A database trend prediction method and system driven by working condition and confidence

Abstract

This invention discloses a database trend prediction method and system driven by both operating conditions and confidence levels, belonging to the field of database technology. Its core principle is as follows: It identifies the database operating conditions through a parallel perception mechanism with both long and short windows; employs a hierarchical hysteresis routing mechanism driven by both operating condition characteristics and the historical confidence performance of each model; matches the optimal prediction model from a three-level gradient prediction model pool to perform trend prediction; and outputs the prediction interval and confidence level using an online-offline dual-layer collaborative architecture. The actual prediction error and the matching deviation of the confidence interval serve as feedback signals. A three-dimensional routing strategy table enables end-to-end collaborative iteration of routing weights, confidence calibration parameters, and hysteresis control parameters. This method achieves dynamic synergy between prediction accuracy, computational efficiency, and result reliability, effectively suppressing frequent model switching oscillations at operating condition boundaries and improving the system's long-term adaptive capability to changes in business models.

Description

Technical Field

[0001] This invention relates to the field of database prediction technology, specifically to a database trend prediction method and system driven by both operating conditions and confidence levels. Background Technology

[0002] As enterprises continue to deepen their digital transformation, databases, as core information infrastructure, bear the critical functions of data storage, interaction, and processing for business data. Their operational stability directly determines the continuity and service quality of business systems. Database trend prediction is a core supporting technology of the Intelligent Operations and Maintenance (AIOps) system. By modeling the time-series patterns of multi-dimensional performance indicators such as CPU utilization, queries per second, and active connections, it predicts the future operating status of the database, providing data support for operational decisions such as capacity planning, elastic scaling, and proactive fault handling. The increasing dynamism and complexity of current business workloads have placed higher demands on the accuracy, real-time performance, and reliability of database trend prediction, driving prediction technology to evolve from a fixed single-model architecture towards a dynamic, adaptive, and self-evolving approach.

[0003] Existing database trend prediction technologies still suffer from two core limitations. First, fixed model architectures cannot adapt to the differentiated needs of dynamic operating conditions. Database operations can exhibit three distinct operating conditions: stable normalcy, gradual trend changes, and sudden fluctuations. Traditional solutions employ a unified model architecture, either sacrificing prediction accuracy for computational efficiency with lightweight statistical models or consuming significant computing power for fitting accuracy with deep learning models. Neither can achieve a dynamic optimal balance under dynamically changing load scenarios. Second, there are obstacles to the reliable implementation of prediction results. Most existing solutions only output single-point prediction values, and even when confidence indicators are included, they are merely supplementary information in reports, not linked to the decision-making authority of automated operations and maintenance. Operations and maintenance personnel cannot quantify the reliability of the prediction results, causing the prediction technology to remain at the level of reference and suggestion, making it difficult to directly trigger automated operations and maintenance actions.

[0004] In addition, existing technologies generally adopt a single-point optimization approach, making independent improvements to prediction accuracy, computational efficiency, or confidence calculation, without integrating these three aspects into a unified framework to break their inherent coupling constraints, thus failing to achieve system-level synergistic optimization. Furthermore, existing feedback mechanisms only focus on parameter updates for a single model, failing to incorporate all link elements such as routing strategies, hysteresis control parameters, and confidence calibration coefficients into a unified iterative system, lacking system-level continuous self-evolution capabilities. With long-term changes in business models, the performance of the prediction system will gradually degrade, requiring regular manual tuning and maintenance, resulting in high long-term operational costs and making it difficult to adapt to the operational needs of rapid business iterations. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a database trend prediction solution driven by both operating conditions and confidence levels.

[0006] According to a first aspect of the present invention, a database trend prediction method driven by both operating conditions and confidence levels is proposed, comprising: S1. Collect multi-dimensional operational performance indicators from the database and obtain the raw multi-dimensional time-series data after time alignment processing; S2. Perform preprocessing on the original multidimensional time series data to obtain standardized multidimensional time series data; S3. Based on standardized multi-dimensional time series data, the operating condition is identified through a parallel operating condition perception mechanism with long and short windows, and the current operating condition label, load fluctuation comprehensive index, trend slope characteristics and emergency operating condition prediction signal are output. S4. A hierarchical hysteresis loop mechanism driven by both operating conditions and confidence levels is used to determine the final target prediction model from a pre-defined three-level gradient prediction model pool. Specifically, this includes: S41. Based on the current operating condition label, assign corresponding basic routing weights to each prediction model in the three-level gradient prediction model pool. S42. Retrieve the historical confidence performance of each prediction model under the corresponding working condition, generate the confidence adjustment coefficient for each model, and correct the basic routing weight of the corresponding model based on the confidence adjustment coefficient to obtain the corrected routing priority of each model and determine the initial target model. S43. If the emergency condition prediction signal is detected to be valid, the prediction model with the highest complexity in the three-level gradient prediction model pool is directly selected as the final target prediction model, and the step-up switching is performed. S44. If no effective emergency condition prediction signal is detected, the switching judgment is made based on the load fluctuation comprehensive index and trend slope characteristics, combined with the preset graded hysteresis threshold and duration window rules. The current operating model is determined as the final target prediction model, or the initial target model is determined as the final target prediction model. Steps S43 and S44 are mutually exclusive. S5. Call the final target prediction model to perform trend prediction on the standardized multidimensional time series data and output the future multi-step prediction value sequence.

[0007] According to some embodiments, in the method of the first aspect of the present invention, step S3 specifically includes: S31. Perform emergency condition pre-detection on standardized multidimensional time series data through a short window of the first preset duration, calculate instantaneous fluctuation characteristics, and output an effective emergency condition pre-judgment signal when the extreme fluctuation triggering conditions are met. S32. Determine the steady-state operating condition of the standardized multi-dimensional time series data through a long-term window of the second preset duration, extract the comprehensive index of load fluctuation and trend slope features, and output the current operating condition label. The second preset duration is much longer than the first preset duration.

[0008] According to some embodiments, in the method of the first aspect of the present invention, step S32 specifically includes: Perform time series decomposition independently on the standardized multidimensional time series data of each dimension within the long time window to obtain the trend term, period term and residual term corresponding to each dimension; The fluctuation statistics of the residual terms in each dimension are calculated, and the fusion weights of each dimension are allocated by the entropy weight method to obtain a unified load fluctuation comprehensive index. Combining the load fluctuation comprehensive index and the slope characteristics of the trend term, the operating conditions are divided into three categories: stable normal, gradual trend change, and sudden fluctuation.

[0009] According to some embodiments, the method of the first aspect of the present invention further includes a confidence assessment and hierarchical decision-making step SA, specifically including: SA1: Quantitatively evaluate the confidence level of the multi-step predicted value sequence and output the corresponding prediction interval and confidence level. SA2: Execute differentiated operation and maintenance decisions based on confidence level and output corresponding operation and maintenance handling instructions.

[0010] According to some embodiments, in the method of the first aspect of the present invention, step SA1 employs an online-offline two-layer collaborative architecture to perform confidence quantification evaluation: The online real-time estimation process is executed synchronously with the predictive inference process. For different types of predictive models, a corresponding lightweight confidence calculation method is matched, and the prediction interval is output in real time and the confidence level is divided. The offline periodic calibration process generates confidence interval benchmarks for each model under various operating conditions based on historical standardized multidimensional time series data, calculates the coverage deviation of the online estimate and generates calibration coefficients, and reversely corrects the prediction interval width of the online output.

[0011] According to some embodiments, in the method of the first aspect of the present invention, step SA2, the differentiated operation and maintenance decision specifically includes: At a high confidence level, the future multi-step predicted value sequence will be directly connected to the automated operation and maintenance system to trigger corresponding operation and maintenance actions; At a medium confidence level, generate an operation and maintenance suggestion report and send it for manual review; At low confidence levels, a secondary route re-prediction is triggered, mandatory manual review is required, and the corresponding predicted samples are stored in the feedback sample library.

[0012] According to some embodiments, the method of the first aspect of the present invention further includes a full-link self-evolutionary optimization step SB, specifically including: SB1. Obtain the real standardized multidimensional time series data for the corresponding prediction period, calculate the actual prediction error, and screen the effective feedback samples; SB2. Based on the positional matching relationship between the actual prediction error and the corresponding prediction interval, classify the deviation types and implement differentiated optimization strategies; SB3. Synchronously update the optimization results to the preset three-dimensional routing policy table to complete the collaborative iteration of the full-link policy parameters.

[0013] According to some embodiments, in the method of the first aspect of the present invention, in step SB2, the deviation is divided into three categories: overconfidence type, overconservative type, and incompetence type, based on the relationship between confidence level and error location. Overconfidence bias corresponds to scenarios with high confidence levels and actual prediction errors exceeding the prediction range. In this case, the routing weights of the corresponding models are simultaneously reduced and the corresponding samples are included in the incremental training set. Overly conservative bias corresponds to scenarios with low confidence levels and actual prediction errors falling within the prediction range; only the confidence calibration parameters of the corresponding model are adjusted. Insufficient capability bias corresponds to scenarios with low confidence levels and actual prediction errors exceeding the prediction range. In such cases, the routing weights are adjusted simultaneously, and the corresponding samples are included in the incremental training set of a higher-complexity model.

[0014] According to some embodiments, in the method of the first aspect of the present invention, during the preset transition window period after the model switching in step S4, the final predicted value is output using a weighted fusion method, and the calculation formula is as follows:

[0015] in, This represents the predicted value at time t of the original model before the switch. This represents the predicted value of the target model at time t after the switch. The transition weight coefficient is monotonically increasing from 0 to 1, and t is the temporal position within the transition window period.

