset exit condition, it indicates that the target user's electricity consumption behavior has stabilized and the program can exit the reprogramming state. This drift detection method quantifies the degree of change in user electricity consumption behavior relative to the reference state, providing an objective quantitative basis for subsequent reprogramming control.

[0061] Furthermore, such as Figure 2 As shown, the normalized drift intensity is obtained. Subsequently, re-arrangement control is performed based on load behavior drift results and orchestration audit records.

[0062] The preset entry condition corresponds to the drift intensity threshold as follows: The drift intensity threshold corresponding to the preset exit condition is ,in This creates a hysteresis interval. The last re-arrangement time is retrieved from the orchestration audit log. The minimum trigger interval constraint is then determined based on the time interval between the current time and the last re-arrangement time. The minimum trigger interval constraint limits the minimum time interval between two adjacent re-arrangements, preventing multiple re-arrangements from being triggered within a short period due to continuous drift detection.

[0063] when When the minimum trigger interval constraint is met, a re-arrangement is triggered, and the steps of generating the candidate training configuration set, training verification, and determining the target prediction model configuration are re-executed. The target prediction model configuration is then re-determined, and the re-determined target prediction model configuration is used as the currently effective prediction model configuration.

[0064] when However, if the minimum trigger interval constraint is not met, re-arrangement will not be triggered, and the determined target prediction model configuration will be used as the currently effective prediction model configuration.

[0065] when If this happens, the re-orchestration will not be triggered, the re-orchestration state will be exited, and the already determined target prediction model configuration will be used as the currently effective prediction model configuration.

[0066] when When the drift intensity is within the hysteresis range, no re-arrangement is triggered. The current arrangement state is maintained, and the determined target prediction model configuration is used as the currently effective prediction model configuration.

[0067] By setting the entry threshold Exceeding the exit threshold The system defines a hysteresis range. When the drift intensity exceeds the entry threshold, a re-arrangement is triggered. When the drift intensity falls below the exit threshold, the re-arrangement state is exited. When the drift intensity is between these two thresholds, the current state remains unchanged. This avoids frequent re-arrangement triggers or state transitions when the drift intensity fluctuates around a single threshold. A minimum trigger interval constraint further limits the minimum time interval between two adjacent re-arrangements, preventing multiple re-arrangements from being triggered consecutively in a short period due to short-term fluctuations or transient changes, thus achieving a balance between the timeliness of configuration updates and configuration stability. After a re-arrangement is triggered, the previous re-arrangement time in the orchestration audit record is updated, and the triggering reason for this re-arrangement is recorded, completing the version update of the audit record.

[0068] In one alternative implementation, difficult users are identified based on orchestration audit records. When a target user's historical prediction error is consistently higher than the average of similar user profiles, it indicates that the candidate configurations in the current user profile consistently perform poorly in predicting the user's performance, failing to meet accuracy requirements. When a target user experiences a high number of candidate training failures, it indicates that the user's load data characteristics may cause some candidate configurations to fail to converge or produce abnormal results during training. When a target user's behavior frequently drifts, it indicates that the user's electricity consumption patterns are constantly changing, making it difficult for static configurations to adapt in the long term. When at least one of the above conditions is met, the target user is marked as a difficult user.

[0069] For users with high difficulty in predicting behavior, the upper limit of candidate configurations is increased during the next generation of candidate training configuration sets. Based on the candidate profile routing, candidate configurations corresponding to adjacent profile segments or those with historically stable performance are introduced. Candidate configurations corresponding to adjacent profile segments refer to those corresponding to other profile segments that are adjacent to the current first-level behavior segment or second-level behavior sub-segment in terms of behavioral characteristics. For example, for weakly periodic low-zero-value users, some candidate configurations with stable periodic low-zero-value characteristics can be introduced to expand the coverage of users with large fluctuations. Candidate configurations with historically stable performance refer to those selected from orchestration audit records that have demonstrated good performance on the target user's historical data. Candidate expansion is still performed under the constraints of the load behavior profile results. The generation of candidate configurations is still based on the current profile segment and sub-segment, and does not constitute an unbounded search of the entire hyperparameter space. Through the above expansion method, a wider candidate coverage is provided for users with high difficulty in predicting behavior, while avoiding a significant increase in training computation budget due to unconstrained full-space search.

