A power prediction online learning adaptive method and system based on large model supervision

Abstract

The application provides a power prediction online learning adaptive method and system based on large model supervision, relates to the technical field of artificial intelligence and power system, and aims to solve the problem of performance decline of a power prediction model online due to non-stationarity of an environment; the method comprises the following steps: performing power prediction by using a small model; solidifying prediction task information into a structured feedback event and obtaining associated unstructured operation information; identifying an error state of the prediction; when the error state meets a condition, analyzing the feedback event and the unstructured information by using a large model to generate a structured training guidance signal, the large model not directly predicting power; checking the training guidance signal; and according to the error state and the checked signal, making the small model perform a hierarchical adaptive update, and preferentially adopting a strategy level online learning, and then adopting a parameter level learning when the strategy level online learning is not effective.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and power system technology, and in particular to an online learning adaptive method and system for power prediction based on large model supervision. Background Technology

[0002] Power prediction systems play a crucial role in scenarios such as power dispatch and renewable energy consumption. However, the power system operates in a constantly changing and non-stationary environment. Factors such as seasonal changes, sudden weather changes, adjustments to user production plans, and maintenance of power generation equipment can all cause deviations in the patterns learned from historical data. This results in a significant performance degradation of many prediction models that perform well during offline training after being deployed online for a period of time.

[0003] To address this issue, existing technologies typically employ multi-model ensemble learning methods, combined with online update mechanisms to maintain the predictive performance of the models. For example, some solutions integrate multiple different predictive models and adjust the system through two update actions: one is policy-level updates, such as dynamically adjusting the fusion weights of each model based on its recent predictive performance; the other is parameter-level updates, such as updating the parameters of each model or retraining it with new data after the system has run for a fixed period or number of iterations.

[0004] However, the aforementioned existing technologies still have the following shortcomings: First, they do not adequately utilize unstructured information. In actual operation, a large amount of key information exists in the form of unstructured text, such as text describing weather processes, equipment maintenance plans, and production scheduling arrangements. This information can often explain the source of prediction errors, but traditional numerical models cannot directly utilize this information, resulting in poor adaptability of the prediction system. Second, the update triggering mechanism is relatively rigid. Parameter-level retraining is usually triggered based on fixed time periods, rather than according to the severity of the actual problem. This approach cannot intelligently distinguish between short-term random fluctuations and continuous model performance degradation, which may lead to unnecessary high-cost retraining or untimely response to the real problems that need to be solved. Third, there is a lack of refined error attribution. The system usually judges the quality of the model only based on the overall error index, making it difficult to distinguish whether the error stems from failure in a specific scenario, abnormal input data, or degradation of the model's own capabilities, resulting in a lack of targeted countermeasures. Finally, the online adjustment process lacks security control and traceability. Automated online adjustment processes may introduce unknown risks. Without strict verification, version management, and rollback mechanisms, it will be difficult to trace the root cause of the problem and quickly restore the system to stable operation if an adjustment goes wrong. Summary of the Invention

[0005] The purpose of this application is to provide an online learning adaptive method and system for power prediction based on large model supervision. It aims to solve the problem of how to safely, controllably, and efficiently utilize unstructured operational information, including text, in the continuously changing online operating environment of power prediction to achieve continuous adaptation of the prediction system, so as to overcome the problem of model performance degradation caused by environmental non-stationarity.

[0006] To achieve the above objectives, this application provides an online learning adaptive method for power prediction based on large model supervision, comprising the following steps: a large model supervises the updates of small models, specifically including: using one or more small models to perform power prediction; solidifying relevant information of the power prediction task into structured feedback events, and acquiring unstructured operational information associated with the power prediction task, wherein the feedback events at least contain information for identifying the prediction context and prediction results; identifying the current prediction error state; when the error state meets a preset trigger condition, a large model supervises and analyzes the feedback events and the unstructured operational information to generate a structured training guidance signal, wherein the large model does not directly output power prediction values; verifying the training guidance signal to determine whether it meets preset verification rules; and, based on the identified error state and the verified training guidance signal, enabling the small models to perform hierarchical adaptive updates, wherein the hierarchical adaptive updates preferentially employ policy-level online learning, and when policy-level online learning fails to correct the error state below a preset threshold, parameter-level learning is then employed.

[0007] Optionally, the feedback event may further include at least one of the following: input version, model version, policy version, unstructured runtime information, and training guidance signals generated by the large model and adopted in connection with the historical power prediction task.

[0008] Optionally, the structured training guidance signal includes at least one of the following: error attribution, scene label, sample reweighting suggestion, candidate feature suggestion, fusion weight adjustment suggestion, online correction suggestion, and retraining suggestion.

[0009] Optionally, the step of verifying the training guidance signal includes at least one of the following steps: a) performing structured parsing and consistency verification on the training guidance signal; b) performing verification on at least one of permissions, sample size, amplitude, and range.

[0010] Optionally, the policy-level online learning includes: adjusting the fusion weights between the small models, or enabling online prediction correction rules; the parameter-level learning includes: performing local retraining, or performing full retraining.