[0016] According to a second aspect of the present invention, a database trend prediction system driven by both operating conditions and confidence levels is proposed, comprising: The data acquisition module is used to: collect multi-dimensional performance indicators of the database and obtain raw multi-dimensional time-series data after time alignment processing; The data preprocessing module is used to: perform preprocessing on the raw multidimensional time series data to obtain standardized multidimensional time series data; The working condition identification module is used to: identify working conditions based on standardized multi-dimensional time series data through a working condition perception mechanism with parallel long and short windows, and output the current working condition label, load fluctuation comprehensive index, trend slope characteristics, and emergency working condition prediction signal. The model routing module is used to determine the final target prediction model from a pre-defined three-level gradient prediction model pool using a hierarchical hysteresis loop routing mechanism driven by both load conditions and confidence levels. Specifically, it is used for: Based on the current operating condition label, assign corresponding basic routing weights to each prediction model in the three-level gradient prediction model pool; retrieve the historical confidence performance of each prediction model under the corresponding operating condition, generate the confidence adjustment coefficient for each model, and correct the basic routing weights of the corresponding models based on the confidence adjustment coefficients to obtain the corrected routing priority of each model and determine the initial target model. If a valid emergency condition prediction signal is detected, the prediction model with the highest complexity in the three-level gradient prediction model pool is directly selected as the final target prediction model, and a skip-level upward switch is performed; if no valid emergency condition prediction signal is detected, the switch is determined based on the load fluctuation comprehensive index and trend slope characteristics, combined with the preset hierarchical hysteresis threshold and duration window rules, and the current running model is determined as the final target prediction model, or the initially selected target model is determined as the final target prediction model. The trend prediction module is used to: call the final target prediction model to perform trend prediction on standardized multidimensional time series data and output a sequence of predicted values ​​for the next multiple steps.

[0017] Compared with the prior art, the present invention provides the following beneficial effects: 1. Achieve dynamic synergy between prediction accuracy, computational efficiency, and result reliability, while ensuring operational stability during model switching.

[0018] To address the issues of fixed model architectures failing to adapt to dynamic database conditions, lack of reliability control in single-factor routing, and the tendency for boundary oscillations during multi-level model switching, a routing decision mechanism with dual inputs of operating condition characteristics and historical confidence levels is employed. This mechanism matches three-level gradient prediction models on demand, significantly reducing computational overhead under stable operating conditions and ensuring nonlinear prediction accuracy under complex operating conditions. It also provides a safety net for prediction reliability from the model selection stage. Furthermore, through a hierarchical hysteresis state machine and differentiated switching rules of "skipping up and going down step by step," it completely suppresses frequent model oscillations at operating condition boundaries while also considering response speed in extreme and sudden scenarios.

[0019] 2. It balances the response speed to sudden operating conditions with the accuracy of steady-state condition determination.

[0020] To address the issues of strong lag and insufficient single-index judgment dimensions in traditional single-window time series decomposition algorithms, a parallel perception architecture with long and short windows is adopted. The short window quickly identifies extreme sudden loads and outputs pre-trigger signals, while the long window achieves accurate classification of steady-state operating conditions through objective fusion of multivariate independent time series decomposition and entropy weight method. This approach fills the perception gap in sudden scenarios without increasing steady-state computational overhead, thereby improving the comprehensiveness and accuracy of operating condition identification.

[0021] 3. It simultaneously satisfies the real-time nature and statistical accuracy of confidence estimation, supporting the reliable and automated implementation of prediction results.

[0022] To address the issues of traditional confidence calculation methods where real-time performance and accuracy are mutually exclusive, and where confidence scores are disconnected from operational decisions, a two-tiered architecture combining online lightweight estimation and offline periodic calibration is employed. This architecture ensures millisecond-level output latency while enhancing the statistical rigor of confidence intervals. Furthermore, a differentiated decision-making mechanism that links confidence score grading with operational permissions elevates confidence scores from an auxiliary evaluation indicator to a basis for decision control, thus bridging the gap between prediction results and automated operational implementation.

[0023] 4. Improve the targeting and effectiveness of feedback optimization to achieve long-term autonomous adaptation at the system level.

[0024] To address the issues of confusion in traditional feedback optimization, limited single-point parameter iteration, and gradual performance degradation as business evolves, this approach uses the matching deviation between the actual prediction error and the confidence interval as the core feedback signal. This accurately distinguishes between two types of problems: insufficient model capability and confidence calibration deviation, thus avoiding ineffective model retraining. Furthermore, a three-dimensional routing strategy table enables coordinated iteration of all parameters across the entire chain, such as routing weights, calibration parameters, and hysteresis thresholds. This allows the system to autonomously adapt to long-term changes in business models, reducing the cost of manual tuning.

[0025] 5. Eliminate numerical shocks from model switching and improve the continuity of the prediction sequence and the stability of system operation.

[0026] To address the issues of predicted value jumps and disruptions to the continuity of operational decisions caused by hard model switching, a soft switching transition between the old and new models is achieved by weighted fusion output of confidence levels during the switching transition period. This avoids the interference of sudden numerical changes during switching on the operational scheduling and early warning logic, thereby improving the stability of the solution's engineering implementation. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating an embodiment 1000 of a database trend prediction method driven by both working conditions and confidence levels according to the present invention. Figure 2 for Figure 1 A flowchart illustrating step S3 in Example 1000; Figure 3 for Figure 2A flowchart illustrating sub-step S32 within step S3; Figure 4 for Figure 1 A flowchart illustrating step S4 in Example 1000; Figure 5 This is a flowchart illustrating an embodiment 2000 of a database trend prediction method driven by both working conditions and confidence levels according to the present invention. Figure 6 This is a flowchart illustrating an embodiment 3000 of a database trend prediction method driven by both working conditions and confidence levels according to the present invention. Figure 7 This is a schematic diagram of an embodiment 4000 of a database trend prediction system driven by both operating conditions and confidence levels according to the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Figure 1 This is a flowchart illustrating an embodiment 1000 of the database trend prediction method driven by both operating conditions and confidence levels according to the present invention. Figure 1 As shown, Example 1000 includes steps S1-S5.

[0030] In step S1, multi-dimensional performance metrics of the database are collected and, after time alignment processing, the raw multi-dimensional time-series data is obtained. The application scenario in this embodiment is performance trend prediction for a MySQL relational database. The data sampling period is set to 1 minute, and all time-series data are aligned to Unix whole-minute timestamps. The goal of step S1 is to obtain raw database performance data with complete dimensions and strict time alignment, providing basic input for subsequent operational condition identification and predictive modeling.

[0031] Specifically, in step S1, a lightweight data acquisition probe is deployed to collect performance metrics from the operating system layer and the database kernel layer, respectively, with all metrics sampled at a uniform 1-minute sampling period.

[0032] There are four system-level performance metrics: CPU utilization, memory utilization, disk I / O throughput, and network bandwidth utilization. The data comes from the operating system's performance counters and the / proc file system interface.

[0033] There are five database-level performance metrics, including: queries per second, transactions per second, active connections, cache hit rate, and row lock wait time. The data comes from the SHOW GLOBAL STATUS performance view and the INFORMATION_SCHEMA system library of the MySQL database.

[0034] After data collection, time alignment is performed: using whole-minute Unix timestamps as the baseline, the sampled values ​​of all metrics are mapped to their corresponding timestamp positions; for sample points with microsecond-level offsets, nearest neighbor matching is used to align them to the standard timestamp, ensuring that all dimensional metrics correspond one-to-one at the same timestamp, forming the original multidimensional time series data. Each row of the original multidimensional time series data corresponds to a timestamp, and each column corresponds to a performance metric.

[0035] The above nine indicators comprehensively depict the database's operating status from three dimensions: resource consumption, load intensity, and resource competition, covering the core discriminative features of three typical operating conditions. The unified sampling period and time alignment ensure the mathematical rationality of subsequent multi-dimensional feature fusion and avoid feature calculation deviations caused by time misalignment.

[0036] In step S2, preprocessing is performed on the original multidimensional time-series data to obtain standardized multidimensional time-series data. The goal of step S2 is to eliminate missing values, outliers, and dimensional differences in the original data, thereby improving the stability and accuracy of subsequent operational condition identification and predictive modeling. The preprocessing sequentially performs three steps: missing value imputation, outlier filtering, and standardization. The object of processing is the original multidimensional time-series data, and the output is standardized multidimensional time-series data. Step S2 specifically includes: (1) Filling missing values First, the missing data types are categorized based on the number of consecutively missing sampling points: those with ≤3 consecutively missing points are classified as short-term missing data, and linear interpolation is used for completion. The formula for linear interpolation is:

[0037] in, , These are the valid sampled values ​​before and after the missing point, where i and j are the corresponding time indices. This is the fill value for the k-th missing point. Points with more than 3 consecutive missing points are considered long-term missing and are filled using the historical average method, which involves retrieving valid sample values ​​of the corresponding indicator at the same time over the past 7 days and calculating the arithmetic mean as the fill value.

[0038] (2) Outlier filtering An outlier points are identified using a dual verification mechanism of the 3σ criterion and the isolated forest algorithm.

[0039] The 3σ criterion is calculated as follows: for a single indicator, a 24-hour sliding window with a window length of 1440 sampling points is used to calculate the mean μ and standard deviation σ. Values ​​exceeding this criterion are considered valid. Sampling points within a certain range are marked as suspected anomalies.

[0040] The parameters of the Isolation Forest algorithm are set as follows: the number of base evaluators is 100, the subsampling size is 256, and the contamination parameter is set to 0.01. The algorithm calculates anomaly scores for all dimensions of data, and sampling points with scores higher than the threshold are marked as suspected anomalies.

[0041] Sampling points that are simultaneously marked as suspected anomalies by both methods are determined to be valid anomalies, and the arithmetic mean of the two valid sampling points before and after that point is used to replace it.