[0070] The order of candidate training is reordered using orchestration audit logs. Historical performance information for each candidate configuration under the target user or similar user demographics is obtained from the orchestration audit logs. This historical performance information includes the validation evaluation value of the candidate configuration in the historical validation window, whether training was successful, and the convergence speed during training. When a candidate configuration performs stably for a long period under the target user or similar user demographics, its priority is increased in subsequent training, ensuring it receives priority training opportunities when resources are limited. When a candidate configuration fails to train multiple times under the same user demographics or exhibits unstable generalization performance, its priority is decreased, allowing it to be trained later when resources are limited. Through this historical prior reordering mechanism, training resources are prioritized for candidate configurations more likely to produce good results within a limited training budget, improving the hit efficiency of the candidate training configuration set and reducing resource waste caused by ineffective candidate training.

[0071] In an exemplary residential load forecasting scenario, the target user's load data sampling granularity is 15 minutes, containing 96 sampling points within a calendar day. After obtaining the target user's historical load time-series data for the most recent several months, the timestamp sorting, data validity check, training and verification window construction, leakage prevention verification, and sliding behavior window extraction are completed according to the aforementioned method. This scenario is only used to illustrate the implementation process of the present invention and does not constitute a limitation on the sampling granularity, window length, user type, or number of candidate configurations.

[0072] In this scenario, if a target user exhibits a high zero-value ratio, a large coefficient of variation, and a high tail ratio within the sliding behavior window, their load behavior profile can be identified as a high-zero-value spike type. Based on this load behavior profile, the policy router generates a set of candidate training configurations that match the high-zero-value spike type. These candidate configurations can include combinations of configurations such as higher model capacity, richer feature sets, and spike protection post-processing strategies, while retaining basic fallback candidate configurations. Each candidate configuration is trained and validated according to the aforementioned unified training and validation protocol, under the same training window, the same validation window, and the same leakage prevention constraints, yielding a validation evaluation value for each candidate configuration.

[0073] After obtaining the validation evaluation values ​​for each candidate configuration, if the validation losses of multiple candidate configurations all fall into the tolerance candidate set, the candidate configuration with the lowest validation loss is not directly selected. Instead, the target prediction model configuration is determined by combining the segment-conditional complexity of each candidate configuration with the complexity preference corresponding to the high-zero-value peak profile. Therefore, when the validation loss difference is not significant, the selection instability caused by validation window noise can be reduced, ensuring that the final determined target prediction model configuration matches the current load behavior profile results.

[0074] During subsequent operation, if the target user experiences a significant increase in peak load during the evening rush hour due to the addition of high-power electrical equipment, resulting in an increase in the tail fraction, tail ratio, or residual tail change in the updated sliding behavior window, the system obtains the load behavior drift result based on the degree of distribution difference between the current behavior fingerprint and the reference behavior fingerprint. When the normalized drift intensity reaches the preset entry condition, and the time interval between the current moment and the last re-arrangement moment meets the minimum trigger interval constraint, the system triggers re-arrangement, re-executes the steps of candidate training configuration set generation, training verification, and target prediction model configuration determination, and uses the newly determined target prediction model configuration as the currently effective prediction model configuration. Simultaneously, the system writes the drift trigger reason, re-arrangement moment, and newly determined target prediction model configuration into the orchestration audit log.

[0075] If the normalized drift intensity does not reach the preset entry condition, or if it reaches the preset entry condition but does not meet the minimum trigger interval constraint, the system will not trigger re-arrangement and will use the determined target prediction model configuration as the currently effective prediction model configuration. Thus, this exemplary scenario illustrates that in actual residential load forecasting tasks, the present invention can generate a set of candidate training configurations based on the user's current load behavior profile and perform controlled re-arrangement when the user's electricity consumption behavior changes. However, this example does not limit the applicable data granularity, user type, candidate configuration form, or specific prediction model structure of the present invention.