[0011] Optionally, the method further includes at least one of the following: a) performing versioned association management of input version, model version, policy version and training guidance signal; b) employing a canary release and automatic rollback mechanism for the hierarchical adaptive update.

[0012] To achieve the above objectives, this application also provides an online learning adaptive system for power prediction, comprising: one or more small models for performing power prediction; a feedback event construction module for solidifying relevant information of the power prediction task into structured feedback events and acquiring unstructured operational information associated with the power prediction task, wherein the feedback events at least include information for identifying the prediction context and prediction results; an error state identification module for identifying the current prediction error state; a large model configured to analyze the feedback events and the unstructured operational information to generate a structured training guidance signal when the error state meets a preset trigger condition, wherein the large model does not directly output power prediction values; a guidance signal verification and filtering module for verifying the training guidance signal to determine whether it meets preset verification rules; and an adaptive update module for enabling the small models to perform hierarchical adaptive updates based on the identified error state and the verified training guidance signal, wherein the hierarchical adaptive updates preferentially employ policy-level online learning, and the adaptive update module is further configured to perform parameter-level learning when the policy-level online learning fails to correct the error state to below a preset threshold.

[0013] Optionally, the guidance signal verification and filtering module is configured to perform at least one of the following: a) structured parsing and consistency verification; b) verification of at least one of permissions, sample size, amplitude, and range.

[0014] Optionally, the adaptive update module is configured such that: the policy-level online learning includes: adjusting the fusion weights between the small models, or enabling online prediction correction rules; the parameter-level learning includes: performing local retraining, or performing full retraining.

[0015] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the preceding claims.

[0016] Compared with existing technologies, this application has the following beneficial effects: It improves information utilization and prediction accuracy. Through a collaborative mechanism of "large model supervision and small model execution," unstructured text information such as weather descriptions and maintenance records, which are difficult for traditional models to utilize, is transformed into structured training signals usable by the small model, such as scene labels, sample weights, and candidate features, thereby improving the prediction accuracy of the system in complex scenarios. It ensures system security and controllability. By structuring and multi-step verifying the guidance signals generated by the large model, and combining version management, canary release, and automatic rollback mechanisms, the output of the large model is effectively constrained, ensuring the interpretability, observability, and rollback capability of the online automatic learning process, thus guaranteeing the stability and security of the system. It achieves efficient response and cost savings. Through a hierarchical adaptive update mechanism, short-term deviations are quickly responded to through lightweight strategy-level adjustments. More costly parameter-level retraining is only triggered when the problem persists or undergoes structural changes, avoiding unnecessary retraining due to misjudgments, thus balancing response speed and operational costs. The system's adaptability and robustness are enhanced. In non-stationary environments, the system can continuously track model performance and perform adaptive optimization through a feedback event-driven closed-loop learning mechanism, avoiding performance degradation caused by static operation after the model is deployed. Furthermore, the risk of single model failure is reduced through dynamic fusion of multiple models. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the structure of the online learning adaptive system provided in the embodiments of this application;

[0019] Figure 2 A flowchart illustrating the online learning adaptive method provided in this application embodiment;

[0020] Figure 3 This is a schematic diagram of the guidance signal verification process provided in the embodiments of this application;

[0021] Figure 4 This is a schematic diagram of the hierarchical adaptive update process provided in an embodiment of this application.

[0022] The main reference numerals in the attached figures are explained as follows: 100 - Online learning adaptive system; 110 - Small model; 120 - Large model; 130 - Feedback event construction module; 140 - Error state identification module; 150 - Guidance signal verification and filtering module; 160 - Adaptive update module; 170 - Version management and security control module; S201 - Perform power prediction; S202 - Construct feedback events; S203 - Identify error states; S204 - Determine if analysis is triggered; S205 - Large model generates guidance signals; S206 - Verify guidance signals; S207 - Perform hierarchical adaptive updates; S301 - Receive guidance signals; S302 - Structure and consistency verification; S303 - Permission / sample size / amplitude / range verification; S304 - Hierarchical routing; S305 - Rejection / degradation; S401 - Determine error states; S402 - Prioritize policy-level learning; S403 - Evaluate effects; S404 - Determine if the problem is resolved. S405 - Execute parameter-level learning. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings, so as to better understand the purpose, technical solutions and advantages of this application. It should be noted that the specific embodiments described herein are only used to explain this application, and not to limit this application.

[0024] Please see Figure 1 The diagram illustrates the structure of an online learning adaptive system 100 according to an embodiment of this application. The system 100 is used to implement online continuous adaptive optimization of a power prediction model. As an optional implementation, the system 100 can be deployed on a cloud computing platform, edge computing node, or server cluster within an enterprise. Accordingly, the core architecture of the system 100 embodies the layered collaborative concept of "large model supervision, small model execution, and control module ensuring security."