[0042] (3) Standardization process The Z-score standardization method is used to independently normalize the dimensions of the performance index for each dimension. The calculation formula is as follows:

[0043] Where x is the original value, This represents the average of the indicator over a 24-hour sliding window. This represents the standard deviation of the indicator over the past 24 hours' sliding window. These are the standardized values.

[0044] After processing all dimensions, standardized multidimensional time-series data is generated. Each dimension follows a standard normal distribution, eliminating the impact of dimensional differences on subsequent fusion calculations. Key principle explanation: The three-step preprocessing process covers most data quality issues in the production environment. Linear interpolation and historical averages are adapted to missing data of different lengths. Dual anomaly checks take into account both statistical distribution anomalies and multidimensional feature anomalies. Z-score standardization provides a unified numerical scale for subsequent multi-index entropy weight fusion.

[0045] In step S3, based on standardized multidimensional time-series data, the operating condition is identified through a parallel operating condition perception mechanism with long and short windows, and the current operating condition label, load fluctuation comprehensive index, trend slope characteristics, and emergency operating condition prediction signal are output.

[0046] In some specific embodiments, step S3 employs a parallel computing architecture with both long and short windows, simultaneously performing emergency condition pre-detection and steady-state condition formal determination, balancing response speed in extreme scenarios with determination accuracy in steady-state scenarios. The input to step S3 is the standardized multi-dimensional time-series data, and the output includes the current condition label, load fluctuation comprehensive index, trend slope characteristics, and emergency condition pre-determination signal.

[0047] Optionally, step S3 includes the following steps: performing emergency condition pre-detection on standardized multidimensional time series data through a short window of a first preset duration, calculating instantaneous fluctuation characteristics, and outputting an effective emergency condition pre-judgment signal when the extreme fluctuation triggering conditions are met; performing steady-state condition judgment on standardized multidimensional time series data through a long window of a second preset duration, extracting the comprehensive load fluctuation index and trend slope characteristics, and outputting the current condition label; wherein, the second preset duration is much longer than the first preset duration.

[0048] In step S4, a hierarchical hysteresis loop control mechanism driven by both operating conditions and confidence levels is used to determine the final target prediction model from a pre-defined three-level gradient prediction model pool. The goal of step S4 is to match the optimal prediction model from the pre-defined three-level gradient prediction model pool based on the dual-dimensional input of real-time operating condition characteristics and historical confidence performance, combined with the hierarchical hysteresis control mechanism, to achieve a dynamic balance between prediction accuracy, computational efficiency, and result reliability, while suppressing frequent model oscillations at the operating condition boundaries.

[0049] The inputs to step S4 are the current operating condition label, the comprehensive load fluctuation index, the trend slope characteristics, the emergency operating condition prediction signal, and the current operating model identifier; the outputs are the final target prediction model identifier and the model switching execution command.

[0050] In this embodiment, the three-level gradient prediction model pool is pre-configured with three levels of prediction models, which are arranged in order of increasing computational complexity and fitting ability: the first level is a lightweight temporal decomposition model, the second level is a lightweight GRU network model, and the third level is a complete deep LSTM prediction model. The inference computing power consumption and prediction accuracy of the three levels of models increase sequentially.

[0051] In some specific embodiments, the implementation process of step S4 specifically includes the following sub-steps: In step S41, based on the current operating condition label, the corresponding basic routing weights are assigned to each prediction model in the three-level gradient prediction model pool. In step S42, the historical confidence performance of each prediction model under the corresponding working condition is retrieved, the confidence adjustment coefficients corresponding to each model are generated, the basic routing weights of the corresponding models are corrected based on the confidence adjustment coefficients, the corrected routing priority of each model is obtained, and the initial target model is determined. In step S43, if the emergency condition prediction signal is detected to be valid, the prediction model with the highest complexity in the three-level gradient prediction model pool is directly selected as the final target prediction model, and the step-up switching is performed. In step S44, if no valid emergency condition prediction signal is detected, a switching judgment is made based on the load fluctuation comprehensive index and trend slope characteristics, combined with the preset graded hysteresis threshold and duration window rules. The current operating model is determined as the final target prediction model, or the initially selected target model is determined as the final target prediction model. Steps S43 and S44 are mutually exclusive. Optionally, in some specific embodiments, when a model switch is determined, a weighted fusion method is used to output the prediction result within a preset transition window period to eliminate the numerical jump at the moment of switch. Specifically, within the preset transition window period after the model switch, a weighted fusion method is used to output the final prediction value, and the calculation formula is as follows:

[0052] in, This represents the predicted value at time t of the original model before the switch. This represents the predicted value of the target model at time t after the switch. The transition weight coefficient is monotonically increasing from 0 to 1, and t is the temporal position within the transition window period. Specifically through The calculation is performed, where T is the total length of the transition window period, T=5 in this embodiment; t is the temporal position within the transition window period, taking values ​​from 1 to 5. After the transition period ends, Once the value reaches 1, switch completely to the target model output and stop running the original model.

[0053] In step S4, the linearly weighted soft handover method can avoid the sudden changes in predicted values ​​caused by hard handover, ensure the continuity of the predicted sequence, reduce the interference of model switching on downstream operation and maintenance decisions, and improve the stability of system engineering implementation.

[0054] In step S5, the final target prediction model is invoked to perform trend prediction on the standardized multidimensional time series data, outputting a sequence of predicted values ​​for the next multiple steps. The goal of step S5 is to invoke the corresponding prediction model based on the routing results, perform multi-step trend prediction based on the recent standardized time series window, and output prediction results in a unified format. This step adopts a modular architecture that separates decision-making and execution; the routing module only outputs the model identifier, and the execution layer invokes the corresponding inference logic according to the model type.

[0055] The input to step S5 is the final target prediction model identifier and recent standardized multidimensional time series window data; the output is a sequence of future multi-step prediction values. In this embodiment, the prediction step size is uniformly 10 steps, corresponding to the performance index prediction values ​​for the next 10 minutes.

[0056] In some specific embodiments, in the specific implementation of step S5, the input of the three-level model is standardized multidimensional time series data, and the output is a sequence of predicted values ​​for the corresponding step size, with a unified format.

[0057] 1. Lightweight temporal decomposition model inference The input window length is 30 sampling periods, and the inference process consists of three steps: The first step is to perform STL decomposition on the input time series window to obtain the trend term, period term and residual term. The decomposition parameters are the same as in step S3, and the period length is set to 10 sampling periods.

[0058] The second step is to use the least squares method to fit the linear slope of the most recent 10 sampling points for the trend term, and extrapolate the trend prediction value for the next 10 steps based on this slope; and to perform cyclic extrapolation on the period term based on the period length of 10 sampling points to obtain the period prediction value for the next 10 steps.

[0059] The third step involves fitting the nonlinear changes of the residual terms using a pre-trained gradient boosting regression tree model. The input is the five most recent residual sampling points, and the output is the predicted residual value for the next step. This process is repeated 10 times to obtain the complete residual prediction sequence. The parameters of the gradient boosting regression tree are set as follows: 100 decision trees, maximum tree depth of 3, learning rate of 0.1, and loss function of squared error.

[0060] Finally, the trend forecast, cycle forecast, and residual forecast are added point by point to obtain a sequence of future multi-step forecast values.

[0061] 2. Lightweight GRU network model inference The input window length is 20 sampling periods, the input dimension is 9, corresponding to 9 standardized performance metrics. The model adopts a single-layer GRU architecture with 64 hidden layer units, followed by a fully connected output layer, directly outputting 10-step prediction results.

[0062] During inference, a single forward computation is sufficient to obtain the complete sequence of predicted values ​​for multiple future steps, eliminating the need for iterative generation. The model pre-training uses mean squared error as the loss function, Adam as the optimizer, and a learning rate of 0.001. This model balances nonlinear fitting capability with inference efficiency, keeping the single-sample inference latency within 10 milliseconds.

[0063] 3. Inference of a complete deep LSTM prediction model The input window length is 30 sampling periods, and the input dimension is 9. The model uses a two-layer LSTM architecture, with 128 hidden units per layer, followed by two fully connected layers with dimensions of 64 and 10 respectively, outputting 10-step prediction results. During model training, L2 regularization with a weight decay coefficient of 0.0001 is added to prevent overfitting.

[0064] During inference, a single forward computation is performed to obtain a sequence of predicted values ​​for multiple future steps. The loss function is the mean absolute error, the optimizer is Adam, and the learning rate is set to 0.0005. This model has stronger capabilities in capturing long-term time-series dependencies and nonlinear fitting, and is suitable for complex scenarios such as sudden fluctuations.

[0065] In step S5, the capability gradient of the three-level model is perfectly matched with the complexity of the three types of working conditions, realizing the on-demand allocation of computing resources; the architecture of separating decision and execution makes the model pool scalable, and adding a new model does not require modifying the routing decision logic, but only needs to add the corresponding inference interface in the execution layer.

[0066] According to such Figure 1 The implementation method shown in this invention proposes a core solution that uses a hierarchical hysteresis loop mechanism driven by both operating conditions and confidence levels to match a three-level gradient prediction model. This breaks the structural limitation that a fixed model architecture cannot adapt to dynamic operating conditions, achieving dynamic synergy and optimality among prediction accuracy, computational efficiency, and result reliability. It can significantly reduce computing power consumption and improve prediction response speed under stable operating conditions, while ensuring nonlinear prediction accuracy under complex and fluctuating scenarios. At the same time, hierarchical hysteresis switching control effectively suppresses frequent model oscillations at the operating condition boundary, taking into account both response speed and system stability in extreme and sudden scenarios. This provides an adaptive selection system for database trend prediction that allows for on-demand allocation of computing power and pre-emptive control of reliability.