[0076] Example 2: Based on the same inventive concept as Embodiment 1, this invention also provides an adaptive prediction model orchestration system based on load behavior profiling, used to implement the adaptive prediction model orchestration method based on load behavior profiling described in Embodiment 1, including: The data acquisition and window construction module is used to acquire the historical load time series data of the target user, and construct a time series window on the historical load time series data to obtain a training window, a verification window and a sliding behavior window. The behavior profile extraction module is used to extract load behavior features from the sliding behavior window to obtain load behavior profile results. The candidate configuration generation and verification module is used to generate a set of candidate training configurations based on the load behavior profile results, and to train and verify the candidate configurations in the set of candidate training configurations based on the training window and the verification window to obtain the verification evaluation value corresponding to each candidate configuration. The target configuration determination module is used to perform tolerance screening based on the verification evaluation value, and determine the target prediction model configuration from the candidate training configuration set by combining the candidate configuration complexity and the load behavior profile results. The audit log module is used to record the load behavior profiling results, the candidate training configuration set, the verification evaluation value, and the target prediction model configuration to obtain the orchestration audit log; The drift detection module is used to acquire the updated load time-series data of the target user, perform load behavior drift detection based on the updated load time-series data, and obtain the load behavior drift result. The re-arrangement control module is used to perform re-arrangement control based on the load behavior drift results and the orchestration audit records to obtain the current effective prediction model configuration.

[0077] The modules of the adaptive prediction model orchestration system based on load behavior profiling proposed in this embodiment are used to implement the steps of the adaptive prediction model orchestration method based on load behavior profiling in Embodiment 1. To avoid redundancy, they will not be described again here.

[0078] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied 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.

[0079] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations 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, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. An adaptive prediction model orchestration method based on load behavior profiling, characterized in that, include: To obtain historical load time-series data of the target user, a time-series window is constructed on the historical load time-series data to obtain a training window, a validation window, and a sliding behavior window. The method is as follows: the historical load time-series data is sorted by timestamp and the data validity is checked to obtain effective load time-series data. The effective load time-series data is segmented according to time sequence to obtain the initial training window and the initial validation window; Perform anti-leakage verification on the initial training window and the initial verification window to determine whether the initial training window and the initial verification window meet the anti-leakage constraints; When the conditions are met, the initial training window is determined as the training window, and the initial verification window is determined as the verification window; If the conditions are not met, at least one of the following is adjusted: the window boundary of the initial training window, the window boundary of the initial verification window, and the isolation interval between the initial training window and the initial verification window. The anti-leakage check is then performed again until a training window and a verification window that satisfy the anti-leakage constraint are obtained. Based on the current orchestration time and the preset behavior window length, the effective load time series data is slid-trimmed to obtain the sliding behavior window; Load behavior features are extracted from the sliding behavior window to obtain a load behavior profile. The method is as follows: load behavior features are calculated from the sliding behavior window to obtain a load behavior fingerprint. The load behavior fingerprint is subjected to hierarchical profiling to obtain the load behavior profile result. The load behavior fingerprint includes the coefficient of variation, zero-value ratio, daily autocorrelation, tail quantiles, interquartile range, and tail ratio. The coefficient of variation characterizes the overall load volatility within the sliding behavior window. The zero-value ratio characterizes the proportion of low-load or near-zero-load samples within the sliding behavior window. The daily autocorrelation characterizes the periodic correlation between the current window load and the load lagged by one day. The tail quantiles characterize the intensity of high-load tails and spikes. The interquartile range characterizes the robustness and diffusion of load data. The tail ratio characterizes the exposure of extreme spikes to the high-load tail. Based on the load behavior profile results, a candidate training configuration set is generated, and the candidate configurations in the candidate training configuration set are trained and verified based on the training window and the verification window to obtain the verification evaluation value corresponding to each candidate configuration. Based on the verification evaluation value, tolerance screening is performed, and the target prediction model configuration is determined from the candidate training configuration set by combining the candidate configuration complexity and the load behavior profile results. Record the load behavior profile results, the candidate training configuration set, the validation evaluation value, and the target prediction model configuration to obtain the orchestration audit record; Obtain the updated load time-series data of the target user, and perform load behavior drift detection based on the updated load time-series data to obtain the load behavior drift result; Based on the load behavior drift results and the orchestration audit records, re-orchestration control is performed to obtain the current effective prediction model configuration.