[0025] In one embodiment of this application, system 100 includes one or more small models 110, a large model 120, a feedback event construction module 130, an error state identification module 140, a guidance signal verification and filtering module 150, an adaptive update module 160, and a version management and security control module 170.

[0026] The small model 110 is a functional entity that performs a specific power prediction task. In one embodiment, multiple small models 110 of different types or with different parameter configurations can be deployed simultaneously. For example, the small model 110 may include, but is not limited to, models based on gradient boosting decision trees (such as LightGBM, XGBoost), neural network models (such as Long Short-Term Memory networks), or traditional statistical models (such as autoregressive integral moving average models). These small models 110 receive real-time input data, such as historical power data, numerical weather forecasts, date types, etc., and output power prediction values ​​for one or more future time points. By employing a multi-model ensemble approach, the weights of each model can be dynamically adjusted to improve the robustness of the prediction.

[0027] The feedback event construction module 130 is used to establish the correlation between online prediction and offline analysis. Specifically, after a small model 110 completes a prediction, this module captures the entire chain information of the task and obtains the associated unstructured operational information from external systems (such as operation and maintenance log systems and meteorological service systems), and then organizes this information into a structured feedback event. A feedback event may include: a unique identifier of the prediction task, a timestamp, the version of the input data used, the model version, the strategy version, the prediction output of each small model 110, the final fused prediction value, the actual power value of delayed recovery, a multi-dimensional error summary calculated based on the prediction value and the actual value (such as mean absolute percentage error, root mean square error, etc.), and unstructured operational information text associated with the prediction period, such as "temporary production increase, night shift production line fully operational" or "weather forecast: continuous rain in the next three days".

[0028] The error state identification module 140 is responsible for continuously monitoring and diagnosing the system's predictive performance. This module receives a sequence of feedback events generated by the feedback event construction module 130 and performs statistical analysis on the sequence. To achieve refined diagnosis, this module can maintain multiple observation windows at different time scales, such as short windows (e.g., the past 24 hours), medium windows (e.g., the past 7 days), long windows (e.g., the past 30 days), and windows for similar scenarios (e.g., all historical weekend mornings). By analyzing error indicators (e.g., mean, variance, and duration of exceeding thresholds) within different windows, and combining this with scenario labels and unstructured information in the feedback events, this module can identify the current error state of the prediction, such as whether it is a normal random fluctuation, an anomaly caused by a short-term event, a continuous degradation of model capabilities, or a failure under specific scenarios.

[0029] The large model 120 plays a core analytical and guiding role in the technical solution of this application, responsible for deeply understanding the scenario and providing optimization guidance. Unlike the small model 110, the large model 120 is typically a large language model pre-trained with massive amounts of data, possessing natural language understanding and logical reasoning capabilities. In this application, the large model 120 does not directly participate in high-frequency power prediction calculations, but is positioned as an analysis and guidance component. When the error state identification module 140 determines that the error state meets preset triggering conditions (e.g., the error exceeds 10% for three consecutive days), it initiates an analysis request to the large model 120. This request contains structured input, such as recent error summaries, changes in the performance ranking of each small model 110, and related unstructured operational information text. The large model 120 comprehensively analyzes this information to generate a structured training guidance signal. It is understood that this signal is a candidate suggestion for small model updates, rather than a direct prediction result or control command.

[0030] The guidance signal verification and filtering module 150, as a key component ensuring system security, receives training guidance signals from the large model 120 and performs a series of automated verifications to ensure the rationality, safety, and executability of the recommendations. Specifically, the verification process may include: structured parsing and consistency verification to ensure that the signal conforms to a predefined format and that its attribution does not contradict the structured indicators in the feedback event; and permission, sample size, magnitude, and range verification to determine whether the recommended action is within the automated permissions, whether the relevant data samples are sufficient, whether the adjustment range of the recommendation is within the safety limit, and whether it has clear activation and deactivation conditions. Based on the verification results, the module will route the guidance signals to different execution paths in a tiered manner. For example, low-risk recommendations are directly approved, while high-risk recommendations are rejected or downgraded to manual review.

[0031] The adaptive update module 160 is used to specifically perform model optimization operations. This module receives update instructions approved by the guidance signal verification and filtering module 150, and performs hierarchical adaptive updates based on the currently identified error state. This module follows a core principle: it prioritizes low-cost, low-risk policy-level learning; only when policy-level learning fails to solve the problem does it upgrade to higher-cost, more far-reaching parameter-level learning.

[0032] The version management and security control module 170 is responsible for the traceability, recoverability, and security evolution of the entire system. This module performs unified version-based association management of all key entities in the system, including the versions of input data, the small model 110, the adaptive update strategy, and adopted training guidance signals. In other words, any update operation can be traced back to which version of the guidance signal it was triggered by, which version of the input and model it was based on, and what problem it was designed to solve. Furthermore, this module implements canary release and automatic rollback mechanisms. For example, a new model version or update strategy will first undergo "shadow" evaluation or canary release on a small portion of the traffic. The system will continuously monitor its online performance. If the effect is not as expected or deteriorates, the module can automatically roll back the system to the previous stable version, thereby controlling the risks during the online learning process.