[0067] Figure 2 for Figure 1 A flowchart illustrating step S3 in Example 1000. (See attached diagram.) Figure 2 As shown, step S3 includes steps S31 and S32.

[0068] In step S31, a short-term window of the first preset duration is used to perform pre-detection of sudden operating conditions on standardized multi-dimensional time-series data, calculate instantaneous fluctuation characteristics, and output an effective pre-judgment signal of sudden operating conditions when the extreme fluctuation triggering conditions are met. Step S31 quickly identifies extreme sudden loads based on the short-term window, providing a pre-trigger signal for the cascading switchover of the routing module. The specific implementation process includes: The window length corresponding to the first preset duration is set to 5 sampling periods, which corresponds to 5 minutes, and the sliding step size is 1 sampling period. The latest standardized multidimensional time series data fragments are captured in real time as the calculation input.

[0069] Calculate the following two types of instantaneous fluctuation characteristics: 1. Extreme Value Bias: The difference between the maximum and minimum values ​​of a single-dimensional indicator within a window. The calculation formula is as follows:

[0070] in, Let j be the numerical sequence of the j-th dimension indicator within the short time window. This is the extreme value deviation of the j-th dimension.

[0071] 2. Instantaneous volatility: The arithmetic mean of the absolute values ​​of the first-order differences of a single-dimensional indicator within a window, calculated using the following formula:

[0072] Where n is the number of sampling points within the short-time window. Let be the value of the j-th dimension at time t. This refers to the instantaneous volatility of the j-th dimension. The extreme volatility trigger condition is set as follows: there exists at least one dimensional indicator whose extreme value deviation is greater than the 95th quantile of the historical extreme value deviation for that dimension, and whose instantaneous volatility is greater than the 90th quantile of the historical instantaneous volatility for that dimension, and this state persists for at least two consecutive sampling periods. When the above conditions are met, a valid emergency condition prediction signal is output; otherwise, the emergency condition prediction signal is invalid.

[0073] The short-time window skips complex time-series decomposition operations and directly calculates extreme values ​​and fluctuation characteristics based on the original standardized data. The computational complexity is O(n), which can realize the identification of sudden working conditions at the millisecond level. It solves the problem of strong lag in traditional time-series decomposition algorithms and provides a lead signal for model switching in extreme scenarios.

[0074] In step S32, a steady-state condition determination is performed on the standardized multidimensional time-series data using a long-term window of a second preset duration. The comprehensive load fluctuation index and trend slope features are extracted, and the current operating condition label is output. The second preset duration is significantly longer than the first preset duration. Step S32 achieves accurate operating condition classification based on the long-term window, providing a steady-state operating condition basis for routing decisions. The window length corresponding to the second preset duration is set to 30 sampling periods, corresponding to 30 minutes, which is much longer than the short-term window length.

[0075] Optionally, step S32 further includes: independently performing time series decomposition on the standardized multidimensional time series data of each dimension within the long-term window to obtain the trend term, period term and residual term corresponding to each dimension; calculating the fluctuation statistics of the residual term of each dimension, allocating the fusion weight of each dimension through the entropy weight method, and calculating a unified load fluctuation comprehensive index; combining the load fluctuation comprehensive index and the slope characteristics of the trend term, classifying the operating conditions into three categories: stable normal, gradual trend change and sudden fluctuation.

[0076] According to such Figure 2The embodiment shown in this invention employs a multivariate operating condition perception architecture with parallel long and short windows. The short window quickly identifies sudden loads and outputs pre-trigger signals based on the original time series extreme value features. The long window achieves accurate determination of steady-state operating conditions through the fusion of multi-dimensional independent time series decomposition and objective weighting using the entropy weight method. This not only solves the shortcomings of traditional single-window time series decomposition algorithms, such as strong lag and inability to quickly respond to extreme fluctuations, but also avoids the subjective bias of single-index judgment and manual weighting. It can simultaneously take into account the response speed of sudden operating conditions and the accuracy of steady-state operating condition determination, providing timely and reliable operating condition input basis for subsequent model routing decisions.

[0077] Figure 3 for Figure 2 A flowchart illustrating sub-step S32 within step S3. (See attached flowchart.) Figure 3 As shown, step S32 further includes steps S321-S323.

[0078] In step S321, time series decomposition is independently performed on the standardized multidimensional time series data for each dimension within the long-term window to obtain the trend term, period term, and residual term corresponding to each dimension. Specifically, step S321 involves independently performing STL (Seasonal and Trend decomposition using Loess) time series decomposition on the standardized multidimensional time series data for each dimension within the long-term window to obtain the trend term, period term, and residual term corresponding to each dimension. The specific implementation includes: STL decomposition is based on the locally weighted regression method, which decomposes univariate time series... Addition can be decomposed into three components, satisfying the formula:

[0079] in, The trend term at time t reflects the long-term direction of the indicator's change. The periodic term at time t reflects the periodic fluctuation pattern of the indicator; The residual term at time t reflects the random fluctuation component of the index. In this embodiment, the STL decomposition parameters are set as follows: the period length is set to 10 sampling periods, corresponding to 10 minutes, which corresponds to the short-term high-frequency fluctuation period of the database load; the number of inner loop iterations is set to 2, and the number of outer loop robust iterations is set to 5; the smoothing window length for the trend term is set to 15 sampling periods, and the smoothing window length for the period term is set to 7 sampling periods; robust mode is enabled to reduce the interference of extreme outliers on the decomposition results.

[0080] The above decomposition process is performed independently on the time series data of each dimension, with no interference between the dimensions, and the corresponding trend term sequence, period term sequence and residual term sequence are output.

[0081] In some specific embodiments, in step S321, STL decomposition has the advantages of strong robustness, ability to handle nonlinear trends, and insensitivity to outliers, and is adapted to the fluctuation characteristics of database performance time series; the multivariate independent decomposition method retains the fluctuation characteristics of each dimension, avoids feature loss caused by direct fusion, and provides accurate fluctuation component input for subsequent entropy weight fusion.

[0082] In step S322, the fluctuation statistics of the residual terms for each dimension are calculated, and the fusion weights for each dimension are allocated using the entropy weight method to obtain a unified comprehensive load fluctuation index. Step S322 calculates the fluctuation statistics of the residual terms for each dimension, and automatically allocates the fusion weights for each dimension using the entropy weight method to obtain a unified comprehensive load fluctuation index. Specific implementation methods include: 1. Calculate the unidimensional fluctuation statistic. For the residual term sequence of each dimension, the standard deviation of the sliding window is calculated as the fluctuation intensity characteristic of that dimension, with the window length set to 5 sampling periods, and calculated using the following formula:

[0083] Where m is the number of sampling points within the sliding window. Let be the residual value of the j-th dimension at time t. This represents the mean of the residuals within the window. This refers to the fluctuation intensity of the j-th dimension.

[0084] 2. Calculating Dimension Weights using the Entropy Weight Method The entropy weight method objectively assigns weights based on the information entropy of indicators. The higher the dispersion of an indicator, the lower its information entropy, the greater the amount of effective information it carries, and the higher its corresponding weight. The calculation steps are as follows: A. Constructing the volatility characteristic matrix: Assume there are p performance dimensions, each corresponding to a volatility intensity value, forming a matrix. Since volatility intensity is a positive indicator, the larger the value, the stronger the volatility; therefore, range normalization is directly applied.

[0085] B. Calculate the weight of the j-th dimension:

[0086] C. Calculate the information entropy of the j-th dimension:

[0087] when At that time, it was agreed .

[0088] D. Calculate the weight of the j-th dimension:

[0089] The sum of the weights of all dimensions is 1.

[0090] 3. Calculate the comprehensive load fluctuation index The load fluctuation comprehensive index is obtained by summing the normalized fluctuation intensity of each dimension with its corresponding weight:

[0091] The load fluctuation index ranges from [0,1], with a larger value indicating more severe overall load fluctuations in the database.

[0092] Using the fluctuation intensity of the residual term as the fusion basis, the interference of trend and periodic terms is eliminated, and only the degree of random fluctuation of the load is reflected, which is more in line with the discrimination requirements of working condition classification. The entropy weight method does not require manual setting of weights, avoids subjective bias, and can adapt to the differences in the importance of each indicator under different business scenarios.

[0093] In step S323, combining the load fluctuation composite index and the slope characteristics of the trend term, the operating conditions are divided into three categories: stable normal, gradual trend change, and sudden fluctuation. In some specific embodiments, the implementation process of step S323 includes: First, the trend slope feature is calculated: For the trend term sequence of each dimension, the linear trend of the past 10 sampling points is fitted using the least squares method to obtain the trend slope of each dimension; the weighted average of the trend slopes of all dimensions is then taken, with the weights consistent with those of the entropy weight method, to obtain the overall trend slope k. The formula for calculating the slope fitted by the least squares method is:

[0094] Where n is the number of sampling points within the fitting window. Let be the trend term value at time t. Two-level classification thresholds are set: the first threshold is 0.3, serving as the boundary between stable conditions and gradual trend changes; the second threshold is 0.7, serving as the boundary between gradual trend changes and sudden fluctuations. The specific judgment rules for the three types of operating conditions are as follows: Stable normal state: The comprehensive load fluctuation index is <0.3, and the absolute value of the overall trend slope is <0.01; Gradual trend change: 0.3 ≤ load fluctuation comprehensive index < 0.7, and the overall trend slope remains in the same direction for three consecutive sampling periods, continuously positive or continuously negative; Sudden fluctuations: The comprehensive load fluctuation index is ≥0.7, or the absolute value of the overall trend slope changes by more than 0.05.

[0095] After the determination is completed, the corresponding current working condition label is output.