2. The adaptive prediction model orchestration method based on load behavior profiling according to claim 1, characterized in that: The method for performing anti-leakage verification on the initial training window and the initial verification window, and determining whether the initial training window and the initial verification window satisfy the anti-leakage constraints, is as follows: Let the time set corresponding to the initial training window be . The time set corresponding to the initial verification window is The sampling time interval is The preset isolation steps are ; The leakage prevention constraint is deemed satisfied when both the initial training window and the initial verification window simultaneously meet the following conditions: ， ； Otherwise, it is determined that the leakage prevention constraint is not met.

3. The adaptive prediction model orchestration method based on load behavior profiling according to claim 1, characterized in that: The method for performing hierarchical profiling on the load behavior fingerprint to obtain the load behavior profile result is as follows: Based on the load behavior fingerprint, the primary behavior segment corresponding to the target user is determined; Determine whether the first-level behavior segment belongs to the behavior segment that needs to be refined to the second level; When it belongs to the category, determine the corresponding secondary behavior sub-segment of the primary behavior segment; The load behavior profile result is generated based on the first-level behavior segment and the second-level behavior sub-segment; When it does not belong to the category, the load behavior profile result is generated based on the first-level behavior segment.

4. The adaptive prediction model orchestration method based on load behavior profiling according to claim 3, characterized in that: The first-level behavior segments include high zero-value spike type, weak periodic low zero-value type, large sample low zero-value type, stable periodic low zero-value type, high zero-value periodic type, and default safe type; The primary behavior segments that require secondary refinement are the weak-period low-zero value type and the stable-period low-zero value type. When the first-level behavior segment is a weak-period low-zero value type, the first-level behavior segment is further refined based on the daily periodic autocorrelation, coefficient of variation and tail peak characteristics to obtain a second-level behavior sub-segment used to characterize the periodic strength, fluctuation degree and peak tail state. When the first-level behavior segment is a stable periodic low-zero value type, the first-level behavior segment is further refined based on the tail peak characteristics to obtain a second-level behavior sub-segment used to characterize the peak tail state in the stable periodic load.

5. The adaptive prediction model orchestration method based on load behavior profiling according to claim 1, characterized in that: The method for training and validating candidate configurations in the candidate training configuration set based on the training window and the validation window to obtain the validation evaluation value corresponding to each candidate configuration is as follows: Training is performed on each candidate configuration in the candidate training configuration set based on the training window; Based on the verification window, each candidate configuration after training is verified, and the verification evaluation value corresponding to each candidate configuration is calculated according to the preset evaluation rules. The training and verification processes for each candidate configuration use the same training window, the same verification window, and the same leakage prevention constraints. During training, the timing of training termination for each candidate configuration is controlled according to preset training stop conditions; When training resources allow, parallel training is performed on each candidate configuration. When training resources are limited, training is performed sequentially according to the priority of the candidate configurations, and basic fallback candidates are retained.