[0033] Please see Figure 2This is a flowchart illustrating an online learning adaptive method provided in one embodiment of this application. The method is a continuously running closed-loop process, specifically including the following steps: Step S201, Perform power prediction. One or more small models 110 in the system receive real-time input data and generate power prediction results. Step S202, Construct feedback events. The feedback event construction module 130 solidifies the context information of this prediction task (such as input version, model version), prediction results, actual power values ​​of delayed recovery, calculated error summaries, and associated unstructured operational information into a structured feedback event and stores it in a database or event stream. Step S203, Identify error states. The error state identification module 140 continuously analyzes historical feedback event sequences and determines the current system error state through multi-window statistics and pattern recognition. Step S204, Determine whether large model supervised analysis is triggered. Specifically, the system checks whether the currently identified error state meets preset trigger conditions, such as "the average absolute percentage error during the evening peak period is greater than 8% for three consecutive days" or "unstructured events related to 'equipment maintenance' are detected, and the error during the corresponding time period increases significantly." If the conditions are not met, the process returns and waits for the next prediction cycle; if the conditions are met, it indicates that the system may have an anomaly requiring intervention, and the process continues to execute step S205. Step S205: The large model generates guidance signals. The system submits structured information containing recent error summaries, model performance, and related unstructured text to the large model 120. The large model 120 analyzes the information and generates structured training guidance signals as candidate suggestions for small model updates. Step S206: The guidance signals are verified. The guidance signal verification and filtering module 150 performs rigorous multi-step verification on the training guidance signals generated by the large model 120. Step S207: Hierarchical adaptive updates are performed. The adaptive update module 160 performs corresponding update operations on the small model 110 based on the identified error state and the verified training guidance signals. After the update is completed, the small model 110 continues to perform the prediction task of step S201 in the new state, thus forming a complete and continuously optimized closed loop.

[0034] Example 1

[0035] This embodiment details how the system can quickly respond to prediction deviations caused by short-term business events through strategy-level online learning. This scenario corresponds to a situation where a low-cost, low-risk adjustment strategy is preferred.

[0036] In a specific application scenario, the online learning adaptive system 100 of this application is used to predict the short-term power load of a large industrial park over the next 24 hours. The system 100 deploys two small models 110: one based on the LightGBM algorithm and the other based on the XGBoost algorithm. The two models are weighted and fused to output the final prediction result.

[0037] In mid-June of a certain year, the system operated stably. However, starting June 15th, the error status identification module 140 detected that during the evening peak hours (specifically 18:00 to 21:00), the system's load forecasts consistently showed significant underestimation, with the average absolute percentage error climbing from the normal 4% to the range of 8% to 12%. For example, on June 17th, the average absolute percentage error during the evening peak hours reached 12.3%. Simultaneously, analysis revealed that the XGBoost-based model exhibited relatively smaller biases during this period, while the LightGBM-based model showed significant performance degradation.

[0038] According to steps S201 and S202, the feedback event construction module 130 generates corresponding feedback events for each day's forecasting task. For example, for the forecasting task on June 17, the generated feedback event includes the following information: the input version is v12, indicating that the latest numerical weather forecast and historical load data for that day were used; the strategy version is S7, indicating that the current fusion weights are LightGBM accounting for 0.35, XGBoost accounting for 0.30, and other models accounting for a total of 0.35; the forecast curves of the two small models 110 and the final fusion forecast curve are recorded; and the actual load curve collected the next day is also recorded. In particular, the system automatically retrieves and associates an unstructured operation information from the operation and maintenance log system: "Following the notification from the production department, in order to catch up with orders, from June 15 to June 25, Workshop No. 2 will temporarily increase production, and the night shift production line will be fully operational."

[0039] In step S203, the error state identification module 140 analyzed the feedback event sequence from June 15th to 17th and identified the error state of "persistent underestimation during the evening peak hours". In step S204, this state met the preset triggering condition of "the same type of error in the same time period lasting more than 3 consecutive days and the average absolute percentage error being greater than 8%". Therefore, the system triggered the large model supervised analysis process.

[0040] In step S205, the system sends a structured request to the large model 120. This request contains key information, such as: an error summary ("During the evening peak hours of 18:00-21:00, the mean absolute percentage error has increased from 4.1% to 12.3% for three consecutive days, showing a continuous underestimation"), model performance ("The LightGBM model contributes the most to the error, while XGBoost's performance is relatively stable"), and key unstructured text ("Temporary production increase from June 15th to 25th, with all night shift production lines operating"). The large model 120 utilizes its semantic understanding capabilities to establish a causal relationship between "temporary production increase" and "all night shift production lines operating" and "underestimation of evening peak load forecast." Subsequently, it generates a structured training guidance signal, the content of which can be represented in the following JSON format:

[0041] {

[0042] "error_attribution": "production_surge",

[0043] "scene_labels": ["production_surge_evening"],

[0044] "fusion_weight_suggestion": {

[0045] "LightGBM": {"target": 0.20, "reason": "performance_degradation"},

[0046] "XGBoost": {"target": 0.45, "reason": "better_adaptation"}

[0047] },

[0048] "correction_suggestion": {

[0049] "scope": {"time_period": ["18:00", "21:00"], "date_type": ["workday"]},

[0050] "action": "add_percentage",

[0051] "value": 8.0,

[0052] "lifespan": {"end_date": "2023-06-25"}

[0053] },

[0054] "sample_reweighting": {

[0055] "scope": {"scene": "production_surge_evening"},

[0056] "weight_multiplier": 2.5,

[0057] "action_type": "retraining_candidate"

[0058] }

[0059] }

[0060] This guidance includes recommendations across multiple levels: 1. Error Attribution: Attributing the problem to a "production surge." 2. Scene Labeling: Recommending a new scene label, production_surge_evening, for easier future identification. 3. Fusion Weight Adjustment Recommendation: Recommending a reduction in the weight of the underperforming LightGBM model from 0.35 to 0.20, while increasing the weight of the better-performing XGBoost model from 0.30 to 0.45. 4. Online Correction Recommendation: Recommending an 8% upward correction to the final prediction results during weekdays from 18:00 to 21:00, with the rule automatically expiring after June 25th. 5. Sample Reweighting Recommendation: Recommending assigning 2.5 times the weight to anomalous samples from the recent evening rush hours in future model training, but marking them as "retraining candidates," indicating this is a suggestion for parametric learning.

[0061] In step S206, the signal verification and filtering module 150 is instructed to verify the signal. The specific process can be found in [reference needed]. Figure 3The illustrated embodiment. In step S302, the structured analysis and consistency verification of the signal pass. In step S303, the module performs multi-dimensional verification: the magnitude of the fusion weight adjustment (-0.15 and +0.15) is within the preset single adjustment safety limit (e.g., ±0.20); the magnitude of the online correction (+8%) is also within the safety limit (e.g., ±10%), and has a clear effective range and exit conditions. Therefore, these two suggestions are considered low-risk. However, for the sample reweighting suggestion, since its associated action is "retraining", permission and sample size verification are triggered. The module detects that there are only 3 days of samples that currently meet the production_surge_evening scenario label, which is far below the minimum sample size threshold (e.g., 14 days) required to trigger local retraining. In step S304, the hierarchical routing logic determines that the sample reweighting suggestion fails verification due to "insufficient conditions". The suggestion is downgraded to "observation" and not executed for the time being. Its status and reason are recorded in the version management and security control module 170.

[0062] In step S207, the adaptive update module 160 receives the verified fusion weight adjustment and online correction suggestions. The specific update process can be found in [reference needed]. Figure 4 The illustrated embodiment. In step S401, the system determines that the current error state is "event-driven short-term bias," suitable for policy-level learning. Therefore, in step S402, the system performs policy-level learning: a new policy version S8 is generated, in which the fusion weights of LightGBM and XGBoost are adjusted to 0.20 and 0.45 respectively, and an online correction rule is enabled to apply a +8% correction to the prediction results from 18:00 to 21:00 on weekdays. This new policy S8 takes effect online through the canary release mechanism of the version management and security control module 170 (e.g., first effective for 50% of the prediction tasks).

[0063] During the subsequent observation period from June 18th to 20th, the evaluation results in step S403 showed that the average absolute percentage error of the evening peak load forecast under the new strategy S8 decreased to approximately 4.5%, indicating that the problem had been corrected. In step S404, the system determined that the problem had been resolved. By June 26th, after the temporary production increase event ended, the online correction rules automatically expired due to the end of their lifecycle, and the system forecast returned to normal, avoiding unnecessary corrections under normal operating conditions. This embodiment illustrates the process by which the system utilizes strategy-level learning to cope with short-term non-stationary events.

[0064] Example 2

[0065] This embodiment will focus on how the large model 120 extracts valuable signals from unstructured text information and transforms them into structured features that the small model 110 can utilize, thereby supporting deeper and more generalizable model optimization.

[0066] The prediction target in this scenario is the day-ahead power generation of a large photovoltaic power station. Photovoltaic power output prediction is highly sensitive to weather conditions, especially cloud cover and rainfall. Traditional models typically use structured data such as irradiance, temperature, and humidity from numerical weather prediction as input. However, for some complex weather processes, text-based weather forecast descriptions often contain additional information.

[0067] One day, the system received a weather forecast text containing the following description: "Over the next three days, influenced by an upper-level trough and a low-level shear line, the local area will experience a continuous period of cloudy and rainy weather. Clouds will move rapidly, and sky conditions will be complex and changeable. Localized short-term severe convective weather may occur in the afternoon. Please take precautions." In the following days, the system observed significant errors in the photovoltaic power output prediction at multiple times, particularly with the prediction curve being too smooth, failing to capture the drastic fluctuations in power output caused by rapid cloud movement and severe convection.