[0096] Step S323 uses a dual-dimensional approach of fluctuation intensity and trend pattern to determine the load, covering both the severity of load fluctuations and the direction of load change. The classification results are more consistent with the actual operating status of the database. The quantified threshold rules ensure the consistency and reproducibility of the operating condition determination.

[0097] Figure 4 for Figure 1 A flowchart illustrating step S4 in Example 1000. (See attached diagram.) Figure 4 As shown, step S4 includes steps S41-S44.

[0098] In step S41, based on the current operating condition label, corresponding basic routing weights are assigned to each prediction model in the three-level gradient prediction model pool. Optionally, step S41 assigns basic routing weights to the three-level models based on the current operating condition label, reflecting the theoretical fit of different models to the corresponding operating conditions.

[0099] In some specific embodiments, step S41 is implemented as follows: A pre-stored working condition-basic weight mapping table is provided, with the mapping relationship corresponding one-to-one with the three types of working conditions. The weight values ​​range from [0,1], and the sum of the basic weights of the three-level models under the same working condition is 1.

[0100] If the current operating condition label is stable and normal, the basic weight of the first-level lightweight time series decomposition model is 0.7, the basic weight of the second-level lightweight GRU network model is 0.2, and the basic weight of the third-level full-depth LSTM prediction model is 0.1. If the current working condition label is a gradual trend change, the base weight of the first-level lightweight time series decomposition model is 0.2, the base weight of the second-level lightweight GRU network model is 0.6, and the base weight of the third-level full-depth LSTM prediction model is 0.2. If the current operating condition label is sudden fluctuation, the basic weight of the first-level lightweight time series decomposition model is 0.1, the basic weight of the second-level lightweight GRU network model is 0.2, and the basic weight of the third-level full-depth LSTM prediction model is 0.7.

[0101] In step S41, the basic weights are based on the working condition adaptability. The higher the working condition matching degree, the higher the basic weight of the model, ensuring the rationality of model selection in normal scenarios and providing a benchmark for subsequent confidence correction.

[0102] In step S42, the historical confidence performance of each prediction model under the corresponding working condition is retrieved, the confidence adjustment coefficients corresponding to each model are generated, the basic routing weights of the corresponding models are corrected based on the confidence adjustment coefficients, the corrected routing priority of each model is obtained, and the initial target model is determined.

[0103] Step S42 retrieves the historical confidence performance of each model under the corresponding operating condition, generates confidence adjustment coefficients, corrects the basic route weights, obtains the final route priority of each model, and determines the initial target model. Optionally, the specific implementation process of step S42 includes: 1. Historical Confidence Data Retrieval: Using the current operating condition label as an index, retrieve the historical confidence achievement rate of the three models under this operating condition from the 3D routing strategy table. The historical confidence achievement rate is defined as: the proportion of the number of samples in the prediction samples of the same type of operating condition in the past 7 days, where the actual prediction error falls within the corresponding prediction interval, to the total number of valid samples, updated daily.

[0104] 2. Calculation of confidence adjustment coefficient: For the j-th model, its confidence adjustment coefficient is... The calculation formula is:

[0105] in, Let j be the historical compliance rate of the j-th model under the current operating conditions. The preset benchmark compliance rate is set to 0.85 in this embodiment. When When higher than the benchmark, The corresponding weight is positively amplified; when When below the benchmark, The corresponding weight is reduced negatively.

[0106] 3. Calculation of corrected route priority: The final route priority of each model is calculated using a multiplicative correction method, with the following formula:

[0107] in, The base routing weights for the j-th model are... Let be the corrected routing priority for the j-th model.

[0108] 4. Initial Target Model Determination: Select the model with the highest routing priority value among the three-level models as the initial target model.

[0109] Step S42 introduces historical confidence performance as a correction factor into the routing decision, breaking the conventional logic of selecting the model based on a single operating condition. It can dynamically lower the routing priority of models with persistently low confidence and raise the priority of models with high confidence, thus ensuring the reliability of the prediction results from the early stage of the selection process.

[0110] In step S43, if a valid emergency condition prediction signal is detected, the prediction model with the highest complexity in the three-level gradient prediction model pool is directly selected as the final target prediction model, and a skip-level switching is performed. Step S43 executes the skip-level switching logic based on the emergency condition prediction signal to ensure response speed in extreme emergency scenarios.

[0111] In some specific embodiments, in step S43, the status of the emergency condition prediction signal is detected in real time. If the signal is valid, the initial selection result verification and hysteresis judgment process is skipped directly, and the third-level complete depth LSTM prediction model with the highest complexity in the three-level gradient prediction model pool is selected as the final target prediction model, and the step-up switching instruction is output.

[0112] Cross-level switching only supports one-way cross-level switching from a lower-level model to a higher-level model, and does not support cross-level switching downwards.

[0113] The step-by-step triggering mechanism in step S43 bypasses the step-by-step judgment process and can quickly switch to a high-precision model when there is a sudden surge in extreme load, shortening the latency of the pre-fault response and adapting to the prediction needs of extreme scenarios such as sudden database traffic and abnormal access.

[0114] In step S44, if no valid emergency condition prediction signal is detected, a switching determination is made based on the load fluctuation comprehensive index and trend slope characteristics, combined with the preset graded hysteresis threshold and duration window rules. The current operating model is determined as the final target prediction model, or the initially selected target model is determined as the final target prediction model. In step S44, when no cascading switching is triggered, the switching determination is performed based on load fluctuation characteristics and graded hysteresis rules to suppress frequent model oscillations at the operating condition boundary.

[0115] In some specific embodiments, the implementation process of step S44 includes: Two preset hysteresis threshold groups are defined, corresponding to the steady-gradual and gradual-burst operating condition boundaries, respectively. Each group includes an uplink trigger threshold, a downlink recovery threshold, and a corresponding duration window. The steady-gradual boundary threshold group: the uplink trigger threshold is a load fluctuation index ≥ 0.35, corresponding to a duration window of 3 sampling periods; the downlink recovery threshold is a load fluctuation index < 0.25, corresponding to a duration window of 5 sampling periods.

[0116] Gradual-Sudden Boundary Threshold Group: The uplink trigger threshold is a load fluctuation composite index ≥ 0.75, with a corresponding duration window of 2 sampling periods; the downlink recovery threshold is a load fluctuation composite index < 0.65, with a corresponding duration window of 4 sampling periods.

[0117] Optionally, the switching decision follows these rules: Uplink switching determination: When the load fluctuation comprehensive index continuously meets the uplink trigger threshold of the corresponding boundary and the duration reaches the corresponding time window requirement, and the trend slope remains in the same direction, it is determined that the uplink switching condition is met, the initial target model is determined as the final target prediction model, and an uplink switching instruction is output.

[0118] Downlink switching determination: When the load fluctuation composite index continuously falls below the corresponding boundary downlink recovery threshold, and the duration reaches the corresponding time window requirement, the downlink switching condition is deemed met. Downlink switching must be executed step by step; that is, switching from the third-level model can only be done to the second-level model, and switching from the second-level model can only be done to the first-level model. Skipping levels is not allowed. When the condition is met, the initially selected target model is determined as the final target prediction model, and a downlink switching instruction is output.

[0119] Switching conditions not met: If neither the uplink nor the downlink conditions are met, the current operating model will remain unchanged and will be determined as the final target prediction model.

[0120] The hierarchical hysteresis mechanism in step S44 forms a buffer between the uplink trigger threshold and the downlink recovery threshold, and superimposes a duration window constraint, which can effectively eliminate frequent model jumps caused by slight load fluctuations; the hierarchical downlink rule further avoids repeated switching caused by load jitter, ensuring the stability of system operation.

[0121] According to such Figure 4 The implementation shown in this invention employs a dual-drive decision logic that uses operating condition characteristics as the basis for selection and historical confidence as a dynamic correction. It combines a hierarchical hysteresis state machine with differentiated switching rules for skipping up and stepping down to complete model routing. This breaks the limitation of traditional single-factor routing that relies solely on load characteristics. It incorporates prediction reliability into the selection decision-making process from the outset, improving the overall reliability of prediction results from the source. At the same time, it effectively suppresses frequent model jumps at the operating condition boundary through dual threshold buffers and duration constraints. This ensures rapid response in extreme and sudden scenarios while maintaining the stability of steady-state operation, ultimately achieving a dynamic trade-off between prediction accuracy, computational efficiency, and result reliability.

[0122] Figure 5 This is a flowchart illustrating an embodiment 2000 of the database trend prediction method driven by both operating conditions and confidence levels according to the present invention. Figure 5 As shown, Embodiment 2000 includes steps S201-S205, and also includes confidence assessment and grading decision steps SA, specifically steps SA1 and SA2. Among them, steps S201-S205 and... Figure 1 Steps S1-S5 in the illustrated embodiment 1000 are the same and will not be repeated here.

[0123] The goal of step SA is to quantitatively evaluate the reliability of the prediction results and match the corresponding level of operation and maintenance (O&M) handling permissions based on the reliability level, thereby maximizing the efficiency of automated O&M while ensuring O&M security. Step SA takes as input a sequence of future multi-step prediction values, historical standardized multi-dimensional time-series data, and current operating condition labels, and outputs a prediction interval, confidence level, and O&M handling instructions. It consists of two sub-steps: SA1 Confidence Quantification Evaluation and SA2 Tiered O&M Decision Making.

[0124] In step SA1, the confidence level of the future multi-step predicted value sequence is quantitatively evaluated, and the corresponding prediction interval and confidence level are output.