6. The adaptive prediction model orchestration method based on load behavior profiling according to claim 1, characterized in that: The method for determining the target prediction model configuration from the candidate training configuration set based on the verification evaluation value, tolerance screening, and combining the candidate configuration complexity and the load behavior profile results is as follows: A reference evaluation value is determined based on the verification evaluation value of each candidate configuration, and a tolerance candidate set is constructed based on the reference evaluation value and a preset tolerance threshold. The tolerance candidate set includes candidate configurations whose verification evaluation value is not greater than the sum of the reference evaluation value and the preset tolerance threshold. In the tolerance candidate set, the target prediction model configuration is obtained by selecting from the complexity metric of each candidate configuration and the complexity preference corresponding to the load behavior profile result.

7. The adaptive prediction model orchestration method based on load behavior profiling according to claim 1, characterized in that: The method for obtaining the updated load time-series data of the target user, and performing load behavior drift detection based on the updated load time-series data to obtain the load behavior drift result is as follows: The sliding behavior window is updated based on the updated load time series data, and the load behavior fingerprint is re-extracted from the updated sliding behavior window to obtain the current behavior fingerprint; Obtain a reference behavior fingerprint, which is obtained from the orchestration audit record, and the reference behavior fingerprint is a load behavior fingerprint corresponding to the determined target prediction model configuration; The load behavior drift result is determined based on the degree of distribution difference between the current behavior fingerprint and the reference behavior fingerprint.

8. The adaptive prediction model orchestration method based on load behavior profiling according to claim 1, characterized in that: The method for re-arrangement control based on the load behavior drift results and the orchestration audit records is as follows: When the load behavior drift result indicates that the drift intensity reaches the preset entry condition and meets the minimum trigger interval constraint, the re-arrangement is triggered, and the steps of candidate training configuration set generation, training verification and target prediction model configuration determination are re-executed, the target prediction model configuration is re-determined, and the re-determined target prediction model configuration is used as the currently effective prediction model configuration. When the load behavior drift result indicates that the drift intensity reaches the preset entry condition but does not meet the minimum trigger interval constraint, re-arrangement is not triggered, and the determined target prediction model configuration is used as the currently effective prediction model configuration. When the load behavior drift result indicates that the drift intensity is lower than or equal to the preset exit condition, the re-arrangement is not triggered, the re-arrangement state is exited, and the determined target prediction model configuration is used as the currently effective prediction model configuration. When the load behavior drift result indicates that the drift intensity is between the preset entry condition and the preset exit condition, re-arrangement is not triggered, and the determined target prediction model configuration is used as the currently effective prediction model configuration. Wherein, the drift intensity threshold corresponding to the preset entry condition is higher than the drift intensity threshold corresponding to the preset exit condition; The minimum trigger interval constraint is used to limit the minimum time interval between two adjacent re-arrangements, and the minimum trigger interval constraint is determined based on the last re-arrangement time in the orchestration audit record.

9. An adaptive prediction model orchestration system based on load behavior profiling, used to implement the adaptive prediction model orchestration method based on load behavior profiling as described in any one of claims 1 to 8, characterized in that, include: The data acquisition and window construction module is used to acquire the historical load time series data of the target user, and construct a time series window on the historical load time series data to obtain a training window, a verification window and a sliding behavior window. The behavior profile extraction module is used to extract load behavior features from the sliding behavior window to obtain load behavior profile results. The candidate configuration generation and verification module is used to generate a set of candidate training configurations based on the load behavior profile results, and to train and verify the candidate configurations in the set of candidate training configurations based on the training window and the verification window to obtain the verification evaluation value corresponding to each candidate configuration. The target configuration determination module is used to perform tolerance screening based on the verification evaluation value, and determine the target prediction model configuration from the candidate training configuration set by combining the candidate configuration complexity and the load behavior profile results. The audit log module is used to record the load behavior profiling results, the candidate training configuration set, the verification evaluation value, and the target prediction model configuration to obtain the orchestration audit log; The drift detection module is used to acquire the updated load time-series data of the target user, perform load behavior drift detection based on the updated load time-series data, and obtain the load behavior drift result. The re-arrangement control module is used to perform re-arrangement control based on the load behavior drift results and the orchestration audit records to obtain the current effective prediction model configuration.

Images