[0068] follow Figure 2 As shown in the process, the feedback event constructed by the system in step S202 not only includes the usual predicted value, actual value, and error, but also specifically includes the textual description of the aforementioned weather forecast. In step S204, due to the prediction error exceeding the threshold, large model supervised analysis is triggered.

[0069] In step S205, the large model 120 receives an analysis request containing the aforementioned weather text. The large model 120 utilizes its natural language processing capabilities to parse this text. It can not only identify keywords such as "rainy" and "strong convection," but also understand words describing the dynamics and uncertainty of the process, such as "continuity," "fast cloud movement," and "complex and changeable." Based on this understanding, the large model 120 generates a training guidance signal that prominently outputs a section of feature suggestions, as follows:

[0070] { "error_attribution": "complex_weather_process", "feature_suggestions": [ { "feature_name": "continuous_rain_days", "type": "integer", "description": "Number of consecutive rainy days, used to capture the cumulative impact of persistent weather on photovoltaic panel efficiency and ground surface temperature. Calculated based on historical weather records." "source": "weather_text_analysis" }, { "feature_name": "irradiance_drop_flag", "type": "binary", "description": "Irradiance drop markers are used to identify events such as strong convection or rapid passage of thick clouds that may cause a sudden and significant drop in power. They can be generated by combining cloud imagery or higher-frequency irradiance data." "source": "weather_text_analysis" }, { "feature_name": "cloud_movement_speed_quantified", "type": "float", "description": "A quantitative indicator of 'fast cloud movement.' It is suggested to explore the possibility of extracting this feature from more detailed meteorological data (such as radar echo maps)." "source": "weather_text_analysis" } ], ... }

[0071] It can be seen that this guidance signal transforms vague, descriptive textual information into three specific, actionable, structured feature suggestions:

[0072] 1. continuous_rain_days: An integer feature used to quantify the concept of "continuous rain".

[0073] 2. irradiance_drop_flag: A binary flag used to capture the transient effects that "strong convection" may bring.

[0074] 3. cloud_movement_speed_quantified: A floating-point feature that attempts to quantify the dynamic process of "cloud movement speed".

[0075] In step S206, the guidance signal verification and filtering module 150 verifies these feature suggestions. The verification focuses on confirming the logical rationality, type correctness, and feasible computational source of these feature definitions. Since these are only "suggestions," the risk is low, and therefore the verification passes.

[0076] Understandably, validated feature suggestions do not immediately trigger model updates. Instead, they are stored in a feature engineering candidate library and associated with the current error event. The version management and security control module 170 records the source of these candidate features (generated by the large model 120 on a specific date for a specific weather event). The system can automatically or manually trigger an offline feature validation task. Data engineers or algorithm engineers can develop specific feature calculation logic based on the descriptions and suggestions from the large model 120. For example, `continuous_rain_days` can be implemented by iterating through historical daily weather conditions; `irradiance_drop_flag` can be implemented by detecting the rate of change of minute-level irradiance data.

[0077] In subsequent parameter-level learning tasks (such as local retraining or full retraining as described in Example 3), these validated new features will be incorporated into the training input of the small model 110. This enables the small model 110 to learn the complex nonlinear relationship between the dynamic characteristics of weather processes and photovoltaic power output.

[0078] The effect of this embodiment is that by transforming textual semantics such as "continuous rain" and "strong convection" into structured features, the small model 110 obtains richer input information. When encountering similar weather forecast descriptions again in the future, even if the numerical weather prediction of irradiance still has deviations, the model can use these new features to predict drastic fluctuations in output, thereby making more accurate and robust predictions and improving the model's generalization ability to specific complex weather events.

[0079] Example 3

[0080] This embodiment aims to illustrate how a system can intelligently upgrade to more costly parameter-level learning (e.g., local retraining) when it faces structural and continuous environmental changes and policy-level learning is insufficient to solve the problem.

[0081] Scenario Description: The target of the forecast is the net load of a certain integrated energy station. This energy station is equipped with an energy storage system. For a long time, the control strategy of the energy storage system was "economic optimization," that is, charging when the electricity price is low and discharging when the electricity price is high. The small model 110 in the system has learned the load curve under this mode. However, starting from the 1st of a certain month, in order to respond to the peak-shaving demand of the power grid, the energy storage control strategy of the station was switched to "peak shaving priority" for an extended period, that is, forced discharging during peak load periods and forced charging during off-peak periods. This change has caused a continuous and structural change in the shape of the net load curve of the station.

[0082] In the initial stage of strategy switching, the system prediction error increased sharply. The error state identification module 140 identified the error state as "continuous degradation". Based on... Figure 4 In the illustrated hierarchical update process, in step S402, the adaptive update module 160 first attempts policy-level learning, such as adjusting the fusion weights between multiple small models 110 and enabling some online correction rules. These measures alleviate the error to some extent, but the effect is limited, and the mean absolute percentage error remains high at over 15%, far exceeding the normal threshold of 5%.