[0125] Optionally, step SA1 employs an online-offline dual-layer collaborative architecture to perform quantitative confidence assessment: the online real-time estimation stage is executed synchronously with predictive inference, matching corresponding lightweight confidence calculation methods for different types of predictive models, and outputting prediction intervals and classifying confidence levels in real time. The offline periodic calibration stage generates confidence interval benchmarks for each model under various operating conditions based on historical standardized multidimensional time-series data, calculates the coverage deviation of the online estimation and generates calibration coefficients, and reversely corrects the width of the prediction interval output online.

[0126] Specifically, the implementation process of step SA1 includes: 1. Online real-time estimation process Online real-time estimation and predictive inference are performed simultaneously. Lightweight confidence calculation methods are matched for different types of predictive models, and the single-sample inference latency is controlled within 50 milliseconds.

[0127] For lightweight time series decomposition models, quantile regression is used to calculate confidence intervals: during the training phase of the gradient boosting tree model with residual fitting, the residual outputs at the 0.025 and 0.975 quantiles are fitted simultaneously; during inference, the predicted residual values ​​at the corresponding quantiles are directly output, and the predicted values ​​of the trend and periodic terms are superimposed to obtain the upper and lower bounds of the 95% confidence interval. The calculation process is completely synchronized with the main prediction, with no additional delay.

[0128] For both the lightweight GRU network model and the full deep LSTM prediction model, the MC-Dropout method is used to calculate the confidence interval: during the model training phase, a Dropout layer is embedded in the fully connected layer before the output layer, and the Dropout rate is set to 0.2; during the inference phase, the Dropout layer is kept on, and 20 independent forward computations are performed to obtain 20 sets of prediction result sequences, which constitute the empirical distribution of the prediction values; the 0.025 and 0.975 quantiles are calculated for the 20 results at each prediction step, respectively, as the upper and lower bounds of the 95% confidence interval at the corresponding time.

[0129] After completing the interval calculation, confidence levels are assigned based on the relative width of the intervals. The formula for calculating the relative width of the intervals is:

[0130] in, The upper bound of the interval is predicted at time t. The lower bound of the interval is predicted at time t. Let t be the predicted value at time t, and N be the total number of prediction steps.

[0131] Two threshold levels are set: the first threshold is 0.1, and the second threshold is 0.3. An interval relative width less than 0.1 is considered high confidence, an interval relative width between 0.1 and 0.3 is considered medium confidence, and an interval relative width greater than 0.3 is considered low confidence.

[0132] 2. Offline periodic calibration process Offline periodic calibration is performed regularly during the daily off-peak business hours, with the execution window being from 2:00 AM to 4:00 AM. The calibration coefficients are updated based on historical data from the past 7 days.

[0133] First, the historical samples are grouped according to the working condition label and model type, with each group having a valid sample size of no less than 200 groups. For each group of samples, the Bootstrap resampling method is used to generate a confidence interval benchmark. Specifically, 1000 samplings with replacement are performed, with each sampling size being 80% of the original sample size. Based on the samples from each sampling, the corresponding model is trained and the prediction interval is calculated. The interval coverage of all sampling results is statistically analyzed to obtain the statistical benchmark coverage of the model under the working condition.

[0134] Next, calculate the coverage bias of the online estimate: statistically analyze the actual coverage of the online output confidence interval over the past 24 hours, calculate the difference between the actual coverage and the nominal confidence level of 0.95, and obtain the coverage bias.

[0135] Finally, the interval width calibration coefficient is generated, and the calculation formula is as follows:

[0136] in, The standard normal quantile corresponds to a 95% confidence level. This represents the standard normal quantile corresponding to the actual online coverage. When the actual online coverage is lower than the nominal level, the calibration coefficient k is greater than 1, widening the prediction interval; when the actual online coverage is higher than the nominal level, the calibration coefficient k is less than 1, narrowing the prediction interval.

[0137] The calibration coefficients are stored separately according to operating conditions and model. During online inference the following day, the calibration coefficients are used to correct the range width of the online output. The correction formula is as follows:

[0138]

[0139] in, , These are the upper and lower bounds of the corrected prediction interval, respectively.

[0140] The two-layer collaborative architecture in step SA1 resolves the contradiction of the traditional Bootstrap method's high computational overhead and inability to adapt to online real-time scenarios. It shifts the application scenario of Bootstrap from online inference to offline benchmark calibration, retaining its advantage of statistical accuracy while meeting the low latency requirements of database operation and maintenance scenarios. The lightweight estimation method with model-specific adaptation is fully matched with the three-level model routing mechanism, ensuring the real-time performance and reasonableness of confidence output under models of different complexities.

[0141] In step SA2, differentiated operation and maintenance decisions are made based on the confidence level, and corresponding operation and maintenance handling instructions are output.

[0142] Optionally, in step SA2, the differentiated operation and maintenance decision specifically includes: 1. At a high confidence level, the future multi-step predicted value sequence directly connects to the automated operation and maintenance system to trigger corresponding operation and maintenance actions, specifically including: In the future, the multi-step predicted value sequence and the corresponding corrected prediction interval will be directly connected to the automated operation and maintenance system. The system will automatically trigger corresponding operation and maintenance actions based on the prediction results, including three categories: elastic expansion of computing resources, dynamic optimization of database parameters, and smooth scheduling of business traffic. No manual review is required before the action is executed. The entire execution process records operation logs and effect data, and the execution results are automatically verified after execution.

[0143] 2. At a medium confidence level, generate an operations and maintenance suggestion report and send it for manual review. This includes: The system automatically generates a standardized operation and maintenance suggestion report, which includes the predicted value curve for the future period, the 95% confidence interval range, the basis for judging the corresponding operating conditions, the explanation of the uncertainty range, and the suggested operation and maintenance actions. The report is pushed to the operation and maintenance personnel's terminal. After manual review and confirmation of the rationality of the prediction and the necessity of the action, a decision is made on whether to execute the corresponding operation and maintenance operation.

[0144] 3. At low confidence levels, trigger secondary route re-prediction and mandatory manual review, and store the corresponding predicted samples in the feedback sample library, specifically including: Three types of actions are triggered simultaneously: First, the secondary routing logic is triggered to reroute the current prediction task to a prediction model with a higher level of complexity for secondary prediction; Second, the current prediction result is marked as a high uncertainty state, and manual review is required before any automated operation and maintenance actions can be triggered; Third, the time series window, operating condition label, predicted value sequence, and confidence interval information corresponding to this prediction are completely stored in the feedback sample library for subsequent system self-evolution iteration.

[0145] The hierarchical decision-making mechanism in step SA2 directly binds confidence level with operation and maintenance decision-making authority, solving the problem of prediction implementation from a mechanism perspective. It improves operation and maintenance response efficiency through full automation in high-confidence scenarios and ensures system operation security through manual backup in low-confidence scenarios. The secondary routing design in low-confidence scenarios further realizes the reverse driving of confidence level on the routing process, strengthens the closed-loop nature of the dual-drive mechanism, and at the same time reserves effective optimization samples for the self-evolution process.

[0146] According to such Figure 5 In the implementation method shown, step SA of the proposed solution constructs a two-layer collaborative confidence assessment architecture. In principle, it uses historical benchmark distribution to correct the coverage bias of the prediction interval generated by online estimation. This effectively solves the problems of inaccurate interval width and coverage drift caused by the lack of global prior information in single online confidence estimation, thus ensuring the statistical reliability and real-time performance of the quantified prediction uncertainty. Based on this, it implements high, medium, and low-level funnel-shaped differentiated operation and maintenance decisions according to the precisely quantified confidence level. Technically, it achieves precise decoupling and hierarchical linkage between prediction credibility and automated execution permissions. The final technical effect not only significantly improves the automated execution efficiency of operation and maintenance actions in high-confidence scenarios, but also effectively avoids the risk of misoperation caused by model overconfidence through manual review and forced re-prediction mechanisms in medium and low-confidence scenarios, making the prediction results truly possess high-availability engineering guidance value for production environments.

[0147] Figure 6 This is a flowchart illustrating an embodiment 3000 of the database trend prediction method driven by both operating conditions and confidence levels according to the present invention. Figure 6 As shown, Example 3000 includes steps S301-S305, confidence assessment and hierarchical decision-making step SA, and also includes a full-link self-evolutionary optimization step SB, specifically steps SB1-SB3. Among these, steps S301-S305 and... Figure 5 Steps S1-S5 in Embodiment 2000 are the same, and step SA is also the same as step SA in Embodiment 2000, so they will not be described again here.

[0148] The goal of step SB is to use the deviation between the predicted results and the actual operating data as feedback signals to accurately distinguish the root causes of the deviation, execute differentiated optimization strategies, and achieve integrated collaborative iteration of all-link parameters through a three-dimensional routing strategy table, enabling the system to autonomously adapt to long-term changes in business models. In some specific embodiments, the input of step SB is the real standardized multi-dimensional time-series data for the corresponding prediction period, historical prediction records, corresponding prediction intervals and confidence levels, and the current version of the three-dimensional routing strategy table. The output is the updated three-dimensional routing strategy table and the incremental training sample set for the model. Specifically, it is divided into three sub-steps: SB1 effective sample selection, SB2 deviation classification and differentiated optimization, and SB3 collaborative iteration of all-link parameters.

[0149] In step SB1, obtain the actual standardized multidimensional time series data for the corresponding prediction period, calculate the actual prediction error, and screen valid feedback samples. Obtain the actual operating data for the corresponding prediction period, calculate the prediction error, exclude exogenous interference samples, and screen valid feedback samples with optimization value.

[0150] Optionally, the specific implementation process of step SB1 includes: After the prediction period ends, the original performance data of the corresponding time period is retrieved and processed through a preprocessing procedure that is completely consistent with step S302 to obtain real standardized multidimensional time series data that corresponds one-to-one with the prediction sequence timestamps.