[0083] In step S403, after continuous observation and effect evaluation over a medium window (e.g., 14 days), the system finds that the online strategy adjustment is ineffective. In step S404, the system determines that the problem has not been resolved. Simultaneously, the feedback event construction module 130 records the feedback that the strategy-level adjustment is ineffective and retrieves the unstructured information "the energy storage strategy was switched to peak shaving priority on month X, 1" from the operation and maintenance records.

[0084] Due to the failure of policy-level learning, and with sufficient evidence (high error for 14 consecutive days, clear external events), the system decides to escalate the issue. Step S205 is triggered again, and the large model 120 receives an analysis request containing a detailed sequence of feedback events and operational records from the past 14 days. After analysis, the large model 120 attributes the error to "energy storage policy switching" and generates a training guidance signal suggesting parameter-level learning. This signal contains a clear retraining suggestion, as follows:

[0085] { "error_attribution": "storage_strategy_change", "retraining_suggestion": { "type": "local_retraining", "reason": "Structural change in load pattern, policy-level learningineffective.", "target_samples": { "start_date": "2023-XX-01", "end_date": "latest" }, "sample_reweighting": { "scope": "all_target_samples", "weight_multiplier": 3.0 }, "feature_candidates": [ { "feature_name": "storage_strategy_mode", "type": "categorical", "values": ["economic_optimization", "peak_shaving_priority"], "source": "manual_input_or_log_parsing" } ] }

[0086] }

[0087] The guidance specifically recommends: 1. Recommendation type: Explicitly recommends "local retraining". 2. Target samples: Select all feedback events since the strategy switch (from January 1st to present) as training samples. 3. Sample reweighting: Recommend increasing the weights of samples in these new modes by a factor of 3 to force the model to focus on the new patterns. 4. Candidate features: Recommend introducing a new classification feature, storage_strategy_mode, to fundamentally distinguish between the two different operating modes.

[0088] In step S206, the signal verification and filtering module 150 verifies the signal. In step S303, the permission verification confirms that the system allows partial retraining; the sample size verification confirms that the 14 days of data meets the minimum retraining sample requirement. Therefore, the retraining suggestion is determined to be "passed" in step S304.

[0089] In step S207, after receiving the instruction, the adaptive update module 160 performs parameter-level learning in step S405. It creates a local retraining task and updates the parameters of the worst-performing small models 110 using the samples, sample weights, and new features specified in the guidance signal.

[0090] The newly trained model version will not be deployed immediately. The version management and security control module 170 will first organize an offline time-series backtest, using data after the switching strategy to compare and verify the new and old model versions. After confirming that the new model has improved performance in the new scenario (e.g., the mean absolute percentage error drops below 5%) and that its performance on historical data is not significantly impaired, the new model version will be gradually deployed online through a canary release mechanism to replace the old version.

[0091] The result of this embodiment is that, through this evidence-driven local retraining guided by the large model 120, the small model 110 learned a new load pattern under the 'peak shaving first' strategy, thereby solving the prediction bias problem. This process illustrates that when the system's policy-level learning is ineffective, it will switch to parameter-level learning to adapt to environmental changes.

[0092] Example 4

[0093] This embodiment will demonstrate in detail how the guidance signal verification and filtering module 150 plays its security role. Through a strict verification process, it rejects or downgrades suggestions from the large model 120 that are of high risk or do not meet the conditions, thereby ensuring the stability and security of the entire system.

[0094] Scenario Description: A brand-new prediction target (e.g., a new charging station) has just been connected to the power prediction system and has only been running for 3 days. As this is a new service, there is little historical data accumulated, and the equipment may be in a break-in period, resulting in relatively large prediction errors in the initial stage.

[0095] After the system finished running for three days, the error status identification module 140 detected an average absolute percentage error of 20% over the past three days. Simultaneously, the feedback event construction module 130, when constructing the feedback event, associated it with an operational note: "New site online, data collection in progress." According to step S204, the high error met the conditions for triggering analysis, so the system initiated an analysis request to the large model 120.

[0096] In step S205, the large model 120 analyzes the high-error data and operational notes from the past three days. Due to the observed "high errors," the large model 120 might generate a seemingly reasonable suggestion: immediately weight these high-error samples and perform a local retraining of the model to quickly "adapt" to the new site. It generates a training guidance signal containing a retraining suggestion.

[0097] In step S206, the guidance signal is verified and filtered. This verification process can be referred to... Figure 3 The process is shown below. In step S301, the guidance signal verification and filtering module 150 receives the guidance signal suggesting retraining. In step S302, the module first performs structured parsing and consistency verification. Assuming that the signal format generated by the large model 120 fully conforms to the predefined JSON structure, and that its motivation for "suggesting retraining" (high error) is consistent with the input error index, this step passes the verification. Next, the process proceeds to step S303, where permissions, sample size, amplitude, and range are verified. Among these, the permission verification confirms that the system configuration allows automated execution of local retraining, and this step passes. However, in the sample size verification stage, the verification module checks the samples on which the retraining suggestion depends and finds that it only has 3 days of data. The system's preset security rules clearly stipulate that the minimum continuous sample size threshold required to trigger any form of retraining is 14 days. Since 3 days is much less than 14 days, the sample size verification item clearly returns "failure".