[0151] Calculate the prediction error and interval hit rate for a single sample, including: 1. Single-step absolute error: For each prediction step size t, calculate the true value. Compared with the predicted value absolute error ; 2. Overall Average Relative Error: Calculate the average relative error over the entire prediction window using the following formula:

[0152] Where N is the total number of prediction steps; 3. Interval Hit Rate: The percentage of actual values ​​falling within the corresponding prediction interval within the statistical prediction window out of the total number of steps. The formula is:

[0153] Where I(・) is an indicator function, which takes the value 1 when the condition is met and 0 when the condition is not met. , These are the lower and upper bounds of the prediction interval at time t, respectively. Valid sample screening is performed; samples that simultaneously meet the following three conditions are considered valid feedback samples: No exogenous event interference: Match the operation and maintenance event database. There are no human or external events such as database version changes, hardware failures, business promotions, or traffic shifts within the corresponding time period, thus excluding prediction deviations caused by reasons other than the model itself. Data integrity meets the standard: the integrity of the actual data for the corresponding time period is no less than 90%, the number of consecutive missing points does not exceed 3, and the reliability of error calculation is guaranteed; Operating condition stability meets the standard: the operating condition label remains consistent within the corresponding prediction period, and the duration of the operating condition is no less than 5 sampling periods, excluding samples during the transition period of operating condition switching.

[0154] After screening, valid samples proceed to the subsequent deviation classification stage, while invalid samples are only archived and do not participate in optimization iterations.

[0155] In step SB1, prediction bias caused by exogenous events is not a defect in the capabilities of the model and strategy. Including it in the optimization will interfere with the direction of parameter iteration. Strict sample screening can ensure the purity of the feedback signal and avoid invalid optimization and parameter shift. Multi-dimensional error indicators cover both numerical bias and interval coverage, providing a comprehensive basis for subsequent bias classification.

[0156] In step SB2, the deviation types are classified according to the positional matching relationship between the actual prediction error and the corresponding prediction interval, and a differentiated optimization strategy is executed. Step SB2 classifies the deviation types based on the confidence level and the interval hit rate, and executes differentiated optimization strategies for different root causes of deviation, avoiding the inefficiency caused by blindly uniform optimization.

[0157] In step SB2, based on the relationship between confidence level and error location, biases are categorized into three types: overconfident, overconservative, and undercapable. Overconfident bias corresponds to scenarios with high confidence levels and actual prediction errors exceeding the prediction range; the routing weights of the corresponding model are simultaneously reduced, and the corresponding samples are included in the incremental training set. Overconservative bias corresponds to scenarios with low confidence levels and actual prediction errors falling within the prediction range; only the confidence calibration parameters of the corresponding model are adjusted. Undercapable bias corresponds to scenarios with low confidence levels and actual prediction errors exceeding the prediction range; the routing weights are simultaneously adjusted, and the corresponding samples are included in the incremental training set of a model with a higher level of complexity.

[0158] In some specific embodiments, during the implementation of step SB2, the interval hit rate threshold is set to 0.8. When the interval hit rate of the sample is ≥0.8, it is determined that the actual prediction error falls entirely within the prediction interval; when the interval hit rate is <0.8, it is determined that the actual prediction error exceeds the prediction interval. Based on the confidence level, the bias is divided into three categories, corresponding to different optimization strategies: 1. Overconfidence bias Corresponding scenario: The confidence level is high and the interval hit rate is <0.8, which means that the model is too optimistic about its own predictive ability, the given confidence interval is too narrow, and the prediction accuracy does not meet expectations.

[0159] Optimization strategy: First, reduce the basic routing weight of the model under the corresponding working condition by 0.05 per adjustment, and set the lower limit of the weight to 0.05 to avoid the model losing its chance of being selected due to excessively low weight. Second, add the sample to the incremental training set of the corresponding model. The samples are stored according to the working condition. Perform incremental fine-tuning of the model once every weekend. The fine-tuning uses a small learning rate of 1 / 10 of the original learning rate for gradient updates to avoid damaging the original generalization ability of the model.

[0160] 2. Overly conservative bias Corresponding scenario: The confidence level is low and the interval hit rate is ≥0.8, which means that the confidence interval estimate is too broad and the actual predictive ability of the model is better than the prediction. There is only a confidence level calibration bias and no model fitting ability defect.

[0161] Optimization strategy: Without modifying the model parameters, only adjust the confidence calibration coefficient of the model under the corresponding working condition, with a single adjustment of 0.05 to narrow the prediction interval width; the lower limit of the calibration coefficient adjustment is set to 0.6 to avoid insufficient coverage due to excessive interval narrowing.

[0162] 3. Insufficient ability bias Corresponding scenario: The confidence level is low and the interval hit rate is <0.8, which means that the model's fitting ability is insufficient to adapt to the complexity of the current working condition, and the confidence level judgment accurately identifies the capability deficiency.

[0163] Optimization strategy: First, adjust the routing weights, lower the base routing weight of the current model under the corresponding working condition by 0.08, and simultaneously increase the base routing weight of the higher-complexity model by 0.08, so as to guide the higher-precision model to be called first in subsequent similar working conditions; Second, add the sample to the incremental training set of the higher-complexity model to enhance the fitting ability of the high-complexity model under similar complex scenarios.

[0164] For samples with medium confidence levels, regardless of the interval hit rate, no immediate parameter adjustment will be triggered. Instead, they will be included in the statistical sample pool for accumulation. Once the cumulative number of samples of the same type under the same working conditions reaches 50, a unified evaluation and batch optimization will be performed to avoid frequent parameter fluctuations caused by scattered samples.

[0165] In step SB2, the three types of deviations correspond to: 1. Confidence calibration error + insufficient model accuracy; 2. Confidence calibration error only; 3. Insufficient model capability + inaccurate confidence judgment. The three fundamentally different root causes of the problem enable precise targeted optimization through differentiated strategies, avoiding the blind retraining of the model when the error is large, which is a common practice in traditional solutions, and significantly improving the efficiency and effectiveness of feedback iteration.

[0166] In step SB3, the optimization results are synchronously updated to the preset three-dimensional routing policy table, completing the collaborative iteration of the full-link policy parameters. Step SB3 synchronously updates all optimization results to the three-dimensional routing policy table, realizing the integrated collaborative iteration of four types of parameters: routing weight, confidence calibration parameters, hysteresis control parameters, and prediction error baseline. Optionally, the specific implementation of step SB3 includes: The three-dimensional routing strategy table is constructed with operating condition characteristics as the first dimension, prediction model as the second dimension, and parameter type as the third dimension. Each operating condition-model node stores four types of parameters: basic routing weight, confidence calibration coefficient, hysteresis control parameter, and prediction error baseline.

[0167] The parameter update uses a moving average smoothing mechanism to avoid sudden parameter changes caused by a single sample. The update formula is as follows:

[0168] in, These are the updated parameter values ​​obtained from this calculation. The parameters are the historical values ​​before the update, and λ is the update smoothing coefficient, with a value of 0.2. This means the new parameters contribute 20% and the historical parameters contribute 80%, ensuring the stability of the system. During each optimization iteration, three types of update operations are executed simultaneously: Route weight and confidence calibration parameter update: Substitute the weight adjustment value and calibration coefficient adjustment value calculated in SB2 into the above smoothing formula to synchronously update the basic route weight and confidence calibration coefficient under the corresponding working condition-model node; Prediction error baseline update: Based on the average relative error of recent valid samples, the prediction error baseline under the corresponding node is updated on a rolling basis. The error baseline serves as an auxiliary reference for confidence level classification. Hysteresis control parameter periodic adjustment: A global calibration of hysteresis parameters is performed every 7 days. Based on the statistical distribution of load fluctuations over the past 7 days, the baseline levels of the two-level hysteresis thresholds are dynamically adjusted. When the overall load fluctuation level increases by more than 10%, the uplink trigger value and downlink recovery value of the two sets of hysteresis thresholds are simultaneously increased to avoid frequent switching at the operating condition boundary. When the overall load fluctuation level decreases by more than 10%, the two sets of hysteresis thresholds are simultaneously decreased to improve the sensitivity of model switching.

[0169] After all parameters are updated, a new version of the 3D routing policy table is generated, which officially takes effect at midnight the next day. The routing module and confidence assessment module will uniformly call the new version of parameters.

[0170] In step SB3, the three-dimensional routing strategy table incorporates the strategy parameters scattered across various modules into a unified storage and iteration framework, breaking the fragmented state of independent optimization by each module. The three types of parameters—routing weight, confidence calibration, and hysteresis control—have an inherent coupling relationship, and their collaborative iteration can produce system-level efficiency gains that cannot be achieved by optimizing any single parameter alone. This is the core carrier for realizing end-to-end self-evolution. The moving average update mechanism ensures the stability of parameter iteration and avoids system operation fluctuations caused by the optimization process.

[0171] According to such Figure 6 The implementation method shown in this invention proposes a solution that constructs a full-link closed-loop self-evolution mechanism. Based on the spatial location matching relationship between actual observation errors and prediction intervals, it accurately attributes deviations to three typical failure modes: overconfidence, overconservatism, and insufficient capability. Furthermore, it strictly implements differentiated routing weight penalties, confidence calibration parameter adjustments, and hierarchical incremental sample allocation strategies according to the error type. This overcomes the limitations of traditional single-precision feedback, transforming offline feedback into a collaborative, structured iteration of model pool complexity and routing strategy parameters. The ultimate technical effect is that the system not only effectively suppresses long-term performance degradation caused by conceptual drift in the operational environment, but also achieves optimal resource allocation for handling difficult samples with highly complex models and stable samples with lightweight models through accurate deviation type attribution. This results in a dynamic balance between long-term online accuracy, robustness, and resource consumption in the trend prediction system, giving the system continuous self-optimization capabilities.

[0172] Figure 7 This is a schematic diagram of an embodiment 4000 of the database trend prediction system driven by both operating conditions and confidence levels according to the present invention. Figure 7 As shown, embodiment 4000 includes a data acquisition module 401, a data preprocessing module 402, a working condition identification module 403, a model routing module 404, and a trend prediction module 405.

[0173] Data acquisition module 401 is used to: collect multi-dimensional performance indicators of the database and obtain raw multi-dimensional time-series data after time alignment processing; Data preprocessing module 402 is used to: perform preprocessing on the original multidimensional time series data to obtain standardized multidimensional time series data; The working condition identification module 403 is used to: identify the working condition based on standardized multi-dimensional time series data through a working condition perception mechanism with parallel long and short windows, and output the current working condition label, load fluctuation comprehensive index, trend slope characteristics and emergency working condition pre-judgment signal. Model routing module 404 is used to: determine the final target prediction model from a pre-defined three-level gradient prediction model pool using a hierarchical hysteresis loop routing mechanism driven by both operating conditions and confidence levels; specifically, it is used for: Based on the current operating condition label, assign corresponding basic routing weights to each prediction model in the three-level gradient prediction model pool; retrieve the historical confidence performance of each prediction model under the corresponding operating condition, generate the confidence adjustment coefficient for each model, and correct the basic routing weights of the corresponding models based on the confidence adjustment coefficients to obtain the corrected routing priority of each model and determine the initial target model. If a valid emergency condition prediction signal is detected, the prediction model with the highest complexity in the three-level gradient prediction model pool is directly selected as the final target prediction model, and a skip-level upward switch is performed; if no valid emergency condition prediction signal is detected, the switch is determined based on the load fluctuation comprehensive index and trend slope characteristics, combined with the preset hierarchical hysteresis threshold and duration window rules, and the current running model is determined as the final target prediction model, or the initially selected target model is determined as the final target prediction model. The trend prediction module 405 is used to: call the final target prediction model to perform trend prediction on standardized multidimensional time series data and output a sequence of future multi-step predicted values.

[0174] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0175] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A database trend prediction method driven by both operating conditions and confidence levels, characterized in that, include: S1. Collect multi-dimensional operational performance indicators from the database and obtain the raw multi-dimensional time-series data after time alignment processing; S2. Perform preprocessing on the original multidimensional time series data to obtain standardized multidimensional time series data; S3. Based on the standardized multi-dimensional time series data, the working condition is identified through a long and short dual-window parallel working condition perception mechanism, and the current working condition label, load fluctuation comprehensive index, trend slope characteristics and emergency working condition pre-judgment signal are output. S4. A hierarchical hysteresis loop mechanism driven by both operating conditions and confidence levels is used to determine the final target prediction model from a pre-defined three-level gradient prediction model pool. Specifically, this includes: S41. Based on the current operating condition label, assign corresponding basic routing weights to each prediction model in the three-level gradient prediction model pool; S42. Retrieve the historical confidence performance of each prediction model under the corresponding working condition, generate the confidence adjustment coefficient for each model, and correct the basic routing weight of the corresponding model based on the confidence adjustment coefficient to obtain the corrected routing priority of each model and determine the initial target model. S43. If the emergency condition prediction signal is detected to be valid, the prediction model with the highest complexity in the three-level gradient prediction model pool is directly selected as the final target prediction model, and a skip-level upward switching is performed. S44. If no valid emergency condition prediction signal is detected, a switching determination is made based on the load fluctuation comprehensive index and the trend slope characteristics, combined with the preset graded hysteresis threshold and duration window rules, to determine the current operating model as the final target prediction model, or to determine the preliminary target model as the final target prediction model; wherein steps S43 and S44 are mutually exclusive. S5. Call the final target prediction model to perform trend prediction on the standardized multidimensional time series data and output a sequence of future multi-step predicted values.

2. The method according to claim 1, characterized in that, Step S3 specifically includes: S31. Perform emergency condition pre-detection on the standardized multidimensional time series data through a short window of the first preset duration, calculate instantaneous fluctuation characteristics, and output an effective emergency condition pre-judgment signal when the extreme fluctuation triggering condition is met. S32. Determine the steady-state operating condition of the standardized multidimensional time series data through a long-term window of the second preset duration, extract the load fluctuation comprehensive index and the trend slope feature, and output the current operating condition label. The second preset duration is much longer than the first preset duration.

3. The method according to claim 2, characterized in that, Step S32 specifically includes: The standardized multidimensional time series data within the long time window are independently decomposed to obtain the trend term, period term and residual term corresponding to each dimension; The fluctuation statistics of the residual terms in each dimension are calculated, and the fusion weights of each dimension are allocated by the entropy weight method to obtain the unified load fluctuation comprehensive index. Based on the combined load fluctuation index and the slope characteristics of the trend term, the operating conditions are divided into three categories: stable normal, gradual trend change, and sudden fluctuation.

4. The method according to claim 1, characterized in that, It also includes confidence assessment and graded decision steps (SA), specifically including: SA1. Quantitatively evaluate the confidence level of the predicted multi-step value sequence and output the corresponding prediction interval and confidence level. SA2. Execute differentiated operation and maintenance decisions based on the confidence level and output corresponding operation and maintenance handling instructions.

5. The method according to claim 4, characterized in that, Step SA1 employs an online-offline two-tier collaborative architecture to perform confidence quantification assessment: The online real-time estimation process is executed synchronously with the predictive inference process. For different types of predictive models, a corresponding lightweight confidence calculation method is matched, and the prediction interval is output in real time and the confidence level is divided. The offline periodic calibration process generates confidence interval benchmarks for each model under various operating conditions based on historical standardized multidimensional time series data, calculates the coverage deviation of the online estimate and generates calibration coefficients, and reversely corrects the width of the prediction interval output online.

6. The method according to claim 4, characterized in that, In step SA2, the differentiated operation and maintenance decision-making specifically includes: At a high confidence level, the future multi-step predicted value sequence is directly connected to the automated operation and maintenance system to trigger corresponding operation and maintenance actions; At a medium confidence level, generate an operation and maintenance suggestion report and send it for manual review; At low confidence levels, a secondary route re-prediction is triggered, mandatory manual review is required, and the corresponding predicted samples are stored in the feedback sample library.

7. The method according to claim 4, characterized in that, It also includes the end-to-end self-evolutionary optimization step SB, which specifically includes: SB1. Obtain the real standardized multidimensional time series data for the corresponding prediction period, calculate the actual prediction error, and screen the effective feedback samples; SB2. Based on the positional matching relationship between the actual prediction error and the corresponding prediction interval, classify the deviation types and execute differentiated optimization strategies; SB3. Synchronously update the optimization results to the preset three-dimensional routing policy table to complete the collaborative iteration of the full-link policy parameters.

8. The method according to claim 7, characterized in that, In step SB2, based on the relationship between the confidence level and the error location, the deviation is divided into three categories: overconfidence, overconservatism, and incompetence. Overconfidence bias corresponds to scenarios with high confidence levels and where the actual prediction error exceeds the prediction range. In such cases, the routing weights of the corresponding model are simultaneously reduced and the corresponding samples are included in the incremental training set. In scenarios where excessively conservative bias corresponds to a low confidence level and the actual prediction error falls within the prediction range, only the confidence calibration parameters of the corresponding model are adjusted. For scenarios where the capability deficiency bias corresponds to a low confidence level and the actual prediction error exceeds the prediction range, the routing weights are adjusted synchronously and the corresponding samples are included in the incremental training set of a higher-complexity model.

9. The method according to claim 1, characterized in that, During the preset transition window period after model switching in step S4, the final predicted value is output using a weighted fusion method, and the calculation formula is as follows: ； in, This represents the predicted value at time t of the original model before the switch. This represents the predicted value of the target model at time t after the switch. The transition weight coefficient is monotonically increasing from 0 to 1, and t is the temporal position within the transition window period.

10. A database trend prediction system driven by both operating conditions and confidence levels, characterized in that, include: The data acquisition module is used to: collect multi-dimensional performance indicators of the database and obtain raw multi-dimensional time-series data after time alignment processing; The data preprocessing module is used to: perform preprocessing on the original multidimensional time series data to obtain standardized multidimensional time series data; The working condition identification module is used to: identify the working condition based on the standardized multi-dimensional time series data through a working condition perception mechanism with parallel long and short windows, and output the current working condition label, load fluctuation comprehensive index, trend slope characteristics and emergency working condition pre-judgment signal. The model routing module is used to determine the final target prediction model from a pre-defined three-level gradient prediction model pool using a hierarchical hysteresis loop routing mechanism driven by both load conditions and confidence levels. Specifically, it is used for: Based on the current operating condition label, assign corresponding basic routing weights to each prediction model in the three-level gradient prediction model pool; retrieve the historical confidence performance of each prediction model under the corresponding operating condition, generate the confidence adjustment coefficient for each model, and correct the basic routing weights of the corresponding models based on the confidence adjustment coefficients to obtain the corrected routing priority of each model and determine the initial target model. If the emergency condition prediction signal is detected as valid, the prediction model with the highest complexity in the three-level gradient prediction model pool is directly selected as the final target prediction model, and a step-up switch is performed; if the emergency condition prediction signal is not detected as valid, the switch is determined based on the load fluctuation comprehensive index and the trend slope characteristics, combined with the preset graded hysteresis threshold and duration window rules, and the current running model is determined as the final target prediction model, or the initially selected target model is determined as the final target prediction model. The trend prediction module is used to: call the final target prediction model to perform trend prediction on the standardized multidimensional time series data, and output a sequence of predicted values ​​for the next multiple steps.

Images