[0098] Because a critical verification item failed, in step S304, the hierarchical routing module determined the guidance signal to be "high risk, insufficient conditions". Therefore, the execution path of this signal was directed to step S306, namely "rejection / downgrade".

[0099] At this point, the system will not execute the retraining suggestion for the large model. Instead, it will perform the following operations: 1. Reject the retraining action: The adaptive update module 160 will not receive any instructions to execute retraining. 2. Record the reason for rejection: The version management and security control module 170 will record this event in detail. It will save the original guidance signal generated by the large model 120 and attach a clear rejection record, such as: "The retraining suggestion with ID [Signal_ID] was rejected because of insufficient sample size (actually 3 days, required 14 days)". This provides data for subsequent system auditing and optimization of the large model's behavior. 3. Execute downgrade or alternative actions: Although the high-risk retraining is rejected, the system is not inactive. The hierarchical routing module can downgrade the original signal, for example, adopting the lower-risk part of the signal (if it exists), or the system can automatically match a preset, more robust response strategy based on the scenario of "initial launch of a new site", such as temporarily not performing any model updates, but shortening the observation window to more closely monitor subsequent error changes and wait for data accumulation.

[0100] The intended effect of this embodiment is that the system avoids a high-risk retraining operation based on insufficient data. If this retraining is performed blindly, the small model 110 is likely to overfit to the noise or random patterns of these 3 days, resulting in a worse generalization ability.

[0101] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An online learning adaptive method for power prediction based on large model supervision, characterized in that, Includes the following steps: The updates of the smaller model are supervised by a larger model, specifically including: Perform power prediction using one or more small models; The relevant information of the power prediction task is solidified into structured feedback events, and the unstructured operational information associated with the power prediction task is obtained. The feedback events contain at least information for identifying the prediction context and prediction results. Identify the current error state of the prediction; When the error state meets the preset triggering conditions, a large model supervises and analyzes the feedback event and the unstructured operation information to generate a structured training guidance signal. The large model does not directly output the power prediction value. The training guidance signal is verified to determine whether it meets the preset verification rules; Based on the identified error state and the verified training guidance signal, the small model performs hierarchical adaptive updates. The hierarchical adaptive updates preferentially employ policy-level online learning. When policy-level online learning fails to correct the error state to below a preset threshold, parameter-level learning is then employed.

2. The method according to claim 1, characterized in that, The feedback event also includes at least one of the following: input version, model version, policy version, unstructured runtime information, and training guidance signals associated with the historical power prediction task and generated and adopted by the large model.

3. The method according to claim 2, characterized in that, The structured training guidance signals include at least one of the following: error attribution, scene label, sample reweighting suggestion, candidate feature suggestion, fusion weight adjustment suggestion, online correction suggestion, and retraining suggestion.

4. The method according to claim 3, characterized in that, The step of verifying the training guidance signal includes: The training guidance signal is subjected to structured analysis and consistency verification; Verify at least one of the following: permissions, sample size, amplitude, and range.

5. The method according to claim 1, characterized in that: The policy-level online learning includes: adjusting the fusion weights between the small models, or enabling online prediction correction rules; the parameter-level learning includes: performing local retraining, or performing full retraining.

6. The method according to claim 1, characterized in that: Versioned association management is implemented for input version, model version, policy version, and training guidance signal; the hierarchical adaptive update adopts a canary release and automatic rollback mechanism.

7. A power prediction online learning adaptive system based on large model supervision, using the online learning adaptive method as described in any one of claims 1-6, characterized in that, include: One or more small models are used to perform power prediction; The feedback event construction module is used to solidify relevant information of the power prediction task into structured feedback events and obtain unstructured operational information associated with the power prediction task. The feedback events at least include information for identifying the prediction context and prediction results. Error state identification module, used to identify the current prediction error state; A large model is configured to analyze the feedback event and the unstructured operational information to generate a structured training guidance signal when the error state meets a preset trigger condition, wherein the large model does not directly output a power prediction value. The guidance signal verification and filtering module is used to verify the training guidance signal to determine whether it meets the preset verification rules. An adaptive update module is used to enable the small model to perform hierarchical adaptive updates based on the identified error state and the validated training guidance signal. The hierarchical adaptive updates preferentially adopt policy-level online learning. The adaptive update module is further configured to perform parameter-level learning when the policy-level online learning fails to correct the error state to below a preset threshold.

8. The system according to claim 7, characterized in that, The guidance signal verification and filtering module is configured to perform verifications of: structured parsing and consistency; and verification of at least one of permissions, sample size, amplitude, and range.

9. The system according to claim 7, characterized in that, The adaptive update module is configured such that: the policy-level online learning includes: adjusting the fusion weights between the small models, or enabling online prediction correction rules; the parameter-level learning includes: performing local retraining, or performing full retraining.

Images